Springer Proceedings in Complexity

Shu-Heng Chen · Ying-Fang Kao

Ragupathy Venkatachalam · Ye-Rong Du
Editors

Complex Systems
Modeling and
Simulation in
Economics and
Finance
Springer Proceedings in Complexity
Springer Complexity

Springer Complexity is an interdisciplinary program publishing the best research

and academic-level teaching on both fundamental and applied aspects of complex
systems—cutting across all traditional disciplines of the natural and life sciences,
engineering, economics, medicine, neuroscience, social, and computer science.
Complex Systems are systems that comprise many interacting parts with the
ability to generate a new quality of macroscopic collective behavior the
manifestations of which are the spontaneous formation of distinctive temporal,
spatial, or functional structures. Models of such systems can be successfully
mapped onto quite diverse “real-life” situations like the climate, the coherent
emission of light from lasers, chemical reaction-diffusion systems, biological
cellular networks, the dynamics of stock markets and of the Internet, earthquake
statistics and prediction, freeway traffic, the human brain, or the formation of
opinions in social systems, to name just some of the popular applications.
Although their scope and methodologies overlap somewhat, one can distinguish
the following main concepts and tools: self-organization, nonlinear dynamics,
synergetics, turbulence, dynamical systems, catastrophes, instabilities, stochastic
processes, chaos, graphs and networks, cellular automata, adaptive systems, genetic
algorithms, and computational intelligence.
The three major book publication platforms of the Springer Complexity program
are the monograph series “Understanding Complex Systems” focusing on the
various applications of complexity, the “Springer Series in Synergetics”, which
is devoted to the quantitative theoretical and methodological foundations, and the
“SpringerBriefs in Complexity” which are concise and topical working reports,
case-studies, surveys, essays, and lecture notes of relevance to the field. In addition
to the books in these two core series, the program also incorporates individual titles
ranging from textbooks to major reference works.

More information about this series at http://www.springer.com/series/11637

Shu-Heng Chen • Ying-Fang Kao
Ragupathy Venkatachalam • Ye-Rong Du
Editors

Complex Systems Modeling

and Simulation in Economics
and Finance

123
Editors
Shu-Heng Chen Ying-Fang Kao
AI-ECON Research Center, Department of AI-ECON Research Center, Department of
Economics Economics
National Chengchi University National Chengchi University
Taipei, Taiwan Taipei, Taiwan

Ragupathy Venkatachalam Ye-Rong Du

Institute of Management Studies Regional Development Research Center
Goldsmiths, University of London Taiwan Institute of Economic Research
London, UK Taipei, Taiwan

ISSN 2213-8684 ISSN 2213-8692 (electronic)

Springer Proceedings in Complexity
ISBN 978-3-319-99622-6 ISBN 978-3-319-99624-0 (eBook)
https://doi.org/10.1007/978-3-319-99624-0

Library of Congress Control Number: 2018960945

© Springer Nature Switzerland AG 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Contents

On Complex Economic Dynamics: Agent-Based Computational

Modeling and Beyond. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Shu-Heng Chen, Ye-Rong Du, Ying-Fang Kao, Ragupathy Venkatachalam,
and Tina Yu

Part I Agent-Based Computational Economics

Dark Pool Usage and Equity Market Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Yibing Xiong, Takashi Yamada, and Takao Terano
Modelling Complex Financial Markets Using Real-Time
Human–Agent Trading Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
John Cartlidge and Dave Cliff
Does High-Frequency Trading Matter? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Chia-Hsuan Yeh and Chun-Yi Yang
Modelling Price Discovery in an Agent Based Model for Agriculture
in Luxembourg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Sameer Rege, Tomás Navarrete Gutiérrez, Antonino Marvuglia,
Enrico Benetto, and Didier Stilmant
Heterogeneity, Price Discovery and Inequality in an Agent-Based
Scarf Economy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Shu-Heng Chen, Bin-Tzong Chie, Ying-Fang Kao, Wolfgang Magerl,
and Ragupathy Venkatachalam
Rational Versus Adaptive Expectations in an Agent-Based Model
of a Barter Economy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Shyam Gouri Suresh
Does Persistent Learning or Limited Information Matter
in Forward Premium Puzzle?. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Ya-Chi Lin

v
vi Contents

Price Volatility on Investor’s Social Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

Yangrui Zhang and Honggang Li
The Transition from Brownian Motion to Boom-and-Bust
Dynamics in Financial and Economic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Harbir Lamba
Product Innovation and Macroeconomic Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . 205
Christophre Georges

Part II New Methodologies and Technologies

Measuring Market Integration: US Stock and REIT Markets . . . . . . . . . . . . . 223
Douglas W. Blackburn and N. K. Chidambaran
Supercomputer Technologies in Social Sciences: Existing
Experience and Future Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Valery L. Makarov and Albert R. Bakhtizin
Is Risk Quantifiable? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Sami Al-Suwailem, Francisco A. Doria, and Mahmoud Kamel

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
About the Editors

Shu-Heng Chen is a Distinguished Professor in the Department of Economics,

National Chengchi University (NCCU), Taipei, Taiwan. He serves as the Director of
the AI-ECON Research Center at NCCU as well as the editor-in-chief of the Journal
of New Mathematics and Natural Computation (World Scientific) and Journal of
Economic Interaction and Coordination (Springer), and the associate editor for
Computational Economics (Springer) and Evolutionary and Institutional Economics
Review (Springer). Prof. Chen holds a Ph.D. in Economics from the University of
California at Los Angeles. His research interests include computational intelligence,
agent-based computational economics, behavioral and experimental economics,
neuroeconomics, computational social sciences, and digital humanities. He has
more than 250 referred publications in international journals and edited book
volumes. He is the author of the book, Agent-Based Computational Economics:
How the Ideas Originated and Where It Is Going (published by Routledge), and
Agent-Based Modeling and Network Dynamics (published by Oxford, co-authored
with Akira Namatame).

Ying-Fang Kao is a computational social scientist and a research fellow at the

AI-Econ Research Center, National Chengchi University, Taiwan. She received her
Ph.D. in Economics from the University of Trento, Italy in 2013. Her research
focuses on the algorithmic approaches to understanding decision-making in eco-
nomics and social sciences. Her research interests include Classical Behavioural
Economics, Computable Economics, Agent-Based Modelling, and Artificial Intelli-
gence.

Ragupathy Venkatachalam is a Lecturer in Economics at the Institute of Man-

agement Studies, Goldsmiths, University of London, UK. He holds a Ph.D. from
the University of Trento, Italy. Prior to this, he has held teaching and research
positions at the Centre for Development Studies, Trivandrum (India) and AI-Econ
Research Center, National Chengchi University, Taipei (Taiwan). His research areas
include Computable Economics, Economic Dynamics, Agent-Based Computational
Economics, and Methodology and History of Economic Thought.

vii
viii About the Editors

Ye-Rong Du is an Associate Research Fellow at the Regional Development

Research Center, Taiwan Institute of Economic Research, Taiwan. He received his
Ph.D. in Economics from the National Chengchi University, Taiwan in 2013. His
research focuses on the psychological underpinnings of economic behavior. His
research interests include Behavioural Economics, Agent-Based Economics, and
Neuroeconomics.
On Complex Economic Dynamics:
Agent-Based Computational Modeling
and Beyond

Shu-Heng Chen, Ye-Rong Du, Ying-Fang Kao, Ragupathy Venkatachalam,

and Tina Yu

Abstract This chapter provides a selective overview of the recent progress in the
study of complex adaptive systems. A large part of the review is attributed to agent-
based computational economics (ACE). In this chapter, we review the frontier of
ACE in light of three issues that have long been grappled with, namely financial
markets, market processes, and macroeconomics. Regarding financial markets,
we show how the research focus has shifted from trading strategies to trading
institutions, and from human traders to robot traders; as to market processes, we
empathetically point out the role of learning, information, and social networks
in shaping market (trading) processes; finally, in relation to macroeconomics, we
demonstrate how the competition among firms in innovation can affect the growth
pattern. A minor part of the review is attributed to the recent econometric computing,
and methodology-related developments which are pertinent to the study of complex
adaptive systems.

Keywords Financial markets · Complexity thinking · Agent-based

computational economics · Trading institutions · Market processes

This book is the post-conference publication for the 21st International Conference
on Computing in Economics and Finance (CEF 2015) held on June 20–22, 2015
in Taipei. Despite being the largest conference on computational economics for
two decades, CEF has never produced any book volume that documents the path-
breaking and exciting developments made in any of its single annual events.

S.-H. Chen () · Y.-F. Kao · T. Yu

AI-ECON Research Center, Department of Economics, National Chengchi University, Taipei,
Taiwan
Y.-R. Du
Regional Development Research Center, Taiwan Institute of Economic Research, Taipei, Taiwan
R. Venkatachalam
Institute of Management Studies, Goldsmiths, University of London, London, UK

© Springer Nature Switzerland AG 2018 1

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_1
2 S.-H. Chen et al.

For many years, the post-conference publications had always been in the form of
journals’ special issues, which, unfortunately, have ceased to continue in recent
years. Consequently, although the voices of CEF had been loud and clear for
many years on many prominent issues, they may have been forgotten as time goes
by. Without proper archives, it will be difficult for the new-comers to trace the
important contributions that the conferences have made in the field of computational
economics.
Two years ago, Springer launched a new series, Springer Proceedings in
Complexity, to publish proceedings from scholarly meetings on topics related to the
interdisciplinary studies of the science of complex systems. The scope of CEF fits
the mission of this series perfectly well. Not only does CEF deal with problems
which are sufficiently complex to defy an analytical solution from Newtonian
Microeconomics [9], but CEF methods also treat economics as a science of
complex systems, which requires complexity thinking both in terms of ontology and
epistemology [22]. Therefore, when Christopher Coughlin, the publishing editor
of the series, invited us contribute a volume, we considered it to be a golden
opportunity to archive the works presented at CEF 2015, in a way similar to what
we had done previously in the form of journals’ special issues.
However, CEF 2015 had a total of 312 presentations, which covered many
aspects of CEF. To include all of them in a single volume is doubtlessly impossible.
A more practical alternative would be to select an inclusive and involving theme,
which can present a sharp focus that is neither too narrow nor too shallow. It is
because of this consideration that we have chosen one of the most active areas
of CEF, namely agent-based computational economics (ACE), as the main theme
of this book and have included ten chapters which contribute to this topic. These
ten chapters are further divided into three distinct but related categories: financial
markets, market processes and the macroeconomy. Although there are other areas of
ACE that have also made important advances, we believe that without tracking the
development of these three research areas, the view of ACE will become partial or
fragmented. These ten chapters, constituting the first part of the book, will be briefly
reviewed in Sect. 1.
In addition to these ten chapters, we include three chapters that present new
methodologies and technologies to study the complex economic dynamics. Three
chapters are contributions of this kind. The first one is an econometric contribution
to the identification of the existence and the extent of financial integration. The
second one addresses the role of supercomputers in developing large-scale agent-
based models. The last one challenges the capability of formal reasoning in
modeling economic and financial uncertainties. It also advocates a reform of the
economic methodology of modeling the real-world economy. These three chapters,
constituting the second part of the book, will be briefly reviewed in Sect. 2.
On Complex Economic Dynamics: Agent-Based Computational Modeling and Beyond 3

1 Agent-Based Computational Economics

One core issue that ACE seeks to address in economics is how well the real economy
performs when it is composed of heterogeneous and, to some extent, boundedly
rational and highly social agents. In fact, a large body of published ACE works can
be connected to this thread. This issue is of particular importance when students of
economics are nowadays still largely trained under Newtonian economics using the
device of a representative agent who is assumed to be fully rational in seeking to
maximize a utility function. As an alternative research paradigm to the mainstream,
ACE attempts to see how our understanding of the economy can become different
or remain the same when these simplifications are removed.
Part of ACE was originated by a group of researchers, including Brian Arthur,
John Holland, Blake LeBaron, Richard Palmer, and Paul Tayler, who developed an
agent-based model called the Santa Fe Artificial Stock Market to study financial
markets [1, 5]. Quite interestingly, their original focus was not so much on the
financial market per se, i.e., the financial market as an institutional parameter,
but on the exploration of trading strategies under evolutionary learning, the co-
evolution of trading strategies and the emergence of market price dynamics. This
focus drove the early ACE research away from the role of trading mechanisms and
institutional arrangements in the financial markets, which was later found to be a
substantially important subject in computational economics and finance. Section 1.1
will summarize the three ACE works that focus on trading institutions, rather than
trading strategies.
Market process theory investigates how a market moves toward a state of general
economic equilibrium and how production and consumption become coordinated.
Agent-based modeling is a modern tool used to analyze the ideas associated with
a theoretical market process. In Sect. 1.2, we will give an overview of six works
that investigate the price discovery process, market dynamics under individual, and
social learning and market herding behaviors using agent-based simulation.
Macroeconomics studies the performance, structure, behavior, and decision-
making of an economy as a whole. ACE is a modern methodology that is applied to
examine the macroeconomy. In Sect. 1.3, we introduce the work using ACE models
to analyze the macroeconomic dynamics under product innovation.

1.1 Financial Markets

Dark Pools is an alternative trading institution to the regular exchanges that have
gained popularity in recent years. In dark pools trading, there is no order book
visible to the public; hence the intention of trading is not known until the order
is executed. This provides some advantages for the institutional traders who can
obtain a better realized price than would be the case if the sale were executed on a
regular exchange. However, there are also disadvantages in that the order may not
4 S.-H. Chen et al.

be executed because of the lack of information for the order’s counter parties. With
the growing usage of dark pools trading, concerns have been raised about its impact
on the market quality. In the chapter, entitled “Dark Pool Usage and Equity Market
Volatility,” Yibing Xiong, Takashi Yamada, and Takao Terano develop a continuous
double-auction artificial stock market that has many real-world market features.
Using various institutional parametric setups, they conduct market simulations to
investigate the market’s stability under dark pools trading.
The institutional-level parameters that they investigated are:
• Dark pool usage probability (0, 0.2 or 0.4);
• Market-order proportion (0.3–0.8);
• Dark pool cross-probability (0.1–1.0), which is the probability that the buy orders
and sell orders in the dark pool are being crossed at the mid-price of the exchange.
A lower cross-probability indicates a relatively longer order execution delay in
the dark pool.
Their simulation results indicated that the use of mid-price dark pools decreases
market volatility, which makes sense because the transaction is not visible to the
public until the order is completed. The transactional impact on the market stock
prices is therefore minimized. Moreover, they found that the volatility-suppressing
effect is stronger when the dark pool usage is higher and when the market-order
proportion submitted to the dark pool is lower.1 They also reported that the dark
pool cross-probability did not have any effects on the market volatility.
Another trend in recent financial markets is the use of computer algorithms
to perform high frequency trading (HFT). Since computer programs can execute
trades much faster than humans, stocks and other instruments exhibit rapid price
fluctuations (fractures) over sub-second time intervals. One infamous example is
the flash crash on May 6, 2010 when the Dow Jones Industrial Average (DJIA)
plunged by around 7% (US$1 trillion) in 5 min, before recovering most of the fall
over the following 20 min. To understand the impact of HFT on financial markets,
in the chapter, entitled “Modelling Complex Financial Markets Using Real-Time
Human-Agent Trading Experiments,” John Cartlidge and Dave Cliff used a real-
time financial-market simulator (OpEx) to conduct economic trading experiments
between humans and automated trading algorithms (robots).
The institutional-level parameters that they investigated included:
• Robots’ trading speed, which is controlled by the sleep-wake cycle (ts ) of robots.
After each decision (buy, sell, or do nothing) is made, a robot will sleep for ts
milliseconds before waking up to make the next decision. The smaller that ts is,
the faster the robots’ trading speed and the higher their trading frequency.
• Cyclical vs. random markets: In each experiment, there are six pre-generated
assignment permits, each of which contains a permit number and a limit price—
the maximum value at which to buy, or the minimum value at which to sell.

1 See [16] for similar findings using empirical data.

On Complex Economic Dynamics: Agent-Based Computational Modeling and Beyond 5

The lower the permit number is, the farther away the limit price is from the
equilibrium. In a cyclical market, the permits are issued to humans and robots
following the permit numbers. By contrast, the permits are issued in random
order in a random market.
Their simulation results showed that, under all robot and human market setups,
robots outperform humans consistently. In addition, faster robot agents can reduce
market efficiency and this can lead to market fragmentation, where humans trade
with humans and robots trade with robots more than would be expected by chance.
In terms of market type, the cyclical markets gave very different results from those
of random markets. Since the demand and supply in the real-world markets do not
arrive in neat price-ordered cycles like those in the cyclical markets, the results from
cyclical markets cannot be used to explain what happened in the real-world financial
markets. The authors used these two types of markets to demonstrate that, if we want
to understand complexity in the real-world financial markets, we should move away
from the simple experimental economic models first introduced in the 1960s.
In the chapter, entitled “Does High-Frequency Trading Matter?”, Chia-Hsuan
Yeh and Chun-Yi Yang also investigated the impact of HFT on market stability, price
discovery, trading volume, and market efficiency. However, instead of conducting
real-time experiments using humans and robots, they developed an agent-based
artificial stock market to simulate the interaction between HFT and non-HFT agents.
In addition, unlike the robots in the previous chapter that used pre-generated permits
to submit buy and sell orders for price matching, the agents in this study are more
sophisticated in terms of using heuristics to make trading decisions. Moreover, the
agents have learning ability to improve their trading strategies through experiences.
In their agent-based model, the trading speed is implemented as the agents’
capability to process market information for decision-making. Although instant
market information, such as the best bid and ask, is observable for all traders, only
HFT agents have the capability to quickly process all available information and to
calculate expected returns for trading decisions. Non-HFT agents, however, only
have the capability to process the most recent k periods’ information. The smaller
that k is, the greater the advantage that the HFT agents have over non-HFT agents.
The institutional-level parameters that they investigated include:
• The number of HFT agents in the market (5, 15, 20);
• The activation frequency of HFT agents, which is specified by the number of
non-HFT agents (m = 40, 20, 10) that have posted their quotes before an HFT
agent can participate in the market. The smaller that m is, the more active the
HFT agents are in participating in the trading.
Their simulation results indicated that market volatilities are greater when there
are more HFT agents in the market. Moreover, a higher activation frequency of the
HFT agents results in greater volatility. In addition, HFT hinders the price discovery
process as long as the market is dominated by HFT activities. Finally, the market
efficiency is reduced when the number of HFT agents exceeds a threshold, which is
similar to that reported in the previous chapter.
6 S.-H. Chen et al.

1.2 Market Processes

The agriculture market in Luxembourg is thin, in terms of volume turnover, and

the number of trades in all commodities is small. While the information on market
products can be obtained through an annual survey of the farmers, the market
products trading price information is not accessible to the public. In the chapter,
entitled “Modelling Price Discovery in an Agent Based Model for Agriculture
in Luxembourg,” Sameer Rege, Tomás Navarrete Gutiérrez, Antonino Marvuglia,
Enrico Benetto, and Didier Stilmant have proposed an agent-based model to
simulate the endogenous price discovery process under buyers and sellers who are
patient or impatient in submitting their bid/ask quotes.
In this model, agents are farmers whose properties (area, type, crops, etc.) are
calibrated using the available survey data. The model is then used to simulate a
market that contains 2242 farmers and ten buyers to trade 22 crops for four rounds.
In each round, after all buyers and farmers have submitted the quantity and price for
a commodity to buy or sell, the buyer who offers the highest price gets to purchase
the desired quantity. If only partial quantity is satisfied under the offered price, the
unmet quantity is carried over to the remaining rounds. Similarly, the sellers whose
products do not get sold under the offered price are carried over to the remaining
rounds. Based on the trading price in the initial round, buyers and sellers can adjust
their bid/ask prices in the remaining rounds to achieve their trading goals.
Some buyers/sellers are impatient and want to complete the trading in the next
round by increasing/decreasing the bid/ask prices to the extreme, while others are
more patient and willing to gradually adjust the prices during each of the remaining
three rounds. Based on their simulation, they found that the trading quantities and
prices produced by patient and by impatient traders have very different distributions,
indicating that traders’ behaviors in submitting their bids/asks can impact the price
discovery process in an economic market.
In the chapter, entitled “Heterogeneity, Price Discovery and Inequality in an
Agent-Based Scarf Economy,” Shu-Heng Chen, Bin-Tzong Chie, Ying-Fang Kao,
Wolfgang Magerl, and Ragupathy Venkatachalam also used an agent-based model
to investigate the price discovery process of an economic market. However, their
agents are different from those in the previous chapter in that they apply individual
and social learning to revise their subjective prices. The focus of this work is to
understand how agents’ learning behaviors impact the efficacy of price discovery
and how prices are coordinated to reach the Walrasian equilibrium.
The model is a pure exchange economy with no market makers. Each agent has
its own subjective prices for the commodities and agents are randomly matched for
trading. The learning behavior of an agent is influenced by the intensity of choice
λ, which specifies the bias toward the better-performing prices in the past. When λ
is high, the agent trusts the prices that have done well (the prices can be from self
and from other agents) and uses them to adjust its prices for the future trades. If λ
is low, the agent is more willing to take risk incorporating prices that have not done
well in the past for the future trades.
On Complex Economic Dynamics: Agent-Based Computational Modeling and Beyond 7

Their simulation results showed that agents with a low λ (0–3) have their
subjective prices converging close to the Walrasian equilibrium. This means risk-
taking agents are good at discovering prices toward the general equilibrium.
Moreover, some agents with a large λ (>4) also have their market prices converging
to the general equilibrium. The authors analyzed those high λ (>4) agents in more
detail and found those agents to also be imitators who copied prices that have done
well in the past to conduct most of their trades. This strategy enhanced their price
coordination toward the general equilibrium.
In terms of accumulated payoffs, the agents with low λ (0–3) who also mixed
innovation and imitation in adjusting their subjective prices have obtained medium
or high payoffs. Meanwhile, the agents with high λ (>4) who are also imitators have
received very high payoffs. Finally, the high λ (>4) agents who are also reluctant
to imitate other agents’ prices have received abysmal accumulated payoffs. Based
on this emerging inequality of payoffs, the authors suggested that different learning
behaviors among individuals may have contributed to the inequality of wealth in an
economy.
In the chapter, entitled “Rational Versus Adaptive Expectations in an Agent-
Based Model of a Barter Economy,” Shyam Gouri Suresh also investigated market
dynamics under agents with learning ability in a pure exchange or barter economy.
In this direct exchange market, an agent can apply individual or social learning to
predict the productivity level of his next exchange partner. Based on the prediction,
the agent then decides his own productivity level. Under the individual learning
mode, the prediction is based on the productivity level of the agent’s current
exchange partner while in the social learning mode, the prediction is based on the
productivity level of the entire population.
In this model, the productivity level of an agent can be either high or low and
there is a transition table that all agents use to decide their current productivity
level according to their previous productivity. Additionally, an agent can incorporate
his prediction about the productivity level of his next exchange partner to decide
his current productivity level. This prediction can be carried out through either
individual or social learning. Finally, to maximize his utility, an agent only adopts
high productivity when his transition table indicates high productivity and his next
exchange partner is also predicted to have high productivity.
The simulation results showed that the market per capita outputs or average
outputs converged to low productivity under individual learning. This is because
each time when an agent trades with another agent with low productivity, the agent
will decide to produce low outputs in the next period regardless of the productivity
specified by the transition table. This action in turn causes the agent he interacts
with in the next period to produce low outputs in the period subsequent to the next.
When an agent encounters another agent who has produced a high level of outputs,
the agent will only adopt high productivity in the next period if the transition table
also specifies high productivity. As a result, the market average outputs converge to
low productivity.
8 S.-H. Chen et al.

By contrast, the market average outputs converge to high productivity under

social learning, when the population size is large (100 in their case). This is because,
in a large population, the likelihood of the population-wide distribution of produc-
tivity level being extreme enough to cause it to fall below the high-productivity
threshold is low. Consequently, as all agents started with high productivity, the
market average outputs remained high throughout the simulation runs.
In addition to the price discovery process and productivity level prediction,
traders’ learning behaviors might have impacted the forward premium in the foreign
exchange market. In the chapter, entitled “Does Persistent Learning or Limited
Information Matter in the Forward Premium Puzzle?, Ya-Chi Lin investigated
whether the interactions between adaptive learning and limited market information
flows can be used to explain the forward premium puzzle.
The forward premium puzzle in the foreign exchange market refers to the well-
documented empirical finding that the domestic currency is expected to appreciate
when domestic nominal interest rates exceed foreign interest rates [4, 10, 14]. This
is puzzling because economic theory suggests that if all international currencies are
equally risky, investors would demand higher interest rates on currencies expected
to fall, and not to increase in value. To examine if investors’ learning behaviors
and their limited accessibility to market information may explain this puzzle, Lin
designed a model where each agent can learn to predict the expected exchange rates
using either full information (day t and prior) or limited information in the past (day
t − 1 and prior).
In this model, the proportion of agents that have access to full information, n, is
an exogenous parameter. In addition, an agent has a learning gain parameter γ that
reflects the learning strength. They simulated the model under different values of n,
from 0.1 to 1, and γ , from 0.02 to 0.1, and found that the forward premium puzzle
exists under small n for all values of γ . Moreover, when agents were allowed to
choose between using limited or full information for forecasting, all agents switched
to using full information (i.e., n = 1) and the puzzle disappeared for all values of
γ . This suggests that limited information might play a more important role than
learning in explaining the forward premium puzzle. However, regardless of the
values of n and γ , the puzzle disappeared when tested in the multi-period mode.
This indicates that limited information alone is not sufficient to explain the puzzle.
There are other factors involved that will cause the puzzle to occur.
Herding is a well-documented phenomenon in financial markets. For example,
using trading data from US brokerages, Barber et al. [3] and Kumar and Lee [13]
showed that the trading of individual investors is strongly correlated. Furthermore,
based on trading data from an Australian brokerage, Jackson [12] reported that
individual investors moved their money in and out of equity markets in a systematic
manner. To macroscopically study the effects of herding behavior on the stock return
rates and on the price volatility under investors with different interaction patterns,
in the chapter, entitled “Price Volatility on the Investor’s Social Network,” Yangrui
Zhang and Honggang Li developed an agent-based artificial stock market model
with different network structures.
On Complex Economic Dynamics: Agent-Based Computational Modeling and Beyond 9

In their interaction-based herding model, the trading decision of an agent

is influenced by three factors: (1) personal belief; (2) public information, and
(3) neighbors’ opinions. Their work investigated the following institutional-level
parameters:
• Agents’ interaction structures: regular, small-world, scale-free, and random
networks;
• Agents’ trust in their neighbors’ opinions (1–3);
Their simulation results showed that the market volatility is the lowest when
the agents are connected in a regular network structure. The volatility increases
when agents are connected under small-world or scale-free structures. The market
volatility is the highest when agents are connected under a random network
structure. This makes sense as the more irregular the agents’ interaction pattern
is, the higher the price fluctuations and market volatility. In addition, they found
that the more an agent trusts in his neighbors’ opinions, the greater the volatility of
the stock price. This is also expected, as the more weight an agent attaches to his
neighbors’ opinions, the more diverse the trading decisions can be, and hence the
higher that the price volatility becomes.
In the chapter, entitled “The Transition from Brownian Motion to Boom-
and-Bust Dynamics in Financial and Economic Systems,” Harbir Lamba also
investigated herding behaviors in financial markets. However, instead of using a
network model, he proposed a stochastic particle system where each particle is
an agent and agents do not interact with each other. Agents’ herding behavior is
controlled by a herding parameter C, which drives the agents’ states toward the
market sentiment. Using this system, Lamba demonstrated that even a very low
level of herding pressure can cause a financial market to transition to a multi-year
boom-and-bust.
At time t, each agent i in the system can be in one of two possible states,
owning the asset (+1) or not owning the asset (−1), according to its pricing strategy
[Li (t), Ui (t)]. When the asset market price rt falls outside the interval of Li (t)
and Ui (t), agent i switches its state to the opposite state. In addition, when an
agent’s state is different from the state of the majority agents, its pricing strategy
is updated at a rate of C|σ |, where σ is the market sentiment, defined as the average
state of all agents. Hence, agents have a tendency to evolve toward the state of the
majority agents. Finally, the market price rt is the result of exogenous information
and endogenous agent states generated by the agents’ evolving pricing strategies.
Using 10,000 agents to simulate the market for 40 years, their results showed
that even with a low herding parameter value C = 20, which is much lower than
the estimated real market herding pressure of C = 100, the deviations of market
prices away from the equilibrium resemble the characteristics of “boom-and-bust”:
a multi-year period of low-level endogenous activities that convince equilibrium-
believers the system is in an equilibrium state with slowly varying parameters.
There then comes a sudden and large reversal involving cascades of agents switching
states, triggered by the change in market price.
10 S.-H. Chen et al.

1.3 Macroeconomy

Product innovation has been shown to play an important role in a firm’s perfor-
mance, growth, and survival in the modern economy. To understand how product
innovation drives the growth of the entire economy, causing business cycle fluctua-
tions, in the chapter, entitled “Product Innovation and Macroeconomic Dynamics,”
Christophre Georges has developed an agent-based macroeconomic model. In this
model, a hedonic approach is used, where product characteristics are specified and
evaluated against consumer preferences.
The macroeconomic environment consists of a single representative consumer
and m firms whose products are described by characteristics that the consumer
cares about. To satisfy the consumer’s utility function, firms improve their product
characteristic values through R&D investment. If the R&D indeed leads to product
innovation that also recovers the cost, the firm grows. Otherwise, the firm becomes
insolvent and is replaced by a new firm.
A firm can choose to invest or not to invest in R&D activities. The decision is based
on the recent profits of other firms engaging in R&D and then tuned by the firm’s
own intensity parameter γ . When a firm decides to engage in R&D, the probability
that the firm will experience successful product innovation increases.
Using 1000 firms and 50 product characteristics to run simulations, the results
showed that the evolution of the economy’s output (GDP) closely follows the
evolution of the R&D investment spending. Meanwhile, the customer’s utility grows
over time, due to a long-term net improvement in product quality. Moreover, when
the R&D intensity parameter γ is increased, the increased R&D spending drives up
consumption, output, and utility. Finally, ongoing endogenous product innovation
leads to ongoing changes in the relative qualities of the goods and the distribution
of product shares. The distribution tends to become skewed, with the degree of
skewness depending on the opportunities for niching in the product characteristics
space. As the number of firms grows large, the economy’s business cycle dynamics
tends to become dominated by the product innovation cycle of R&D investment.

2 New Methodologies and Technologies for Complex

Economic Dynamics

In addition to the previous ten chapters, this book also includes three chapters, which
may not be directly related to agent-based modeling that may provide some useful
ideas or tools that can help the modeling, simulation, and analysis of agent-based
modeling. We shall also briefly highlight each of them here.
This book is mainly focused on financial markets and market processes. One
issue naturally arising is related to how different markets are coupled or connected,
and to what degree. In the chapter, entitled “Measuring Market Integration: U.S.
Stock and REIT Markets,” Douglas Blackburn and N.K. Chidambaran take up
On Complex Economic Dynamics: Agent-Based Computational Modeling and Beyond 11

the issue of identifying the existence and extent of financial integration. This is
an important methodological issue that empirical studies often encounter, given
the complex relationships and heterogeneity that underpins financial markets. The
authors identify a potential joint hypothesis problem that past studies testing for
financial integration may have suffered from. This problem arises when testing for
the equality of risk premia across markets for a common (assumed) set of risk
factors; nonetheless, there is a possibility that a conclusion claiming a rejection of
integration may actually stem from the markets not sharing a common factor.
Overcoming the joint hypothesis problem means disentangling the two issues and
examining them separately. They present an approach based on factor analysis and
canonical correlation analysis. This approach can be summarized in two steps. First,
one should determine the correct factor model in each market and determine whether
the markets share a common factor. Second, one should develop economic proxies
for the shared common factor and test for the equality of risk premia conditional on
a common factor being present. The equality of risk premia is tested only if common
factors exist. The authors argue that this procedure in fact gives more power to the
tests. They test their method on US REIT and stock markets for 1985–2013.
When one attempts to understand social systems as complex systems, for
instance, through agent-based models, computers and simulations play a very
important role. As the scale and scope of these studies increase, simulations can
be highly demanding in terms of data-storage and performance. This is likely
to motivate more and more researchers to use highly powerful, supercomputers
for their studies as the field matures. In the chapter, entitled “Supercomputer
Technologies in Social Sciences: Existing Experience and Future Perspectives,”
Valery Makarov and Albert Bakhtizin document several forays into supercomputing
in the social science literature.
The authors introduce some open-source platforms that already exist in the
scientific community to perform large-scale, parallel computations. They discuss
their hands-on experience in transforming a pre-existing agent-based model into
a structure that can be executed on supercomputers. They also present their own
valuable experiences and lessons in applying their models to supercomputers. From
their experiences, C++ appears to be more efficient than Java for developing soft-
wares running on supercomputers. The processes and issues related to translating a
Java-based system into a C++ based system are also explained in the chapter.
Social sciences are distinct from natural sciences in terms of the potential of
their theories to have an impact, for better or worse, on the actual lives of people.
The great financial crisis of 2008, as some have argued, is a result of over reliance
on unrealistic models with a narrow world-view, ignoring the complexities of the
financial markets. Should more complex, sophisticated mathematical models be
the solution? In the chapter, entitled “Is Risk Quantifiable?”, Sami Al-Suwailem,
Francisco Doria, and Mahmoud Kamel take up this issue and examine the method-
ological issues related to the use of or over-reliance on “formal” models in the social
sciences, in particular in economics and finance.
The authors question whether the indeterminacy associated with future economic
losses or failures can be accurately modeled and systematically quantified using
12 S.-H. Chen et al.

formal mathematical systems. Using insights from metamathematics—in particular,

Kurt Gödel’s famous theorems on incompleteness from the 1930s—they point to the
inherent epistemological limits that exist while using formal models. Consequently,
they argue that a systematic evaluation or quantification of risk using formal models
may remain an unachievable dream. They draw several examples and applications
from real-world financial markets to strengthen their argument and the chapter
serves as a cautionary message.

3 Conclusion and Outlook

Computational economics is a growing field [6]. With the advancement of technolo-

gies, modern economies exhibit complex dynamics that demand sophisticated meth-
ods to understand. As manifested in this book, agent-based modeling has been used
to investigate contemporary financial institutions of dark pools and high-frequency
trading (chapters “Dark Pool Usage and Equity Market Volatility”, “Modelling
Complex Financial Markets Using Real-Time Human-Agent Trading Experiments”,
and “Does High-Frequency Trading Matter?”). Meanwhile, agent-based modeling
is also used to shed light on the market processes or the price discovery processes by
examining the roles of traders’ characteristics (chapter “Modelling Price Discovery
in an Agent Based Model for Agriculture in Luxembourg”), learning schemes
(chapters “Heterogeneity, Price Discovery and Inequality in an Agent-Based Scarf
Economy” and “Rational Versus Adaptive Expectations in an Agent-Based Model
of a Barter Economy”), information exposure (chapter “Does Persistent Learning
or Limited Information Matter in Forward Premium Puzzle?”), social networks
(chapter “Price Volatility on Investor’s Social Network”), and herding pressure
(chapter “The Transition from Brownian Motion to Boom-and-Bust Dynamics in
Financial and Economic Systems”). Each of these efforts made is a contribution to
enhancing our understanding and awareness of market complexity. Given this extent
of complexity, markets may not perform well for many reasons, not just economic
ones, but also psychological, behavioral, sociological, cultural, and even humanistic
ones. Indeed, market phenomena have constituted an interdisciplinary subject for
decades [11, 15, 17–19]. What agent-based modeling can offer is a framework that
can integrate these interdisciplinary elements into a coherent body of knowledge.
Furthermore, agent-based modeling can also help modern economies that have
been greatly influenced by the big data phenomenon [7]. By applying computational
methods to big data, economists have addressed microeconomic issues in the
internet marketplaces, such as pricing and product design. For example, Michael
Dinerstein and his co-authors [8] ranked products in response to a consumer’s
search to decide which sellers get more business as well as the extent of price
competition. Susan Athey and Denis Nekipelov [2] modeled advertiser behavior and
looked at the impact of algorithm changes on welfare. To work with big data, Google
chief economist Hal Varian proposed machine learning tools as new computational
methods for econometrics [20]. What will the impact of machine learning be
On Complex Economic Dynamics: Agent-Based Computational Modeling and Beyond 13

on economics? “Enormous” answered Susan Athey, Economics of Technology

Professor at Stanford Graduate School of Business. “Econometricians will modify
the methods and tailor them so that they meet the needs of social scientists primarily
interested in conducting inference about causal effects and estimating the impact of
counterfactual policies,” explained Athey [21]. We also expect the collaborations
between computer scientists and econometricians to be productive in the future.

Acknowledgements The authors are grateful for the research support in the form of the Taiwan
Ministry of Science and Technology grants, MOST 104-2916-I-004-001-Al, 103-2410-H-004-
009-MY3, and MOST 104-2811-H-004-003.

References

1. Arthur, W. B., Holland, J., LeBaron, B., Palmer, R., & Tayler, P. (1997). Asset pricing under
endogenous expectations in an artificial stock market. In W. B. Arthur, S. Durlauf, & D. Lane
(Eds.), The economy as an evolving complex system II (pp. 15–44). Reading, MA: Addison-
Wesley.
2. Athey, S., & Nekipelov, D. (2010). A Structural Model of Sponsored Search Advertising
Auctions. Sixth ad auctions workshop (Vol. 15).
3. Barber, B. M., Odean, T., & Zhu, N. (2009). Systematic noise. Journal of Financial Markets,
12(4), 547–569. Amsterdam: Elsevier.
4. Bilson, J. F. O. (1981), The ‘speculative efficiency’ hypothesis. Journal of Business, 54(3),
435–451.
5. Chen, S. H. (2017). Agent-based computational economics: How the idea originated and where
it is going. Routledge.
6. Chen, S. H., Kaboudan, M., & Du, Y. R. (Eds.), (2018). The Oxford handbook of computational
economics and finance. Oxford: Oxford University Press.
7. Chen, S. H., & Venkatachalam, R. (2017). Agent-based modelling as a foundation for big data.
Journal of Economic Methodology, 24(4), 362–383.
8. Dinerstein, M., Einav, L., Levin, J., & Sundaresan, N. (2018), Consumer price search and
platform design in internet commerce. American Economic Review, 108(7), 1820–59.
9. Estola, M. (2017). Newtonian microeconomics: A dynamic extension to neoclassical micro
theory. Berlin: Springer.
10. Fama, E. (1984). Forward and spot exchange rates. Journal of Monetary Economics, 14(3),
319–338.
11. Halteman, J., & Noell, E. S. (2012). Reckoning with markets: The role of moral reflection in
economics. Oxford: Oxford University Press.
12. Jackson, A. (2004). The aggregate behaviour of individual investors, working paper. http://ssrn.
com/abstract=536942
13. Kumar, A., & Lee, C. M. C. (2006). Retail investor sentiment and return comovements. Journal
of Finance, LXI(5). https://doi.org/10.1111/j.1540-6261.2006.01063.x
14. Longworth, D. (1981). Testing the efficiency of the Canadian-U.S. exchange market under the
assumption of no risk premium. The Journal of Finance, 36(1), 43–49.
15. Lonkila, M. (2011). Networks in the Russian market economy. Basingstoke: Palgrave Macmil-
lan.
16. Petrescu, M., Wedow, M., & Lari, N. (2017). Do dark pools amplify volatility in times of
stress? Applied Economics Letters, 24(1), 25–29.
17. Rauch, J. E., & Casella, A. (Eds.), (2001). Networks and markets. New York: Russell Sage
Foundation.
14 S.-H. Chen et al.

18. Staddon, J. (2012). The malign hand of the markets: The insidious forces on wall street that
are destroying financial markets—and what we can do about it. New York: McGraw Hill
Professional.
19. Tuckett, D. (2011). Minding the markets: An emotional finance view of financial instability.
Berlin: Springer.
20. Varian, H. R. (2014, Spring). Big data: New tricks for econometrics. Journal of Economic
Perspectives, 28(2), 3–28.
21. What Will The Impact Of Machine Learning Be On Economics? https://www.forbes.com/.../
what-will-the-impact-of-machine-learning-be-on-economics/), Forbes, Jan 27, 2016.
22. Wolfram, S. (2002). A new kind of science (Vol. 5). Champaign: Wolfram Media.
 Part I
Agent-Based Computational Economics
Dark Pool Usage and Equity Market
Volatility

Yibing Xiong, Takashi Yamada, and Takao Terano

Abstract An agent-based simulation is conducted to explore the relationship

between dark pool usage and equity market volatility. We model an order-driven
stock market populated by liquidity traders who have different, but fixed, degrees
of dark pool usage. The deviation between the order execution prices of different
traders and the volume weighted average price of the market is calculated in an
attempt to measure the effect of dark pool usage on price volatility. By simulating
the stock market under different conditions, we find that the use of the dark pool
enhances market stability. This volatility-decreasing effect is shown to become
stronger as the usage of the dark pool increases, when the proportion of market
orders is lower, and when market volatility is lower.

Keywords Dark pool · Market volatility · Agent-based model · Behavioral

economics · Order-driven market

1 Introduction

In recent years, equity markets have become decentralized electronic networks, with
increasingly fragmented liquidity over multiple venues. Take the US equity market
as an example: between January 2009 and April 2014, the market shares of NYSE
Euronext and NASDAQ OMX declined by approximately one-third and one-quarter,
respectively [13], whereas off-exchange trading volume increased from one-quarter
to more than one-third of the market.
Over the same period, the off-exchange market trading volume in dark pools
(alternative trading venues where orders placed are not visible to other market
participants) increased from 9% to 15%. Trading in dark pools has advantages and
challenges. When an institutional investor uses a dark pool to sell a block of one

Y. Xiong () · T. Yamada · T. Terano

Tokyo Institute of Technology, Yokohama, Kanagawa, Japan
e-mail: ybxiong@trn.dis.titech.ac.jp; tyamada@trn.dis.titech.ac.jp; terano@dis.titech.ac.jp

© Springer Nature Switzerland AG 2018 17

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_2
18 Y. Xiong et al.

million shares, the lack of transparency actually works in the institutional investors
favor, as it may result in a better realized price than if the sale was executed on
an exchange. This is because dark pool participants do not disclose their trading
intention to the exchange prior to execution—as there is no order book visible to
the public, large orders have a much smaller market impact. In addition, there is the
possibility of price improvement if the midpoint of the quoted bid and ask price is
used for the transaction. This can help the investor save half of the spread. However,
the lack of information about counterparties can lead to the uncertain execution of
orders in the dark pool. In general, dark pools offer potential price improvements,
but do not guarantee execution.
The increasing usage of the dark pool raises concerns about the impact of
dark trading on market quality [18]. Although previous studies offered consistent
conclusions on a variety of issues, i.e., market impact is significantly reduced for
large orders, the relationship between dark pool usage and market volatility remains
unclear. Some studies [5, 20] suggest that a higher dark pool market share is
associated with higher market volatility, whereas others draw the opposite [1, 10]
or more complex conclusions [19]. One potential explanation for such contrasting
results is that these studies are conducted based on different dark pools and market
conditions.
Previous studies concerning dark pool usage and market volatility can be classi-
fied into three categories according to their methodologies: (1) using empirical data
from the market; (2) using an equilibrium model to predict the behavior of market
participants; and (3) using an agent-based model to simulate the market dynamics.
In the first case, because different traders have different dark trading participation
rates and trading strategies, conclusions drawn from various markets are likely to be
inconsistent. For example, using transaction data from 2005to 2007 covering three
block-crossing dark pools (Liquidnet, Posit, and Pipeline), Ready [16] showed that
dark pool usage is lower for stocks with the lowest spreads per share. However,
using daily data collected by SIFMA (Securities Industry and Financial Markets
Association) from 11 anonymous dark pools in 2009, Buti et al. [2] found that the
market share of dark pools is higher for lower-spread and higher-volume stocks [2].
In the second category, the traders are considered to be fully rational in seeking
to maximize a utility function [1, 20]. However, the equilibrium methods of these
models are too abstract to be observed in financial markets [6], especially for dark
trading. In addition, because traders are unlikely to exist as monopolists, their
models are only applicable to short trading periods (single-period optimization
model). Unlike such approaches, we investigate the relationship between dark pool
usage and market volatility through zero-intelligence and repeatedly transacting
traders. The transaction scenarios considered by previous agent-based models are
relatively simple: in each trading session, only one agent submits one share order
that will never be repeated [10]. To model the real market more accurately, we
build a continuous double-auction market that allows multiple traders to trade
continuously in environments with different order sizes, and explore how the use
of the dark pool affects market volatility under different market conditions.
Dark Pool Usage and Equity Market Volatility 19

As of April 2014, there were 45 dark pools of three types in the USA: agency
broker or exchange-owned, broker–dealer owned, and electronic market makers [4].
Agency broker and exchange-owned dark pools adopt a classical pricing mechanism
that matches customer orders at prices derived from lit venues, such as the midpoint
of the National Best Bid and Offer (NBBO). Examples include ITG Posit and
Liquidnet, as well as midpoint dark order types offered by NASDAQ, BATS, and
Direct Edge. The simplicity of this kind of dark pool means it is frequently used for
model building [1, 10, 11, 20]. In this paper, we mainly focus on this type of dark
pool.
Without loss of generality, traders can be considered to use dark pools in a
simple way: each one has a fixed probability of using the dark pool during a
transaction. This zero-intelligence agent design, which relies not on utility functions
but an empirically generated distribution that characterizes the aggregate market
participant behavior, has been widely used to represent agent actions. It was first
introduced by Gode and Sunder [8] in a double-auction market, and its modifications
have come to dominate the agent-based model (ABM) limit order book literature
because of the ability to replicate dynamics such as spread variance and price
diffusion rates [3, 7, 9, 14]. With zero-intelligence agents, we can focus on the
overall influence of dark pool usage in the absence of possible nonlinear interactions
arising from diverse individual trading strategies. Thus, we overcome the problems
encountered when using empirical market data.
After describing a model that represents the usage of dark pools in equity
markets, we answer the following two questions:
• How does the usage of dark pool affect market volatility?
• How does this influence vary according to different market conditions?
By analyzing the simulation results, we find that the use of dark pools decreases
market volatility. Higher usage of dark pools strengthens this effect, as does a lower
proportion of market orders and lower market volatility. However, changes in the
cross-probability of dark pool use do not affect market volatility.
The remainder of this paper proceeds as follows: The next section briefly reviews
the relevant literature considering dark pools and market quality. Section 3 describes
the design of an artificial stock market with a dark pool. In Sect. 4, we describe
simulations carried out to explore the relationship between dark pool usage and
market volatility. Section 5 analyzes the results, and Sect. 6 gives our conclusions.

2 Literature Review

Many empirical studies have examined the relationship between dark trading and
market quality. Ready [16] investigated the determinants of trading volume for
NASDAQ stocks in three dark pools, and his results suggest that dark pool usage
is lower for stocks with lower spreads, but higher for higher market volatility.
Similarly, Ray [15] modeled the decision of whether to use a crossing network
20 Y. Xiong et al.

(CN) or a traditional quoting exchange, and derived hypotheses regarding the

factors that affect this decision. He then tested these hypotheses on realized CN
volumes, and found that the likelihood of using CNs increases and then decreases
as the relative bid–ask spread and other measures of market liquidity increase.
Nimalendran and Ray [12] analyzed a proprietary high-frequency dataset, and found
that the introduction of a CN resulted in increased bid–ask spreads and provided
short-term technical information for informed traders. In contrast, Ye [19] modeled
the market outcome when an informed trader can split trades between an exchange
and a CN (dark pool). He found that the CN reduces price discovery and volatility,
and that the dark pool share decreases with increasing volatility and spread.
Although the results of empirical studies largely depend on the data available,
equilibrium models can examine the influence exerted by the dark pool on market
quality, regardless of data limitations. Buti et al. [1] modeled a dynamic financial
market in which traders submit orders either to a limit order book (LOB) or to a dark
pool. They demonstrated that dark pool market share is higher when the LOB is
deeper and has a narrower spread, when the tick size is large, and when traders seek
protection from price impacts. Further, though the inside quoted depth of the LOB
always decreases when a dark pool is introduced, the quoted spreads can narrow for
liquid stocks and widen for illiquid ones. Zhu [20] formed a different opinion. In his
model, the dark pool share is likely to increase and then decrease with increasing
volatility and spread, and higher usage of the dark pool results in a higher spread
and has a greater market impact.
Agent-based simulations can also be applied to test hypotheses regarding the
influence of dark trading. Mo et al. [11] described the costs and benefits of trading
small orders in dark pool markets through agent-based modeling. The simulated
trading of 78 selected stocks demonstrated that dark pool market traders can
obtain better execution rates when the dark pool market has more uninformed
traders relative to informed traders. In addition, trading stocks with larger market
capitalization yields better price improvement in dark pool markets. In another
study, Mizuta et al. [10] built an artificial market model and found that as the dark
pool was used more often, the markets stabilized. In addition, higher usage of the
dark pool reduced the market impacts.
Table 1 summarizes these contrasting findings regarding dark pool usage and
market quality, grouped by different analytical methods.
To develop a more realistic model representing the trading activity in real equity
markets, we introduce a number of improvements on the basis of previous work. For
example, in previous studies, traders only submit their orders once [10] or within a
few trading periods [20], but our model allows the continuous submission of orders
toward the exchange or dark pool. In addition, many studies limit orders to one share
(Buti et al. [2]), but our model handles different order sizes obeying some statistical
property and simulates the process of splitting orders into smaller chunks. Other
important characteristics of real markets are also considered, such as the cancelation
of orders, partially executed orders, and the change of order urgencies according to
price movement. With these new features, we are able to explore various factors that
affect the influence of the dark pool on market volatility.
Dark Pool Usage and Equity Market Volatility 21

Table 1 Selected studies and findings on dark pool usage and market quality (DPS: dark pool
share)
Findings
Method Paper Market condition → DPS DPS → Market quality
Empirical study spread↓ → DPS↓
Ready [16]
volatility↑ → DPS↑
Ray [15] volatility↑ → DPS(↑) DPS↑ → spread↑
market impact↑
spread↓ → DPS↑
Ye [19]
volatility↑ → DPS↓
Equilibrium model depth↑ tick size↑ → DPS↑ DPS↑ → volume↑
spread↓ → DPS↑ DPS↑ → inside quoted
Buti ([2])
depth↓
stock liquidity↑ → DPS↑ DPS↑ → (liquidity)
spread↓
large orders↑ → DPS↑ → (illiquidity)spread↑
volatility↑ spread↑ → DPS↑() DPS↑ → spread↑ mar-
Zhu [20]
ket impact↑
uniformed traders↑ → DPS↑ DPS↑ → adverse
selection↓
Agent-based model DPS↑ → volatility↓
Mizuta [10] tick size↑ → DPS↑
DPS↑ → market
impact↓

3 Model

Our model is based on a simple limit order-driven market model developed by

Maslov [9]. This agent-based model is chosen because many realistic features of
the financial market are generated by only a few parameters. In the original model,
a new trader appears and attempts to make a transaction at each time step. There is an
equal probability of this new trader being a seller or a buyer, and fixed probabilities
of trading one unit of stock at the market price or placing a limit order. Our model
employs the idea of zero-intelligence traders who make their decisions based on
exogenous distributions. To make this model more realistic, we allow multiple
traders to continuously transact in one trading session. The order sizes follow a
log-normal distribution, and large orders are split into small pieces. Furthermore,
order urgency is introduced to determine order price, and this is updated according
to market conditions. Finally, a midpoint dark pool is added as an alternative trading
venue.
Thus, in our model, the market consists of an exchange and a dark pool, and one
single stock is traded by liquidity traders. Their order types, order sizes, and order
urgencies are assigned exogenously, but will be updated based on price movements
in the market and the execution situations of individual orders. After receiving a
trading task, the trader splits his/her order into smaller ones and submits them
22 Y. Xiong et al.

Fig. 1 Intraday trading process

successively. For each submission, the order will be submitted to either the exchange
or the dark pool. The submission prices of orders to the exchange are determined
by both the midprice of the exchange and the order urgency, and the latter will be
adjusted according to price movements and order execution. Orders submitted to the
dark pool do not have a specified price, as they will be executed at the midprice of
the exchange when the dark pool is crossing and when there are enough orders on
the other side of the dark pool. The intraday trading process of a trader is illustrated
in Fig. 1.

3.1 Trading Sessions and Traders

In the model, intraday trading sessions are set as S, a total of D trading days are
considered in the scenario.
The total number of liquidity traders is T , and they are evenly divided into M
groups. Each group is distinguished from others with different but fixed levels of
dark pool usage, denoted as U1 to UM .
Dark Pool Usage and Equity Market Volatility 23

3.2 Order Type, Order Size, and Order Urgency

Trading tasks are generated before simulation. In the task pool, there is an equal
number of buy and sell order types. The order size follows a log-normal distribution.
The order urgency is divided into three levels (low = −1, middle = 0, high = 1). In
the initial stage, the total number of orders of each urgency level follows an a : b :
c distribution. The order urgency is updated after each intraday transaction. Each
trader randomly picks one task from the pool as their trading task at the beginning
of the day.

3.3 Order Submission and Cancelation

When submitting orders, traders use a time-weighted average price with random-
ization (TWAP)-like order placement strategy to mitigate the market impact. Each
trader first splits his/her task order into F equal fractions, then randomly selects
F trading sessions in the day and submits one piece of the order in each chosen
session.
If a prior order has been submitted to the exchange but not fully executed by the
time a new order is about to be submitted to the exchange, the old order will be
canceled and the remaining amount will be added to the new one. However, orders
submitted to the dark pool will not be canceled. This is because orders submitted to
the dark pool follow the time priority mechanism.

3.4 Order Submitted Price

The order submitted price is mainly determined by order urgency. Take buy orders
as an example, and suppose the current best ask is A0 and best bid is B0 . If the
urgency is low (−1), the trader will place it at Δ ticks away from the best bid in the
LOB (B0 −Δ×tick), where Δ ∼ U [0, λ] (λ is the maximum value for the “offset”)
is an integer that follows an even distribution. At the medium urgency level (0), the
trader will place the order at the midquote (0.5 × (A0 + B0 )). This is also the price
executed in the dark pool. If the urgency is high (1), the trader will aggressively take
liquidity up to a limit price impact, denoted as P Imax (P Imax refers to the highest
price impact a trader can suffer), so the submitting price Ps will be

Ps = 0.5 × (A0 + B0 ) × (1 + P Imax ). (1)

This is a market order that attempts to absorb sell orders with prices lower than
the submitting price. The relationship between order urgency and order price is
illustrated in Table 2.
24 Y. Xiong et al.

Table 2 Order urgency and submitting price

Buy order Sell order
Urgency Submit price Probability Submit price Probability
−1 B0 1/(λ + 1) A0 1/(λ + 1)
... 1/(λ + 1) ... 1/(λ + 1)
B0 − tick × λ 1/(λ + 1) A0 + tick × λ 1/(λ + 1)
0 0.5 × (A0 + B0 ) 0.5 × (A0 + B0 )
1 0.5 × (A0 + B0 ) × (1 + P Imax ) 0.5 × (A0 + B0 ) × (1 − P Imax )

3.5 Order Execution in Exchange and Dark Pool

Order execution in the exchange follows the traditional price-then-time mechanism,

and unexecuted orders are stored in the LOB. Order execution in the dark pool
follows the time priority mechanism. Each trader is assumed to have a fixed
probability of using the dark pool, denoted as Ui (the probability of using the dark
pool is exogenously specified). Hence, when a trader decides to submit an order,
he/she has a probability of Ui of placing it on the dark pool. At the end of each
trading session, buy orders and sell orders in the dark pool have a probability of cp
of being crossed at the midprice (0.5 × (A0 + B0 )) of the exchange. Unexecuted
orders in the dark pool are left for the next crossing.
There are two types of midpoint dark pools in the market. For example, in
Liquidnet, the matching acts continuously, whereas ITG only performs order
matching a few times a day [20]. The probability cp is introduced to consider both
situations.

3.6 Order Urgency Updated After Intraday Transaction

The order urgency will be updated by price movements and the order execution
condition after intraday transactions. For example, if the stock price has increased
by a significant level in one trading day, then the sell order urgency will increase and
the buy order urgency will decrease for the next day. Another adjustment considers
the order execution condition. Suppose that one trader has a task order size for a day
of St , and the order executed that day is Sexe . At the end of the day, if there are still
a considerable number of orders that have not been fully executed ((St − Sexe )/St >
U T , where U T is some threshold for unexecuted orders), this trader may continue
to execute the order the next day, but with higher urgency. Suppose that the open
price and close price for a trading day are Popen and Pclose . At the beginning of the
next day, the order urgency will be adjusted according to the following rules:
• if Pclose > Popen × (1 + θ ) (θ is the threshold for urgency adjustment), traders
will pick a new task. If it is a buy order, its urgency has a 50% probability to
Dark Pool Usage and Equity Market Volatility 25

minus 1 (50% chance of maintaining the original urgency); if it is a sell order, its
urgency has a 50% probability to plus 1.
• if Pclose < Popen × (1 − θ ), traders will pick up a new task. If it is a buy order, its
urgency has a 50% probability to plus 1 (50% chance of maintaining the original
urgency); if it is a sell order, its urgency has a 50% probability to minus 1.
• else if (St − Sexe )/St > U T , traders have an equal probability (50%) of
continuing to execute the previous remaining order and increase its urgency with
1, or dropping it and picking a new one.

4 Experiment

To test how different extents of dark pool usage affect market volatility, T traders
are divided into three groups, each distinguished from the others by different but
fixed levels of dark pool usage, set as U1 ,U2 , and U3 , respectively. The probabilities
of dark order submission of these three groups are denoted as [U1 :U2 :U3 ]. U1 is
assigned as the benchmark probability of 0. U2 and U3 refer to low and high usage
of the dark pool, respectively.
In addition to dark pool usage, there are two other factors that may have
a significant influence on the order executed price in the dark pool, and thus
reflect market volatility. The first is the dark pool cross-probability. A lower cross-
probability indicates a relatively longer order execution delay in the dark pool.
In this case, there may be a significant difference between the midprice in the
submission session and the execution price in the dark pool, especially when the
market is volatile. The second factor is the proportion of market orders. Although
an increase in market orders makes the market more volatile and implies a higher
price improvement in dark trading, this increase also reflects a lack of liquidity in
the main exchange, and implies the same situation in the dark pool. Thus, there is an
increased execution risk for the dark orders. During the simulations, different values
are assigned to these two parameters to examine their influence on market volatility.
Table 3 lists the parameter values used in the simulation.
These parameters fall into three categories. Some are effectively insensitive. For
example, the experiments show that there is no significant influence on the results
for 100–3000 agents. The experiments described in this paper consider a relatively
small simulation time. The model parameters evaluated by this method include T ,
D, S, F , and P0 . Some of the parameters are based on the values used in the previous
models, like λ, θ , and U T [9, 10], whereas others are taken from empirical studies,
such as tick and P Imax [16, 19].
Before conducting the experiments, we first confirm that the model can account
for the main characteristics of financial markets, which reflect empirical statistical
moments of a market pricing series. In validating that markets are efficient, it is
common practice to show there is no predictability in the price returns of assets. To
demonstrate this, the autocorrelation of price returns should show that there is no
26 Y. Xiong et al.

Table 3 Parameters in simulation

Description Symbol Value
Number of traders T 100
Number of groups M 3
Trading days D 10
Intraday trading sessions S 100
Order split fractions F 10
Stock initial price P0 10
Tick size tick 0.01
Dark pool cross-probability cp [0.1,1]
Order urgency distribution [a:b:c] [0.9-MO:0.1:MO]
Market order proportion MO [0.3,0.8]
Dark order submission probability [U1 :U2 :U3 ] [0:0.2:0.4]
Order placement depth λ 4
Max price impact P Imax 0.003
Urgency adjust threshold θ 0.05
Unexecuted order threshold UT 0.2

Autocorrelation function of returns

Sample autocorrelation

0.8
0.6

0.4

0.2
0

−0.2
0 5 10 15 20 25 30 35 40 45 50
Lag

Fig. 2 Autocorrelation of returns for dark pool trading model

correlation in the time series of returns. Figure 2 shows that the price movements
generated by the model are in line with the empirical evidence in terms of an absence
of autocorrelation.
The characteristic of volatility clustering is seen in the squared price returns for
securities that have a slowly decaying autocorrelation in variance. In our model,
the autocorrelation function of squared returns displays a slow decaying pattern, as
shown in Fig. 3.
The autocorrelation values are listed in Table 4.
In addition, it has been widely observed that the empirical distribution of
financial returns has a fat tail. The distribution of returns is shown in Fig. 4. The
kurtosis of this distribution is greater than 3, indicating the existence of a fat tail.
Dark Pool Usage and Equity Market Volatility 27

Autocorrelation function of squared returns

Sample autocorrelation

0.8

0.6

0.4

0.2

−0.2
0 5 10 15 20 25 30 35 40 45 50
Lag

Fig. 3 Autocorrelation of squared returns for dark pool trading model

Table 4 Values of autocorrelation returns and squared returns

Returns 1 0.0282 −0.0253 0.0217 0.0628 0.0007 −0.0591 0.0323 0.0078 . . .
Squared returns 1 0.1578 0.1481 0.1808 0.162 0.1216 0.1426 0.1401 0.0836 . . .

Sample return distribution

800
700 Kurtosis = 6.4
600
500
Count

400
300
200
100
0
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Price return(%)

Fig. 4 Return distribution for dark pool trading model

In each simulation, there are T S = D × S = 1000 trading sessions. Assuming

the stock price at session t is Pt , the return at session t is Rt , which is calculated as:

Rt = Ln(Pt /Pt−1 ) (2)

28 Y. Xiong et al.

The daily volatility (Vol) of the market is calculated as:

 ⎛ ⎞
  n
 
n 
2
 1 1
Vol = S ⎝ Rt −
2 Rt ⎠ (3)
TS −1 T S(T S − 1)
i=1 i=1

Assuming that the trading volume at session t is Vt , the volume weighted average
price (VWAP) of the market is calculated as:

TS
t=1 (Pt × Vt )
VWAP(market) = TS
(4)
t=1 Vt

In the same way, we can calculate the VWAP of trader k, denoted as VWAP(k).
The price slippage of trader k (P S(k)) is calculated as:

(VWAP(k) − VWAP(market))/VWAP(market), sell order

P S(k) = (5)
(VWAP(market) − VWAP(k))/VWAP(market), buy order

Let group(A) refer to the group of n traders (k1 , k2 ,. . . , kn ) in which the dark
pool submission probability is A. V (ki ) refers to the trading volume of trader ki .
The VWAP slippage of group A (VWAPS(A)) is then calculated as:
n
i=1 P S(ki ) × V (ki )
VWAPS(A) = n (6)
i=1 V (ki )

The variance of executed prices (with respect to VWAP) of group A (VEP(A)) is

calculated as the weighted square of the price slippage:
n
i=1 P S(ki ) × V (ki )
2
VEP(A) = n (7)
i=1 V (ki )

Inference
Price volatility measures the extent of price changes with respect to an average price.
For different types of traders, the average prices set here are the same (VWAP of
the market). Higher variance in the executed price of a group indicates that the
execution of orders in this group tends to make the stock price deviate more from
the previous price, thus making the market more volatile [17]. Take VEP(0) as
the benchmark. The impact of group(A) on market volatility with respect to the
benchmark (RVEP(A)) is calculated as:

VEP(A)
RVEP(A) = (8)
VEP(0)
Dark Pool Usage and Equity Market Volatility 29

As the market volatility is the combination of the price volatilities of these three
groups, higher price volatility in one group will increase the price volatility of the
whole market. This leads to:
1. If RVEP(A) > 1, usage of the dark pool makes the market more volatile;
2. If RVEP(A) < 1, usage of the dark pool makes the market less volatile.
The price slippage is a derived measurement based on price volatility. The
volatility level is mainly determined by the liquidity situation of the market, and
the price slippage shows how different traders contribute to this volatility level.
The advantage in using price slippage is its ability to compare differences in the
order execution situations of traders (with different dark pool usages) within one
simulation, rather than measuring volatility differences among different simulations.
Because the order-balance, order-quantity, and order-urgency in different simula-
tions may slightly affect market volatility levels, it is more accurate to compare the
price slippages of different dark pool users within one simulation.

5 Result

During each simulation, cp was assigned a value from the set {0.1, 0.2, . . . , 0.9, 1.0}
and MO took a value from {0.3, 0.4, . . . , 0.7, 0.8}. The simulation was repeated
50 times for each setting, giving a total of 3000 simulations. For each simulation,
the market volatility, dark pool usage, and the variance of executed prices of each
group were recorded. The analysis of variance (ANOVA) was used to test whether
different dark pool usage, dark pool cross-probability, and market order proportion
had a significant effect on the values of VEP and RVEP.
The means of VEP(0), VEP(0.2), and VEP(0.4) over the 3000 simulations were
compared by one-way ANOVA to see whether different dark pool usages led to
different values of VEP. The results in Table 5 indicate that VEP changes with the
dark pool usage.
In addition, two-way ANOVA was applied to test the effects of dark pool cross-
probability and market order proportion on the mean of RVEP(0.2) and RVEP(0.4).
Table 6 presents the analysis results for RVEP(0.2).

Table 5 One-way ANOVA to analyze the effect of dark pool usage on VEP
ANOVA table
Source of variation SS df MS F P -value F crit
Between groups 12,534 2 6267 9.62 <0.01 3.0
Within groups 5,862,478 8997 652
Total 5,875,012 8999
SS: sum of the squared errors; df: degree of freedom, MS: mean squared error
30 Y. Xiong et al.

Table 6 Two-way ANOVA to analyze the effect of dark pool cross-probability and market order
proportion on RVEP(0.2)
ANOVA table
Source of variation SS df MS F P -value F crit
Cross-probability 0.51 9 0.057 0.78 0.63 1.88
Market order 5.48 5 1.10 15.2 <0.01 2.22
Interaction 3.20 45 0.07 0.99 0.50 1.37
Within 211 2940 0.07
Total 221 2999
SS: sum of the squared errors; df: degree of freedom; MS: mean squared error

Table 7 Linear regression Regression statistics

(1) (Objective variable:
RVEP. Explanation variables: Multiple R 0.92
dark pool usage, R square 0.84
cross-probability, market Adjusted R square 0.84
order proportion, volatility) Standard error 0.02
Observation 6000

Table 8 Linear regression (2) (Objective variable: RVEP. Explanation variables: dark pool usage,
cross-probability, market order proportion, volatility. SE: standard error)
Coefficients SE t Stat P -value Lower 95% Upper 95%
Intercept 0.73 0.003 247 0 0.72 0.73
DP usage −0.24 0.003 −68 <0.01 −0.24 −0.23
Cross-prob. 0.001 0.0012 0.88 0.38 −0.001 0.003
Market order 0.3 0.006 50 <0.01 0.29 0.32
Volatility 1.33 0.12 11 <0.01 1.10 1.57

Table 6 illustrates that the value of RVEP is affected by the proportion of market
orders but is not affected by the cross-probability of the dark pool.
Next, we analyzed how RVEP changes according to these factors. Taking RVEP
as the objective variable, we conducted a linear regression to observe the effect of
dark pool usage, cross-probability, market order proportion, and volatility on RVEP.
Among these factors, market volatility is a statistic. RVEP and the market volatility
were calculated after each simulation, and the analysis investigates the relationship
between the two. The following tables present the linear regression results.
Tables 7 and 8 indicate that the market volatility and market order proportion
are important and have a positive relationship with RVEP. Moreover, the dark pool
usage exhibits a negative relationship with RVEP, which suggests that higher usage
of the dark pool leads to smaller RVEP values and tends to stabilize the market. The
dark pool cross-probability does not have a significant effect on RVEP.
Figures 5, 6, 7 plot RVEP(0.2) and RVEP(0.4) according to different dark pool
cross-probabilities, market order proportions, and market volatilities, respectively.
Dark Pool Usage and Equity Market Volatility 31

RVEP(0.2) & RVEP(0.4) wrt Cross probability

0.96
RVEP(0.2)
0.94 RVEP(0.4)

0.92

0.9
Ratio

0.88

0.86

0.84

0.82

0.8
0.14 0.3 0.46 0.64 0.8 0.96
Dark cross probability

Fig. 5 RVEP with respect to dark pool cross-probability

RVEP(0.2) & RVEP(0.4) wrt Market order proportion

1
RVEP(0.2)
RVEP(0.4)
0.95

0.9
Ratio

0.85

0.8

0.75

0.7
0.3 0.4 0.5 0.6 0.7 0.8
Market order proportion

Fig. 6 RVEP with respect to market order proportion

In Fig. 5, the results are ordered by the cross-probability of the 3000 simulations
and the RVEP is shown as a moving average (over intervals of 500). The red solid
line and blue dashed line are the RVEP values given by low and high usage of the
dark pool, respectively. Both RVEP(0.2) and RVEP(0.4) are less than 1 for all cross-
probabilities, which indicates that using the dark pool decreases market volatility.
In addition, the value of RVEP(0.4) is consistently lower than that of RVEP(0.2),
suggesting that higher usage of the dark pool makes the market more stable. Both
32 Y. Xiong et al.

RVEP(0.2) & RVEP(0.4) wrt Volatility

1
RVEP(0.2)
RVEP(0.4)
0.95

0.9
Ratio

0.85

0.8

0.75

0.7
0.62 0.80 1.04 1.44 2.10 2.69
Volatility (%)

Fig. 7 RVEP with respect to market volatility

curves fluctuate within small ranges, indicating that the dark pool cross-probability
does not affect market volatility. These fluctuations are caused by different market
order proportions.
The market tends to be stable when there is a balance between buy orders and
sell orders. In this case, the price will move up and down around the midprice. Such
order-balance guarantees the executions in the dark pool, and these executions at
the midprice of the exchange stabilize the price movements. However, the market
tends to be more volatile when there exists an extreme order-imbalance. When these
imbalanced orders are submitted to the exchange, they cause rapid price changes
and larger spreads. However, if such an imbalance occurs in the dark pool, only a
few executions are made until a new balance is formed between buy orders and sell
orders. In this sense, orders submitted to the dark pool tend to inhibit the trading
volume when the price is changing rapidly, but enhance the trading volume when
the price is relatively stable.
According to the above analysis, increased usage of the midpoint dark pool leads
to a less volatile market. Although different midpoint dark pools have different
crossing mechanisms, the results indicate that the crossing time may not have a
significant influence on market volatility.
Figure 6 shows an upward trends in RVEP with an increase in market order
proportion. Moreover, the difference between RVEP(0.2) and RVEP(0.4) decreases
as the market order proportion rises. This indicates that when the proportion of
market orders is small, higher usage of the dark pool decreases market volatility
(corresponding to lower RVEP values and a larger difference between RVEP(0.2)
and RVEP(0.4)). For larger numbers of market orders, higher usage of the dark pool
has less of an effect on decreasing the market volatility (corresponding to higher
RVEP values and a smaller difference between RVEP(0.2) and RVEP(0.4)).
Dark Pool Usage and Equity Market Volatility 33

Similar to Figs. 6, 7 reveals the relationship between RVEP and market volatility.
This graph shows that when market volatility is low, higher usage of the dark pool
has a definite suppressing effect. However, when market volatility is high, higher
usage of the dark pool causes a less noticeable suppression.
When market volatility is low, there is a balance between buy orders and sell
orders, and executions in the midpoint dark pool tend to make the market stable.
When market volatility is high, only a small number of executions will happen in the
dark pool, so this volatility-decreasing effect is weakened. The difference between
RVEP(0.2) and RVEP(0.4) decreases under high market volatility, possibly because
the higher usage of the dark pool will not lead to many more executions when the
market is very volatile and there is an extreme order-imbalance.
This volatility-decreasing effect of the dark pool is consistent with the findings
of Buti et al. [1, 2] and Mizuta et al. [10]. Buti et al. [1, 2] stated that increased
dark pool trading activity tends to be associated with lower spreads and lower return
volatilities. Similarly, Mizuta et al. [10] found that as dark pool use increases, the
markets become more stable. In contrast, Ray [15] reported that following dark pool
transactions, the bid–ask spreads tend to widen and price impacts tend to increase.
This may be because the CN in his sample was relatively small: the average dark
pool usage in Ray’s sample was 1.5%, whereas we considered a usage of 20%
(which is closer to the present market rate of 15%). In another model, Zhu [20]
suggested that a higher dark pool market share is associated with wider spreads and
higher price impacts. This difference may be because Zhu’s conclusion is based
on the assumption that all traders are fully rational. Our model, however, used
zero-intelligence agents to investigate the relationship between dark pool usage and
market volatility, and did not consider the impact of specific trading strategies.

6 Conclusion

We used an agent-based model to analyze the relationship between dark pool

usage and market volatility. We focused on one typical type of dark pool in which
customer orders are matched at the midpoint of the exchange’s bid and ask prices.
Simulation results showed that the use of this midpoint dark pool decreases market
volatility. This volatility-suppressing effect becomes stronger when the usage of the
dark pool is higher, when the proportion of market orders is lower, and when market
volatility is lower. However, changes in the cross-probability of the dark pool were
not found to have any effects.

References

1. Buti, S., Rindi, B., & Werner, I. M. (2011). Dark Pool Trading Strategies. Charles A Dice
Center Working Paper (2010-6).
34 Y. Xiong et al.

2. Buti S, Rindi, B., & Werner, I. M. (2011). Diving into Dark Pools. Charles A Dice Center
Working Paper (2010-10).
3. Challet, D., & Stinchcombe, R. (2001). Analyzing and modeling 1+ 1D markets. Physica A:
Statistical Mechanics and its Applications, 300(1), 285–299.
4. Cheridito, P., & Sepin, T. (2014). Optimal trade execution with a dark pool and adverse
selection. Available at SSRN 2490234.
5. Degryse, H., De Jong, F., Van Kervel, V. (2011). The impact of dark and visible fragmentation
on market quality. SSRN eLibrary.
6. Farmer, J. D., & Foley, D. (2009). The economy needs agent-based modelling. Nature,
460(7256), 685–686.
7. Farmer, J. D., Patelli, P., & Zovko, I. I. (2005). The predictive power of zero intelligence in
financial markets. Proceedings of the National Academy of Sciences of the United States of
America, 102(6), 2254–2259.
8. Gode, D. K., & Sunder, S. (1993). Allocative efficiency of markets with zero-intelligence
traders: Market as a partial substitute for individual rationality. Journal of Political Economy,
101, 119–137.
9. Maslov, S. (2000). Simple model of a limit order-driven market. Physica A: Statistical
Mechanics and Its Applications, 278(3), 571–578.
10. Mizuta, T., Matsumoto, W., Kosugi, S., Izumi, K., Kusumoto, T., & Yoshimura, S. (2014).
Do dark pools stabilize markets and reduce market impacts? Investigations using multi-
agent simulations. In 2104 IEEE Conference on Computational Intelligence for Financial
Engineering & Economics (CIFEr) (pp. 71–76). New York: IEEE.
11. Mo, S. Y. K., Paddrik, M., & Yang, S. Y. (2013). A study of dark pool trading using an
agent-based model. In 2013 IEEE Conference on Computational Intelligence for Financial
Engineering & Economics (CIFEr) (pp. 19–26). New York: IEEE.
12. Nimalendran, M., & Ray, S. (2011), Informed trading in dark pools. SSRN eLibrary.
13. Preece, R., & Rosov, S. (2014). Dark trading and equity market quality. Financial Analysts
Journal, 70(6), 33–48.
14. Preis, T., Golke, S., Paul, W., & Schneider, J. J. (2006). Multi-agent-based order book model
of financial markets. Europhysics Letters, 75(3), 510.
15. Ray, S. (2010). A match in the dark: understanding crossing network liquidity. Available at
SSRN 1535331.
16. Ready, M. J. (2010). Determinants of volume in dark pools. Working Paper, University of
Wisconsin-Madison.
17. Satchell, S., & Knight, J. (2011). Forecasting volatility in the financial markets. Oxford:
Butterworth-Heinemann.
18. Securities and Exchange Commission. (2010). Concept release on equity market structure.
Federal Register, 75(13), 3594–3614.
19. Ye, M. (2011). A glimpse into the dark: Price formation, transaction cost and market share
of the crossing network. Transaction Cost and Market Share of the Crossing Network June 9
(2011).
20. Zhu, H. (2013). Do dark pools harm price discovery? Review of Financial Studies. 27(3), 747–
789.
Modelling Complex Financial Markets
Using Real-Time Human–Agent Trading
Experiments

John Cartlidge and Dave Cliff

Abstract To understand the impact of high-frequency trading (HFT) systems on

financial-market dynamics, a series of controlled real-time experiments involving
humans and automated trading agents were performed. These experiments fall at
the interdisciplinary boundary between the more traditional fields of behavioural
economics (human-only experiments) and agent-based computational economics
(agent-only simulations). Experimental results demonstrate that: (a) faster financial
trading agents can reduce market efficiency—a worrying result given the race
towards zero-latency (ever faster trading) observed in real markets; and (b) faster
agents can lead to market fragmentation, such that markets transition from a regime
where humans and agents freely interact to a regime where agents are more likely
to trade between themselves—a result that has also been observed in real financial
markets. It is also shown that (c) realism in experimental design can significantly
alter market dynamics—suggesting that, if we want to understand complexity in
real financial markets, it is finally time to move away from the simple experimental
economics models first introduced in the 1960s.

Keywords Agent-based computational economics · Automated trading ·

Continuous double auction · Experimental economics · High-frequency trading ·
Human–agent experiments · Robot phase transition · Trading agents

1 Introduction

In recent years, the financial markets have undergone a rapid and profound transfor-
mation from a highly regulated human-centred system to a less-regulated and more
fragmented computerised system containing a mixture of humans and automated
trading systems (ATS)—computerised systems that automatically select and execute

J. Cartlidge () · D. Cliff

Department of Computer Science, University of Bristol, Bristol, UK
e-mail: john.cartlidge@bristol.ac.uk; dc@cs.bris.ac.uk

© Springer Nature Switzerland AG 2018 35

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_3
36 J. Cartlidge and D. Cliff

a trade with no human guidance or interference. For hundreds of years, financial

trading was conducted by humans, for humans, via face-to-face (or latterly tele-
phone) interactions. Today, the vast majority of trades are executed electronically
and anonymously at computerised trading venues where human traders and ATS
interact. Homogeneous human-only markets have become heterogeneous human-
ATS markets, with recent estimates suggesting that ATS now initiate between 30%
and 70% of all trades in the major US and European equity markets [25].
As computerisation has altered the structure of financial markets, so too the
dynamics (and systemic risk) have changed. In particular, trading velocity (the
number of trades that occur in unit time) has dramatically increased [25]; stocks and
other instruments exhibit rapid price fluctuations (fractures) over subsecond time-
intervals [36]; and widespread system crashes occur at astonishingly high speed.
Most infamously, the flash crash of 6th May 2010 saw the Dow Jones Industrial
Average (DJIA) plunge around 7% ($1 trillion) in 5 min, before recovering most of
the fall over the following 20 min [21, 37]. Alarmingly, during the crash, some major
company stocks (e.g., Accenture) fell to just one cent, while others (e.g., Hewlett-
Packard) increased in value to over $100,000. These dynamics were unprecedented,
but are not unique. Although unwanted, flash crashes are now an accepted feature
of modern financial markets.1
To accurately model financial systems, it is now no longer sufficient to consider
human traders only; it is also necessary to model ATS. To this end, we take a
bottom-up, agent-based experimental economics approach to modelling financial
systems. Using purpose built financial trading platforms, we present a series of
controlled real-time experiments between human traders and automated trading
agents, designed to observe and understand the impact of ATS on market dynamics.
Conducted at the University of Bristol, UK, these experiments fall at the interdisci-
plinary boundary between the more traditional fields of experimental economics (all
human participants) and agent-based computational economics (all agent simulation
models) and offer a new insight into the effects that agent strategy, agent speed,
human experience, and experiment design have on the dynamics of heterogeneous
human–agent markets.
Results demonstrate that: (a) the speed of financial agents has an impact on
market efficiency—in particular, it is shown that faster financial trading agents
can lead to less efficient markets, a worrying result given the race towards zero-
latency (ever faster trading) observed in real markets; (b) faster agents can lead to
market fragmentation, such that markets transition from a regime where humans
and agents freely interact to a regime where agents are more likely to trade between
themselves—a result that has also been observed in real financial markets; (c)
experiment design, such as discrete-time (where participants strictly act in turns)
versus real-time systems (where participants can act simultaneously and at any

1 Flash crashes are now so commonplace that during the writing of this chapter, a flash crash
occurred in the FX rate of the British Pound (GBP). On 7 Oct 2016, GBP experienced a 6% drop
in 2 min, before recovering most of the losses [53]—a typical flash crash characteristic.
Modelling Financial Markets Using Human–Agent Experiments 37

time), can dramatically affect results, leading to the conclusion that, where possible,
a more realistic experiment design should be chosen—a result that suggests it is
finally time to move away from Vernon Smith’s traditional discrete time models of
experimental economics, first introduced in the 1960s.
This chapter is organised as follows. Section 2 introduces the argument that
financial systems are inherently complex ecosystems that are best modelled using
agent-based approaches rather than neoclassical economic models. Understanding
the causes and consequences of transient non-linear dynamics—e.g., fractures and
flash crashes that exacerbate systemic risk for the entire global financial system—
provides the primary motivation for this research. In Sect. 3, the agent-based exper-
imental economics approach—i.e., human–agent financial trading experiments—is
introduced and contextualised with a chronological literature review. Section 4
introduces the trading platform used for experiments, and details experiment
design and configuration. Empirical results presented in Sect. 5 demonstrate market
fragmentation—a significantly higher proportion of agent-only and human-only
trading in markets containing super-humanly fast agents. Since the experimental
market we construct is too constrained to exhibit fractures directly, in Sect. 6 we
interpret this result as proxy evidence for the robot phase transition associated with
fractures in real markets. In Sect. 7, conclusions are drawn, and some avenues for
future research are outlined.

2 Motivation

There exists a fundamental problem facing financial-market regulators—current

understanding of the dynamics of financial systems is woefully inadequate; there is
simply no sound theoretical way of knowing what the systemic effect of a structural
change will be [7]. Therefore, when policy makers introduce new market regulation,
they are effectively trial and error testing in the live markets. This is a concerning
state of affairs that has negative ramifications for us all, and it provides adequate
motivation for the research presented here.
In this section, it is argued that our lack of understanding is a symptom
of the dominant neoclassical economic paradigms of rational expectations and
oversimplified equilibrium models. However, a solution is proposed. It has been
compellingly argued elsewhere that economic systems are best considered through
the paradigm of complexity [41]. Agent-based models—dynamic systems of het-
erogeneous interacting agents—present a way to model the financial economy as a
complex system [22] and can naturally be extended to incorporate human (as living
agent) interactions. In addition, the converging traditions of behavioural [38] and
experimental [48] economics can address non-rational human behaviours such as
38 J. Cartlidge and D. Cliff

overconfidence and fear using controlled laboratory experiments. Here, we present

a hybrid approach—mixed human–agent financial trading experiments—that we
believe offers a path to enlightenment.

2.1 Complex Economic Systems

Neoclassical economics relies on assumptions such as market efficiency, simple

equilibrium, agent rationality, and Adam Smith’s invisible hand. These concepts
have become so ingrained that they tend to supersede empirical evidence, with
many economists subliminally nurturing an implicit Platonic idealism about market
behaviour that is divorced from reality. As Robert Nelson argued in his book,
Economics as Religion, it is almost “as if the marketplace has been deified” [42].
Consequently, no neoclassical framework exists to understand and mitigate wild
market dynamics such as flash crashes and fractures. It is necessary, therefore,
to develop “a more pragmatic and realistic representation of what is going on in
financial markets, and to focus on data, which should always supersede perfect
equations and aesthetic axioms” [7]. Disturbingly, despite global capitalism’s
existential reliance on well-functioning financial markets, there exist no mature
models to understand and predict issues of systemic risk [14]. Policy makers,
therefore, are essentially acting in the dark, with each new regulatory iteration
perturbing the market in unanticipated ways.
Fuelled by disillusionment with orthodox models, and a desire to address the
inadequacies of naïve policy making, there is a trend towards alternative economic
modelling paradigms: (1) Non-equilibrium economics focuses on non-equilibrium
processes that transform the economy from within, and include the related and
significantly overlapping fields of evolutionary economics (the study of processes
that transform the economy through the actions of diverse agents from experi-
ence and interactions, using an evolutionary methodology, e.g., [43]), complexity
economics (seeing the economy not as a system in equilibrium, but as one in
motion, perpetually constructing itself anew, e.g., [2]), circular and cumulative
causation (CCC) (understanding the real dynamic and self-reinforcing aspects
of economic phenomena, e.g., [5]), and network effects and cascading effects
(modelling the economy as a network of entities connected by inter-relationships)
[3, 6, 20, 46]; (2) Agent-based models potentially present a way to model the
financial economy as a complex system, while taking human adaptation and learning
into account [7, 22, 23, 41]; (3) Behavioural economics addresses the effects of
social, cognitive, and emotional factors on the economic decisions of individuals
[3, 38]; (4) Experimental economics is the application of experimental methods to
study economic questions. Data collected in experiments are used to test the validity
of economic theories, quantify the effects, and illuminate market mechanisms [48].
Modelling Financial Markets Using Human–Agent Experiments 39

Following this movement away from traditional economic models, in the

research presented here, markets are modelled using an approach that straddles
the interdisciplinary boundary between experimental economics and agent-based
computational economics. Controlled real-time experiments between human traders
and automated financial trading agents (henceforth, referred to simply as agents, or
alternatively as robots) provide a novel perspective on real-world markets, through
which we can hope to better understand, and better regulate for, their complex
dynamics.

2.2 Broken Markets: Flash Crashes and Subsecond Fractures

As algorithmic trading has become common over the past decade, automated
trading systems (ATS) have been developed with truly super-human performance,
assimilating and processing huge quantities of data, making trading decisions, and
executing them, on subsecond timescales. This has enabled what is known as
high-frequency trading (HFT), where ATS take positions in the market (e.g., by
buying a block of shares) for a very short period of perhaps 1 or 2 s or less, before
reversing the position (e.g., selling the block of shares); each such transaction may
generate relatively small profit measured in cents, but by doing this constantly
and repeatedly throughout the day, steady streams of significant profit can be
generated. For accounts of recent technology developments in the financial markets,
see [1, 29, 40, 44].
In February 2012, Johnson et al. [35] published a working paper—later revised
for publication in Nature Scientific Reports [36]—that immediately received
widespread media attention, including coverage in New Scientist [26], Wired [39],
and Financial News [45]. Having analysed millisecond-by-millisecond stock-price
movements over a 5 year period between 2006 and 2011, Johnson et al. argued that
there is evidence for a step-change or phase transition in the behaviour of financial
markets at the subsecond timescale. At the point of this transition—approximately
equal to human response times—the market dynamics switch from a domain where
humans and automated robot (i.e., agent) trading systems freely interact with one
another to a domain newly identified by Johnson et al. in which humans cannot
participate and where all transactions result from robots interacting only among
themselves, with no human traders involved.2 Here, we refer to this abrupt system-
wide transition from mixed human-algorithm phase to a new all-algorithm phase,
the robot phase transition (RPT).
At subsecond timescales, below the robot transition, the robot-only market
exhibits fractures—ultrafast extreme events (UEEs) in Johnson et al.’s parlance,

2 The primary reason for no human involvement on these timescales is not because of granularity in

decision making—i.e., limitations in human abilities to process information, e.g., [12]—but rather
that humans are simply too slow to react to events happening, quite literally, in the blink of an eye.
40 J. Cartlidge and D. Cliff

akin to mini flash crashes—that are undesirable, little understood, and intriguingly
appear to be linked to longer-term instability of the market as a whole. In Johnson et
al.’s words, “[w]e find 18,520 crashes and spikes with durations less than 1500 ms
in our dataset. . . We define a crash (or spike) as an occurrence of the stock price
ticking down (or up) at least ten times before ticking up (or down) and the price
change exceeding 0.8% of the initial price. . . Their rapid subsecond speed and
recovery. . . suggests [UEEs are] unlikely to be driven by exogenous news arrival”
[36].
In other words, while fractures are relatively rare events at human time scales—
those above the RPT—at time scales below the RPT, fractures are commonplace,
occurring many thousands of times over a 5 year period (equivalent to more than
ten per day when averaged uniformly). This is interesting. The price discovery
mechanism of markets is generally assumed to be driven by the actions of buyers
and sellers acting on external information, or news. For instance, the announcement
of poor quarterly profits, a new takeover bid, or civil unrest in an oil producing
region will each affect the sentiment of buyers and sellers, leading to a shift in price
of financial instruments. The prevalence of ATS means that markets can now absorb
new information rapidly, so it is not unusual for prices to shift within (milli)seconds
of a news announcement. However, fractures are characterised by a shift in price
followed by an immediate recovery, or inverse shift (e.g., a spike from $100 to
$101; returning to $100). To be driven by news, therefore, fractures would require
multiple news stories to be announced in quick succession, with opposing sentiment
(positive/negative) of roughly equal net weighting. The speed and frequency of
fractures makes this highly unlikely. Therefore, fractures must be driven by an
endogenous process resulting from the interaction dynamics of traders in the market.
Since fractures tend to occur only below the RPT, when trading is dominated by
robots, it is reasonable to conclude that they are a direct result of the interaction
dynamics of HFT robot strategies.
What Johnson et al. have identified is a phase transition in the behaviour of
markets in the temporal domain caused by fragmentation of market participants—
i.e., at time scales below the RPT, the only active market participants are HFT
robots, and the interactions between these robots directly result in fractures that
are not observed over longer time scales above the RPT. Intriguingly, however,
Johnson et al. also observe a correlation between the frequency of fractures and
global instability of markets over much longer time scales. This suggests that there
may be a causal link between subsecond fractures and market crashes. “[Further,
data] suggests that there may indeed be a degree of causality between propagating
cascades of UEEs and subsequent global instability, despite the huge difference in
their respective timescales . . . [Analysis] demonstrates a coupling between extreme
market behaviours below the human response time and slower global instabilities
above it, and shows how machine and human worlds can become entwined
across timescales from milliseconds to months . . . Our findings are consistent with
an emerging ecology of competitive machines featuring ‘crowds’ of predatory
algorithms, and highlight the need for a new scientific theory of subsecond financial
phenomena” [36].
Modelling Financial Markets Using Human–Agent Experiments 41

This discovery has the potential for significant impact in the global financial
markets. If short-term micro-effects (fractures) can indeed give some indication of
longer-term macro-scale behaviour (e.g., market crashes), then it is perhaps possible
that new methods for monitoring the stability of markets could be developed—e.g.,
using fractures as early-warning systems for impending market crashes. Further, if
we can better understand the causes of fractures and develop methods to avoid their
occurrence, then long-term market instability will also be reduced. This provides
motivation for our research. To understand fractures, the first step is to model the
RPT.
Here, we report on using a complementary approach to the historical data
analysis employed by Johnson et al. [35, 36]. We conduct laboratory-style exper-
iments where human traders interact with algorithmic trading agents (i.e., robots)
in minimal experimental models of electronic financial markets using Marco De
Luca’s OpEx artificial financial exchange (for technical platform details, see [19,
pp. 26–33]). Our aim is to see whether correlates of the two regimes suggested by
Johnson et al. can occur under controlled laboratory conditions—i.e., we attempt
to synthesise the RPT, such that we hope to observe the market transition from a
regime of mixed human–robot trading to a regime of robot-only trading.

3 Background

Experimental human-only markets have a rich history dating back to Vernon Smith’s
seminal 1960s research [48]. “Before Smith’s experiments, it was widely believed
that the competitive predictions of supply/demand intersections required very
large numbers of well-informed traders. Smith showed that competitive efficient
outcomes could be observed with surprisingly small numbers of traders, each with
no direct knowledge of the others’ costs or values” [32]. This was a significant
finding, and it has spawned the entire field of experimental economics; whereby
markets are studied by allowing the market equilibration process to emerge from
the interacting population of actors (humans and/or agents), rather than assuming
an ideal market that is trading at the theoretical equilibrium. By measuring the
distance between the experimental equilibrium and the theoretical equilibrium,
one can quantify the performance of the market. Further, by altering the rules of
interaction (the market mechanism) and varying the market participants (human or
agent), one can begin to understand and quantify the relative effects of each. This is
a powerful approach and it is one that we adopt for our experimental research.3
The following sections present a detailed background. Section 3.1 introduces
the continuous double auction mechanism used for experiments; Sect. 3.2 provides
metrics for evaluating the performance of markets; and Sect. 3.3 presents a review
of previous human–agent experimental studies.

3 For a more thorough background and literature review, refer to [19, pp. 6–25].
42 J. Cartlidge and D. Cliff

3.1 The Continuous Double Auction

An auction is a mechanism whereby sellers and buyers come together and agree on
a transaction price. Many auction mechanisms exist, each governed by a different
set of rules. In this chapter, we focus on the Continuous Double Auction (CDA),
the most widely used auction mechanism and the one used to control all the
world’s major financial exchanges. The CDA enables buyers and sellers to freely
and independently exchange quotes at any time. Transactions occur when a seller
accepts a buyer’s bid (an offer to buy), or when a buyer accepts a seller’s ask
(an offer to sell). Although it is possible for any seller to accept any buyer’s bid,
and vice-versa, it is in both of their interests to get the best deal possible at any
point in time. Thus, transactions execute with a counter party that offers the most
competitive quote.
Vernon Smith explored the dynamics of CDA markets in a series of Nobel
Prize winning experiments using small groups of human participants [47]. Splitting
participants evenly into a group of buyers and a group of sellers, Smith handed
out a single card (an assignment) to each buyer and seller with a single limit price
written on each, known only to that individual. The limit price on the card for
buyers (sellers) represented the maximum (minimum) price they were willing to pay
(accept) for a fictitious commodity. Participants were given strict instructions to not
bid (ask) a price higher (lower) than that shown on their card, and were encouraged
to bid lower (ask higher) than this price, regarding any difference between the price
on the card and the price achieved in the market as profit.
Experiments were split into a number of trading days, each typically lasting a few
minutes. At any point during the trading day, a buyer or seller could raise their hand
and announce a quote. When a seller and a buyer agreed on a quote, a transaction
was made. At the end of each trading day, all stock (sellers assignment cards) and
money (buyer assignment cards) was recalled, and then reallocated anew at the start
of the next trading day. By controlling the limit prices allocated to participants,
Smith was able to control the market’s supply and demand schedules. Smith found
that, typically after a couple of trading days, human traders achieved very close to
100% allocative efficiency, a measure of the percentage of profit in relation to the
maximum theoretical profit available (see Sect. 3.2). This was a significant result:
few people had believed that a very small number of inexperienced, self-interested
participants could effectively self-equilibrate.

3.2 Measuring Market Performance

An ideal market can be perfectly described by the aggregate quantity supplied by

sellers and the aggregate quantity demanded by buyers at every price point (i.e., the
market’s supply and demand schedules; see Fig. 1). As prices increase, in general
there is a tendency for supply to increase, with increased potential revenues from
Modelling Financial Markets Using Human–Agent Experiments 43

Fig. 1 Supply and demand curves (here illustrated as straight lines) show the quantities supplied
by sellers and demanded by buyers at every price point. In general, as price increases, the quantity
supplied increases and the quantity demanded falls. The point at which the two curves intersect is
the theoretical equilibrium point, where Q0 is the equilibrium quantity and P0 is the equilibrium
price

sales encouraging more sellers to enter the market; while, at the same time, there is
a tendency for demand to decrease as buyers look to spend their money elsewhere.
At some price point, the quantity demanded will equal the quantity supplied. This
is the theoretical market equilibrium. An idealised theoretical market has a market
equilibrium price and quantity (P0 , Q0 ) determined by the intersection between the
supply and demand schedules. The dynamics of competition in the market will tend
to drive transactions towards this partial equilibrium point.4 For all prices above P0 ,
supply will exceed demand, forcing suppliers to reduce their prices to make a trade;
whereas for all prices below P0 , demand exceeds supply, forcing buyers to increase
their price to make a trade. Any quantity demanded or supplied below Q0 is called
intra-marginal; all quantity demanded or supplied in excess of Q0 is called extra-
marginal. In an ideal market, all intra-marginal units and no extra-marginal units are
expected to trade.
In the real world, markets are not ideal. They will always trade away from
equilibrium at least some of the time. We can use metrics to calculate the
performance of a market by how far from ideal equilibrium it trades. In this chapter,
we make use of the following metrics:
Smith’s Alpha
Following Vernon Smith [47], we measure the equilibration (equilibrium-finding)
behaviour of markets using the coefficient of convergence, α, defined as the root
mean square difference between each of n transaction prices, pi (for i = 1 . . . n)
over some period, and the P0 value for that period, expressed as a percentage of the
equilibrium price:

4 The micro-economic supply and demand model presented only considers a single commodity,
ceteris paribus, and is therefore a partial equilibrium model. The market is considered indepen-
dently from other markets, so this is not a general equilibrium model.
44 J. Cartlidge and D. Cliff


 n
100  
α= 1 (pi − P0 )2 (1)
P0 n
i=1

In essence, α captures the standard deviation of trade prices about the theoretical
equilibrium. A low value of α is desirable, indicating trading close to P0 .
Allocative Efficiency
For each trader, i, the maximum theoretical profit available, πi∗ , is the difference
between the price they are prepared to pay (their limit price) and the theoretical
market equilibrium price, P0 . Efficiency, E, is used to calculate the performance of
a group of n traders as the mean ratio of realised profit, πi , to theoretical profit, πi∗ :

1  πi
n
E= (2)
n πi∗
i=1

As profit values cannot go below zero (traders in these experiments are not allowed
to enter into loss-making deals, although that constraint can easily be relaxed), a
value of 1.0 indicates that the group has earned the maximum theoretical profit
available, πi∗ , on all trades. A value below 1.0 indicates that some opportunities
have been missed. Finally, a value above 1.0 means that additional profit has been
made by taking advantage of a trading counterparty’s willingness to trade away from
P0 . So, for example, a group of sellers might record an allocative efficiency of, say,
1.2 if their counterparties (a group of buyers) consistently enter into transactions at
prices greater than P0 ; in such a situation, the buyers’ allocative efficiency would
not be more than 0.8.
Profit Dispersion
Profit dispersion is a measure of the extent to which the profit/utility generated by a
group of traders in the market differs from the profit that would be expected of them
if all transactions took place at the equilibrium price, P0 . For a group of n traders,
profit dispersion is calculated as the root mean square difference between the profits
achieved, πi , by each trader, i, and the maximum theoretical profit available, πi∗ :

 n
1 
πdisp = (πi − πi∗ )2 (3)
n
i=1

Low values of πdisp indicate that traders are extracting actual profits close to
profits available when all trades take place at the equilibrium price P0 . In contrast,
higher values of πdisp indicate that traders’ profits differ from those expected at
equilibrium. Since zero-sum effects between buyers and sellers do not mask profit
dispersion, this statistic is attractive [28].
Modelling Financial Markets Using Human–Agent Experiments 45

Delta Profit
Delta profit is used to calculate the difference in profit maximising performance
between two groups, x and y, as a percentage difference relative to the mean profit
of the two groups, πx , πy :

2(πx − πy )
ΔP (x − y) = (4)
πx + πy

Delta profit directly measures the difference in profit gained by two groups. In
a perfect market, we expect ΔP (x − y) = 0, with both groups trading at the
equilibrium price P0 . A positive (negative) value indicates that group x secures
more (less) profit than group y. Using this measure enables us to determine which,
if either, of the two groups competitively outperforms the other.

3.3 Human vs. Agent Experimental Economics

In 1993, after three decades of human-only experimental economics, a landmark

paper involving a mix of traditional human experimental economics and software-
agent market studies was published in the Journal of Political Economy by Gode
and Sunder (G&S) [28]. G&S were interested in understanding how much of the
efficiency of the CDA is due to the intelligence of traders, and how much is due
to the organisation of the market. To test this, G&S introduced a very simple Zero
Intelligence Constrained (ZIC) trading agent that generate random bid or ask prices
drawn from a uniform distribution, subject to the constraint that prices generated
cannot be loss-making—i.e., sell prices are equal or above limit price, buy prices are
equal or below limit price. G&S performed a series of ZIC-human experiments, with
results demonstrating that the simple ZIC agents produced convergence towards the
theoretical equilibrium and had human-like scores for allocative efficiency (Eq. (2));
suggesting that market convergence towards theoretical equilibrium is an emergent
property of the CDA market mechanism and not the intelligence of the traders.
Indeed, G&S found that the only way to differentiate the performance of humans
and ZIC traders was by using their profit dispersion statistics (Eq. (3)). These results
were striking and attracted considerable attention.
In 1997, Dave Cliff [13] presented the first detailed mathematical analysis and
replication of G&S’s results. Results demonstrated that the ability of ZIC traders
to converge on equilibrium was dependent on the shape of the market’s demand
and supply curves. In particular, ZIC traders were unable to equilibrate when
acting in markets with demand and supply curves very different to those used
by G&S. To address this issue, Cliff developed the Zero Intelligence Plus (ZIP)
trading algorithm. Rather than issuing randomly generated bid and ask prices in
the manner of ZIC, Cliff’s ZIP agents contain an internal profit margin from which
bid and ask prices are calculated. When a buyer (seller) sees transactions happen
at a price below (above) the trader’s current bid (ask) price, profit margin is raised,
46 J. Cartlidge and D. Cliff

thus resulting in a lower (higher) bid (ask) price. Conversely, a buyer’s (seller’s)
profit margin is lowered when order and transaction prices indicate that the buyer
(seller) will need to raise (lower) bid (ask) price in order to transact [13, p.43].
The size of ZIP’s profit margin update is determined using a well-established
machine learning mechanism (derived from the Widrow–Hoff Delta rule [56]).
Cliff’s autonomous and adaptive ZIP agents were shown to display human-like
efficiency and equilibration behaviours in all markets, irrespective of the shape of
demand and supply.
Around the same time that ZIP was introduced, economists Steve Gjerstad
and his former PhD supervisor John Dickhaut independently developed a trading
algorithm that was later named GD after the inventors [27]. Using observed market
activity—frequencies of bids, asks, accepted bids, and accepted asks—resulting in
the most recent L transactions (where L = 5 in the original study), GD traders
calculate a private, subjective “belief” of the probability that a counterparty will
accept each quote price. The belief function is extended over all prices by applying
cubic-spline interpolation between observed prices (although it has previously
been suggested that using any smooth interpolation method is likely to suffice
[19, p.17]). To trade, GD quotes a price to buy or sell that maximises expected
surplus, calculated as price multiplied by the belief function’s probability of a
quote being accepted at that price. Simulated markets containing GD agents were
shown to converge to the competitive equilibrium price and allocation in a fashion
that closely resembled human equilibration in symmetric markets, but with greater
efficiency than human traders achieved [27]. A modified GD (MGD) algorithm,
where the belief function of bid (ask) prices below (above) the previous lowest
(highest) transaction price was set to probability zero, was later introduced to
counter unwanted price volatility.
In 2001, a series of experiments were performed to compare ZIP and MGD in
real-time heterogeneous markets [52]. MGD was shown to outperform ZIP. Also in
2001, the first ever human–agent experiments—with MGD and ZIP competing in
the same market as human traders—were performed by Das et al., a team from IBM
[15]. Results had two major conclusions: (a) firstly, mixed human–agent markets
were off-equilibrium—somehow the mixture of humans and agents in the market
reduces the ability of the CDA to equilibrate; (b) secondly, in all experiments
reported, the efficiency scores of humans were lower than the efficiency scores
of agents (both MGD and ZIP). In Das et al.’s own words, “. . . the successful
demonstration of machine superiority in the CDA and other common auctions could
have a much more direct and powerful impact—one that might be measured in
billions of dollars annually” [15]. This result, demonstrating for the first time in
human-algorithmic markets that agents can outperform humans, implied a future
financial-market system where ATS replace humans at the point of execution.
Despite the growing industry in ATS in real financial markets, in academia there
was a surprising lack of further human–agent market experiments over the following
decade. In 2003 and 2006, Grossklags & Schmidt [30, 31] performed human–
agent market experiments to study the effect that human behaviours are altered
by their knowledge of whether or not agent traders are present in the market. In
Modelling Financial Markets Using Human–Agent Experiments 47

2011, De Luca & Cliff successfully replicated Das et al.’s results, demonstrating
that GDX (an extension of MGD, see [51]) outperforms ZIP in agent–agent and
agent–human markets [17]. They further showed that Adaptive Aggressive (AA)
agents—a trading agent developed by Vytelingum in 2006 that is loosely based on
ZIP, with significant novel extensions including short-term and long-term adaptive
components [54, 55]—dominate GDX and ZIP, outperforming both in agent–agent
and agent–human markets [18]. This work confirmed AA as the dominant trading-
agent algorithm. (For a detailed review of how ZIP and AA have been modified
over time, see [49, 50].) More recent human–agent experiments have focused
on emotional arousal level of humans, monitoring heart rate over time [57] and
monitoring human emotions via EEG brain data [8].
Complementary research comparing markets containing only humans against
markets containing only agents—i.e., human-only or agent-only markets rather
than markets in which agents and humans interact—can also shed light on market
dynamics. For instance, Huber, Shubik, and Sunder (2010) compare dynamics
of three market mechanisms (sell-all, buy-all, and double auction) in markets
containing all humans against markets containing all agents. “The results suggest
that abstracting away from all institutional details does not help understand dynamic
aspects of market behaviour and that inclusion of mechanism differences into theory
may enhance our understanding of important aspects of markets and money, and
help link conventional analysis with dynamics” [33]. This research stream reinforces
the necessity of including market design in our understanding of market dynamics.
However, it does not offer the rich interactions between humans and ATS that we
observe in real markets, and that only human–agent interaction studies can offer.

4 Methodology

In this section, the experimental methodology and experimental trading platform

(OpEx) are presented. Open Exchange (OpEx) is a real-time financial-market
simulator specifically designed to enable economic trading experiments between
humans and automated trading algorithms (robots). OpEx was designed and devel-
oped by Marco De Luca between 2009 and 2010 while he was a PhD student at
the University of Bristol, and since Feb. 2012 is freely available for open-source
download from SourceForge, under the terms of the Creative Commons Public
License.5 Figure 2 shows the Lab-in-a-box hardware arranged ready for a human–
agent trading experiment. For a detailed technical description of the OpEx platform,
refer to [19, pp. 26–33].
At the start of each experiment, 6 human participants were seated at a terminal
around a rectangular table—with three buyers on one side and three sellers
opposite—and given a brief introduction and tutorial to the system (explaining

5 OpEx download available at: www.sourceforge.net/projects/open-exchange.

48 J. Cartlidge and D. Cliff

Fig. 2 The Lab-in-a-box hardware ready to run an Open Exchange (OpEx) human versus agent
trading experiment. Six small netbook computers run human trader Sales GUIs, with three buyers
(near-side) sitting opposite three sellers (far-side). Netbook clients are networked via Ethernet
cable to a network switch for buyers and a network switch for sellers, which in turn are connected
to a router. The central exchange and robots servers run on the dedicated hardware server (standing
vertically, top-left), which is also networked to the router. Finally, an Administrator laptop (top
table, centre) is used to configure and run experiments. Photograph: © J. Cartlidge, 2012

the human trading GUI illustrated in Fig. 3), during which time they were able to
make test trades among themselves while no robots were present in the market.
Participants were told that their aim during the experiment was to maximise profit
by trading client orders (assignments or alternatively named permits to distinguish
that traders will simultaneously have multiple client orders to work, whereas in the
traditional literature, a new assignment would only be received once the previous
assignment had been completed) that arrive over time. For further details on the
experimental method, refer to [9, pp. 9–11].
Trading Agents (Robots)
Agent-robots are independent software processes running on the multi-core hard-
ware server that also hosts the central exchange server. Since agents can act at any
time—there is no central controller coordinating when, or in which order, an agent
can act—and since the trading logic of agents does not explicitly include temporal
information, in order to stop agents from issuing a rapid stream of quotes, a sleep
timer is introduced into the agent architecture. After each action, or decision to not
act, an agent will sleep for ts milliseconds before waking and deciding upon the
next action. We name this the sleep-wake cycle of agents. For instance, if ts = 100,
the sleep-wake cycle is 0.1 s. To ensure agents do not miss important events during
sleep, agents are also set to wake (i.e., sleep is interrupted) when a new assignment
permit is received and/or when an agent is notified about a new trade execution. The
parameter ts is used to configure the “speed” of agents for each experiment.
Modelling Financial Markets Using Human–Agent Experiments 49

Fig. 3 Trading GUI for a human buyer. New order assignments (or permits) arrive over time in
the Client Orders panel (top-left) and listed in descending order by potential profit. Assignments
are selected by double-clicking. This opens a New Order dialogue pop-up (top-centre) where bid
price and quantity are set before entering the new bid into the market by pressing button BUY.
The market Order Book is displayed top-right, with all bids and asks displayed. Bid orders that the
trader currently has live in the market are listed in the Orders panel (middle), and can be amended
from here by double-clicking. When an order executes it is removed from the orders panel and
listed in the Trades history panel (bottom). For further GUI screen shots, refer to [9, Appendix C]

Trading agents are configured to use the Adaptive Aggressive (AA) strategy
logic [54, 55], previously shown to be the dominant trading agent in the literature
(see Sect. 3.3). AA agents have short-term and long-term adaptive components.
In the short term, agents use learning parameters β1 and λ to adapt their order
aggressiveness. Over a longer time frame, agents use the moving average of
the previous N market transactions and a learning parameter β2 to estimate the
market equilibrium price, p̂0 . The aggressiveness of AA represents the tendency to
accept lower profit for a greater chance of transacting. To achieve this, an agent
with high (low) aggression will submit orders better (worse) than the estimated
equilibrium price p̂0 . For example, a buyer (seller) with high aggression and
estimated equilibrium value p̂0 = 100 will submit bids (asks) with price p > 100
(price p < 100). Aggressiveness of buyers (sellers) increases when transaction
prices are higher (lower) than p̂0 , and decreases when transaction prices are lower
(higher) than p̂0 . The Widrow–Hoff mechanism [56] is used by AA to update
aggressiveness in a similar way that it is used by ZIP to update profit margin (see
Sect. 3.3). For all experiments reported here, we set parameter values β1 = 0.5,
λ = 0.05, N = 30, and β2 = 0.5. The convergence rate of bids/asks to transaction
price is set to η = 3.0.
50 J. Cartlidge and D. Cliff

Table 1 Permit schedule for market efficiency experiments.

1 2a 3 4 5 6
Buyer 1 350 (0) 250 (4) 220 (7) 190 (09) 150 (14) 140 (16)
Buyer 2 340 (1) 270 (3) 210 (8) 180 (10) 170 (12) 130 (17)
Buyer 3 330 (2) 260 (4) 230 (6) 170 (11) 160 (13) 150 (15)
Seller 1 50 (0) 150 (4) 180 (7) 210 (09) 250 (14) 260 (16)
Seller 2 60 (1) 130 (3) 190 (8) 220 (10) 230 (12) 270 (17)
Seller 3 70 (2) 140 (4) 170 (6) 230 (11) 240 (13) 250 (15)
Six permit types are issued to each market participant, depending on their role. For each role (e.g.,
Buyer 1), there are two traders: one human (Human Buyer 1) and one robot (Robot Buyer 1). Thus,
there are 12 traders in the market. Permit values show limit price—the maximum value at which
to buy, or minimum value at which to sell—and the time-step they are issued (in parentheses).
The length of each time-step is 10 s, making one full permit cycle duration 170 s. During a 20-min
experiment there are seven full cycles
a Type 2 permits were accidentally issued to Buyer1/Seller1 at time-step 4 rather than time-step 5

Exploring the Effects of Agent Speed on Market Efficiency: April–June 2011

All experiments were run at the University of Bristol between April and July 2011
using postgraduate students in non-financial but analytical subjects (i.e., students
with skills suitable for a professional career in finance, but with no specific trading
knowledge or experience). Participants were paid £20 for participating and a further
£40 bonus for making the most profit, and £20 bonus for making the second
highest profit. Moving away from the artificial constraint of regular simultaneous
replenishments of currency and stock historically used, assignment permits were
issued at regular intervals. AA agents had varying sleep-wake cycle: ts = 100,
and ts = 10,000. We respectively label these agents AA-0.1 to signify a sleep-
wake cycle of 0.1 s, and AA-10 to signify a sleep-wake cycle of 10 s. A total of 7
experiments were performed, using the assignment permit schedules presented in
Table 1. The supply and demand curves generated by these permits are shown in
Fig. 4. We can see that for all experiments, P0 = 200 and Q0 = 126. Since each
human only participates in one experiment, and since trading agents are reset at the
beginning of each run, traders have no opportunity to learn the fixed value of P0
over repeated runs. For further details of experimental procedure, see [11].
Exploring the Robot Phase Transition (RPT): March 2012
Twenty-four experiments were run on 21st March 2012, at Park House Business
Centre, Park Street, Bristol, UK. Participants were selected on a first-come basis
from the group of students that responded to adverts broadcast to two groups: (1)
students enrolled in final year undergraduate and postgraduate module in computer
science that includes coverage of the design of automated trading agents; (2)
members of the Bristol Investment Society, a body of students interested in pursuing
a career in finance. We assume that these students have the knowledge and skills to
embark on a career as a trader in a financial institution. Volunteers were paid £25
for participating, and the two participants making the greatest profit received an iPad
valued at £400. To reduce the total number of participants required, each group were
Modelling Financial Markets Using Human–Agent Experiments 51

400

350

300 Supply

250
Price

P0
200

150

100 Demand
50
Q0
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Quantity

Fig. 4 Stepped supply and demand curves for permit schedule defined in Table 1. Curves show
the aggregate quantity that participants are prepared to buy (demand) and sell (supply) at every
price point. The point at which the two curves intersect is the theoretical equilibrium point for the
market: P0 = 200 is the equilibrium price, and Q0 is the equilibrium quantity. As there are two
traders in each role—one human and one robot—each permit cycle Q0 = 2 × 9 = 18, and over the
seven permit cycles of one full experiment, Q0 = 18 × 7 = 126. The market is symmetric about
P0

used in a session of six separate experiments. Therefore, 24 experiments were run

using the 24 participants. Between experiments, human participants rotated seats,
so each played every role exactly once during the session of 6 experiments. Human
roles were purposely mixed between experiment rounds to reduce the opportunity
for collusion and counteract any bias in market role. Once again, agents used the
AA algorithm with varying sleep-wake cycle, and assignment orders were released
into the market at regular intervals.
Table 2 presents the assignment permit schedules used for each experiment, and
the full supply and demand curves generated by these permits are plotted in Fig. 5.
At each price point—i.e., at each step in the permit schedule—two assignment
permits are sent simultaneously to a human trader and to a robot trader, once every
replenishment cycle. For all experiments, permits are allocated in pairs symmetric
about P0 such that the equilibrium is not altered, and the inter-arrival time of permits
is 4 s. Cycles last 72 s and are repeated eight times during a 10 min experiment.
Therefore, over a full experiment there are 2 × 8 = 16 permits issued at each
price point. The expected equilibrium number of trades for the market, Q0 , is
144 intra-marginal units. Each experiment, P0 is varied in the range 209–272
to stop humans from learning the equilibration properties of the market between
experiments. Agents are reset each time and have no access to data from previous
experiments. In cyclical markets, permits are allocated in strict sequence that is
unaltered between cycles. In random markets, the permit sequence across the entire
run is randomised. For further details on experimental procedure, see [9, 10].
52 J. Cartlidge and D. Cliff

Table 2 Permit schedule for RPT experiments

1 2 3 4 5 6
Buyer 1 77 (1) 27 (4) 12 (7) −9 (10) −14 (13) −29 (16)
Buyer 2 73 (2) 35 (5) 8 (8) −5 (11) −22 (14) −25 (17)
Buyer 3 69 (3) 31 (6) 16 (9) −1 (12) −18 (15) −33 (18)
Seller 1 −77 (1) −27 (4) −12 (7) 9 (10) 14 (13) 29 (16)
Seller 2 −73 (2) −35 (5) −8 (8) 5 (11) 22 (14) 25 (17)
Seller 3 −69 (3) −31 (6) −16 (9) 1 (12) 18 (15) 33 (18)
Six permit types are issued to each market participant, depending on their role. For each role,
there is one human and one robot participant. Permit values show limitprice − P0 . Thus, e.g., if
P0 = 100, a permit of type 4 to Buyer1 would have a limit price of 91. For buyers, limit prices
are the maximum value to bid; and for sellers, limit prices are the minimum value to ask. Numbers
in brackets show the time-step sequence in which permits are allocated. Thus, after 11 time-steps,
Buyer2 and Seller2 each receive a permit of type 4. For all experiments, the inter-arrival time-
step between permits is 4 s. Permits are always allocated in pairs, symmetric about P0 . In cyclical
markets, the sequence is repeated eight times: the last permits are issued to Buyer3 and Seller3 at
time 576 s, and the experiment ends 24 s later. In non-cyclical or “random” markets, the time-step
of permits is randomised across the run. Participants receive the same set of permits in both cyclical
and random markets, but in a different order

Fig. 5 Stepped supply and demand curves for an entire run of the RPT experiments, defined by
the permit schedules shown in Table 2. Curves show the aggregate quantity that participants are
prepared to buy (demand) and sell (supply) at every price point. The point at which the two curves
intersect is the theoretical equilibrium point for the market: Q0 = 144 is the equilibrium quantity,
and P0 is the equilibrium price. Each experiment the value of P0 is varied in the range 209–272 to
avoid humans learning a fixed value of P0 over repeated trials. The market is symmetric about P0 .
Modelling Financial Markets Using Human–Agent Experiments 53

5 Results

Here, we present empirical results from the two sets of experiments: (a) exploring
the robot phase transition, performed in March 2012, and (b) exploring the effects of
agent speed on market efficiency, performed in April–June 2011. Throughout this
section, for detecting significant differences in location between two samples we
use the nonparametric Robust Rank-Order (RRO) test and critical values reported
by Feltovich [24]. RRO is particularly useful for small sample statistics of the kind
we present here, and is less sensitive to changes in distributional assumptions than
the more commonly known Wilcoxon–Mann–Whitney test [24].

5.1 Exploring the Robot Phase Transition: March 2012

Experiments were run using AA agents with sleep-wake cycle times (in seconds)
ts = 0.1 (AA-0.1), ts = 1 (AA-1), ts = 5 (AA-5), and ts = 10 (AA-10). Of the
24 runs, one experienced partial system failure, so results were omitted. Runs with
agent sleep time 5 s (AA-5) are also omitted from analysis where no significant
effects are found. For further detail of results, see [9, 10].

5.1.1 Market Data

OpEx records time-stamped data for every exchange event. This produces rich
datasets containing every quote (orders submitted to the exchange) and trade (orders
that execute in the exchange) in a market. In total, we gathered 4 h of trading data
across the four one-hour sessions, but for brevity we explore only a small set of
indicative results here; however, for completeness, further datasets are presented in
[9, Appendix A]. Figure 6 plots time series of quotes and trades for a cyclical market
containing AA-0.1 agents. The dotted horizontal line represents the theoretical
market equilibrium, P0 , and vertical dotted lines indicate the start of each new
permit replenishment cycle (every 72 s). We see the majority of trading activity
(denoted by filled markers) is largely clustered in the first half of each permit
replenishment cycle; this correlates with the phase in which intra-marginal units
are allocated and trades are easiest to execute. After the initial exploratory period,
execution prices tend towards P0 in subsequent cycles. In the initial period, robots
(diamonds for sellers; inverted triangle for buyers) explore the space of prices. In
subsequent periods, robots quote much closer to equilibrium. Agent quotes are
densely clustered near to the start of each period, during the phase that intra-
marginal units are allocated. In contrast, humans (squares for sellers; triangles for
buyers) tend to enter exploratory quotes throughout the market’s open period.
54 J. Cartlidge and D. Cliff

Fig. 6 Time series of quote and trade prices from a cyclical market containing AA-0.1 agents.
The dotted horizontal line represents the theoretical market equilibrium, P0 . Vertical dotted lines
indicate the start of each new permit replenishment cycle

5.1.2 Smith’s α

We can see the equilibration behaviour of the markets more clearly by plotting
Smith’s α for each cycle period. In Fig. 7 we see mean α (±95% confidence
interval) plotted for cyclical and random markets. Under both conditions, α follows
a similar pattern, tending to approx 1% by market close. However, in the first
period, cyclical markets produce significantly greater α than random markets (RRO,
p < 0.0005).This is due to the sequential order allocation of permits in cyclical
markets, where limit prices farthest from equilibrium are allocated first. This enables
exploratory shouts and trades to occur far from equilibrium. In comparison, in
random markets, permits are not ordered by limit price, thus making it likely that
limit prices of early orders are closer to equilibrium than they are in cyclical markets.

5.1.3 Allocative Efficiency

Tables 3 and 4 display the mean allocative efficiency of agents, humans, and the
whole market grouped by agent type and market type, respectively. Across all
groupings, E(agents) > E(humans). However, when grouped by robot type
(Table 3), the difference is only significant for AA-0.1 and AA-5 (RRO, 0.051 <
p < 0.104). When grouped by market type (Table 4), E(agents) > E(humans)
is significant in cyclical markets (RRO, 0.05 < p < 0.1), random markets (RRO,
0.05 < p < 0.1), and across all 23 runs (RRO, 0.01 < p < 0.025). These results
suggest that agents outperform humans.
Modelling Financial Markets Using Human–Agent Experiments 55

Fig. 7 Mean α (±95% confidence interval) plotted using log scale for results grouped by market
type. In cyclical markets, α values are significantly higher than in random markets during the
initial period (RRO, p < 0.0005). In subsequent periods all markets equilibrate to α < 1% with
no statistical difference between groups

Table 3 Efficiency and profit for runs grouped by robot type

Robot Type Trials E(agents) E(humans) E(market) ΔP (agents − humans)
AA-0.1 6 0.992 0.975 0.984 1.8%
AA-1 5 0.991 0.977 0.984 1.4%
AA-5 6 0.990 0.972 0.981 1.8%
AA-10 6 0.985 0.981 0.983 0.4%
All 23 0.989 0.976 0.983 1.34%
Agents achieve greater efficiency E(agents) > E(humans), and greater profit ΔP (agents −
humans) > 0, under all conditions

Table 4 Efficiency and profit for runs grouped by market type

Market type Trials E(agents) E(humans) E(market) ΔP (agents − humans)
Cyclical 12 0.991 0.978 0.985 1.32%
Random 11 0.987 0.974 0.981 1.36%
All 23 0.989 0.976 0.983 1.34%
Agents achieve greater efficiency E(agents) > E(humans), and greater profit ΔP (agents −
humans) > 0, under all conditions

In Table 3, it can be seen that as sleep time increases the efficiency of agents
decreases (column 3, top-to-bottom). Conversely, the efficiency of humans tends to
increase as sleep time increases (column 4, top-to-bottom). However, none of these
differences are statistically significant (RRO, p > 0.104). In Table 4, efficiency of
agents, humans, and the market as a whole are all higher when permit schedules
56 J. Cartlidge and D. Cliff

are issued cyclically rather than randomly, suggesting that cyclical markets lead to
greater efficiency. However, these differences are also not statistically significant
(RRO, p > 0.104). Finally, when comparing E(agents) grouped by robot type
using only data from cyclical markets (data not shown), AA-0.1 robots attain a
significantly higher efficiency than AA-1 (RRO, p = 0.05), AA-5 (RRO, p = 0.05),
and AA-10 (RRO p = 0.1), suggesting that the very fastest robots are most efficient
in cyclical markets.

5.1.4 Delta Profit

From the right-hand columns of Tables 3 and 4, it can be seen that agents achieve
greater profit than humans under all conditions, i.e., ΔP (agents − humans) > 0.
Using data across all 23 runs, the null hypothesis H0 : ΔP (agents − humans) ≤ 0
is rejected (t-test, p = 0.0137). Therefore, the profit of agents is significantly
greater than the profit of humans, i.e., agents outperform humans across all
runs. Differences in ΔP (agents − humans) between robot groupings and market
groupings are not significant (RRO, p > 0.104).

5.1.5 Profit Dispersion

Table 5 shows the profit dispersion of agents πdisp (agents), humans πdisp (humans),
and the whole market πdisp (market), for runs grouped by market type. It is clear
that varying between cyclical and random permit schedules has a significant
effect on profit dispersion, with random markets having significantly lower
profit dispersion of agents (RRO, 0.001 < p < 0.005), significantly lower profit
dispersion of humans (RRO, 0.025 < p < 0.05), and significantly lower profit
dispersion of the market as a whole (RRO, 0.005 < p < 0.01). These results
indicate that traders in random markets are extracting actual profits closer to profits
available when all trades take place at the equilibrium price, P0 ; i.e., random
markets are trading closer to equilibrium, likely due to the significant difference in
α during the initial trading period (see Sect. 5.1.2). When grouping data by robot
type (not shown), there is no significant difference in profit dispersion of agents,
humans, or markets (RRO, p > 0.104).

Table 5 Profit dispersion for runs grouped by market type

Market type Trials πdisp (agents) πdisp (humans) πdisp (market)
Cyclical 12 89.6 85.4 88.6
Random 11 50.2 57.2 55.6
All 23 70.0 71.9 72.8
Profit dispersion in random markets is significantly lower than in cyclical markets for agents
πdisp (agents), humans πdisp (humans), and the whole market πdisp (market)
Modelling Financial Markets Using Human–Agent Experiments 57

5.1.6 Execution Counterparties

Let aa denote a trade between agent buyer and agent seller, hh a trade between
human buyer and human seller, ah a trade between agent buyer and human seller,
and ha a trade between human buyer and agent seller. Then, assuming a fully
mixed market where any buyer (seller) can independently and anonymously trade
with any seller (buyer), we generate null hypothesis, H0 : the proportion of trades
with homogeneous counterparties—aa trades or hh trades—should be 50%. More
formally:

Σaa + Σhh
H0 : = 0.5
Σaa + Σhh + Σah + Σha

In Fig. 8, box-plots present the proportion of homogeneous counterparty trades

for markets grouped by robot type (AA-0.1, AA-1, and AA-10); the horizontal
dotted line represents the H0 value of 50%. It can clearly be seen that the proportion
of homogeneous counterparty trades for markets containing AA-0.1 robots is
significantly greater than 50%; and H0 is rejected (t-test, p < 0.0005). In contrast,
for markets containing AA-1 and AA-10 robots, H0 is not rejected at the 10% level
of significance. This suggests that for the fastest agents (AA-0.1) the market tends
to fragment, with humans trading with humans and robots trading with robots more
than would be expected by chance. There also appears to be an inverse relationship

Fig. 8 Box-plot showing the percentage of homogeneous counterparty executions (i.e., trades
between two humans, or between two agents). In a fully mixed market, there is an equal chance
that a counterparty will be agent or human, denoted by the horizontal dotted line, H0 . When agents
act and react at time scales equivalent to humans (i.e., when sleep time is 1 s or 10 s), counterparties
are selected randomly—i.e., there is a mixed market and H0 is not rejected (p > 0.1). However,
when agents act and react at super-human timescales (i.e., when sleep time is 0.1 s), counterparties
are more likely to be homogeneous—H0 is rejected (p < 0.0005). This result suggests that, even
under simple laboratory conditions, when agents act at super-human speeds the market fragments.
58 J. Cartlidge and D. Cliff

between robot sleep time and proportion of homogeneous counterparty trades.

RRO tests show that the proportion of homogeneous counterparty trades in AA-0.1
markets is significantly higher than AA-1 markets (p < 0.051) and AA-10 markets
(p = 0.0011); and for AA-1 markets the proportion is significantly higher than AA-
10 (p < 0.104). For full detail of RRO analysis of execution counterparties, see [9,
Appendix A.2.1].

5.2 Effect of Agent Speed on Market Efficiency:

April–June 2011

Experiments were run using AA agents with sleep-wake cycle times (in seconds)
ts = 0.1 (AA-0.1) and ts = 10 (AA-10). A total of 8 experiments were performed.
However, during one experiment, a human participant began feeling unwell and
could no longer take part, so results for this trial are omitted. Here, we present
results of the remaining 7 experiments. For further detail of results, see [11].

5.2.1 Smith’s α

Figure 9 plots mean α (±95% confidence interval) for each permit replenishment
cycle, grouped by robot type: AA-0.1 and AA-10. Under both conditions, α > 10%
in the initial period, and then equilibrates to a value α < 10%. For every period, i,

Fig. 9 Mean Smith’s α (±95% confidence interval) plotted using log scale for results grouped
by robot type. In markets containing fast AA-0.1 robots, α values are significantly higher than in
markets containing slow AA-10 robots. After the initial period, all markets equilibrate to α < 10%.
Modelling Financial Markets Using Human–Agent Experiments 59

Table 6 Efficiency and profit for runs grouped by robot type

Trials E(agents) E(humans) E(market) ΔP (agents − humans)
AA-0.1 3 0.966 0.906 0.936 3.2%
AA-10 4 0.957 0.963 0.960 −0.3%
Agents achieve greater efficiency E(agents) > E(humans), and greater profit ΔP (agents −
humans) > 0, when robots are fast AA-0.1. In contrast, humans achieve greater efficiency
E(humans) > E(agents), and greater profit ΔP (agents − humans) < 0, when robots are
slow AA-10

mean value αi is lower for markets containing AA-10 robots than it is for markets
containing AA-0.1 robots. Using RRO, this difference is significant at every period:
α1 (p < 0.029), α2 (p < 0.029), α3 (p < 0.029), α4 (p < 0.029), α5 (p < 0.114),
α6 (p < 0.057), α7 (p < 0.029). This suggests that markets with slower agents
(AA-10) are able to equilibrate better than markets with faster agents (AA-0.1).

5.2.2 Allocative Efficiency and Delta Profit

Table 6 presents mean allocative efficiency and delta profit for runs grouped by
robot type. The efficiency of agents is similar under both conditions, with no
statistical difference (RRO, p > 0.114). However, runs with slow AA-10 robots
result in significantly higher efficiency of humans (RRO, 0.114 < p < 0.029), and
significantly higher efficiency of the whole market (RRO, 0.114 < p < 0.029). In
markets containing slower AA-10 robots, humans are able to secure greater profit
than agents ΔP (agents − humans) < 0; whereas in markets containing fast AA-
0.1 robots, agents secure more profit than humans ΔP (agents − humans) > 0.
However, the difference in delta profit between the two groups is not significant
(RRO, p > 0.114).
These data provide evidence that markets containing fast AA-0.1 robots are less
efficient than markets containing slow AA-10 robots. However, this does not imply
that AA-10 outperform AA-0.1, as their efficiency shows no significant difference.
Rather, we see that humans perform more poorly when competing in markets
containing faster trader-agents, resulting in lower efficiency for the market as a
whole.

5.2.3 Profit Dispersion

Table 7 presents profit dispersion for runs grouped by robot type. In markets
containing fast AA-0.1 robots, profit dispersion is significantly higher for agents
(RRO, p < 0.114), humans (RRO, p < 0.114), and the market as a whole (RRO,
0.029 < p < 0.057). These data provide evidence that fast AA-0.1 agents result in
higher profit dispersion than slow AA-10 agents, an undesirable result.
60 J. Cartlidge and D. Cliff

Table 7 Profit dispersion for runs grouped by market type

Trials πdisp (agents) πdisp (humans) πdisp (market)
AA-0.1 3 105 236 185
AA-10 4 100 164 139
In markets with slow AA-10 robots, profit dispersion of agents πdisp (agents), humans
πdisp (humans), and the whole market πdisp (market) is significantly lower than in markets with
fast AA-0.1 robots

6 Discussion

Here we discuss results presented in the previous section. First, in Sect. 6.1 we
summarise the main results that hold across all our market experiments presented
in Sect. 5.1. Subsequently, in Sect. 6.2 we discuss potentially conflicting results
from experiments presented in Sect. 5.2. Finally, in Sect. 6.3 we discuss results that
demonstrate significant differences between cyclical and random markets.

6.1 Evidence for the Robot Phase Transition (RPT)

Results in Sect. 5.1.2 show that, across all markets, α values start relatively
high (α ≈ 10%) as traders explore the space of prices, and then quickly reduce,
with markets tending to an equilibration level of α ≈ 1%. This suggests that
the market’s price discovery is readily finding values close to P0 . Further, in
Sects. 5.1.3 and 5.1.4, agents are shown to consistently outperform humans, securing
greater allocative efficiency E(agents) > E(humans), and gaining greater profit
ΔP (agents − humans) > 0. These results demonstrate a well-functioning robot-
human market trading near equilibrium, with robots out-competing humans. This
is an interesting result, but for our purpose of exploring the RPT described by
[35, 36] it only serves as demonstrative proof that our experimental markets are
performing as we would expect. The real interest lies in whether we can observe a
phase transition between two regimes: one dominated by robot–robot interactions,
and one dominated by human–robot interactions. We seek evidence of this by
observing the proportion of homogeneous counterparties within a market; that is,
the number of trade executions that occur between a pair of humans or a pair of
robots, as a proportion of all market trades. Since traders interact anonymously via
the exchange, there can be no preferential selection of counterparties. Therefore,
every buyer (seller) has an equal opportunity to trade with every seller (buyer), as
long as both have a pending assignment. The experimental market is configured to
have an equal number of robot traders and human traders, and an equal number
of identical assignments are issued to both groups. Hence, in the limit, we should
expect 50% of trade counterparties to be homogeneous (both robot, or both human),
and 50% to be heterogeneous (one robot and one human), as traders execute with
counterparties drawn at random from the population.
Modelling Financial Markets Using Human–Agent Experiments 61

From Sect. 5.1.6, our results demonstrate that for markets containing AA-0.1
robots (with sleep-wake cycle ts = 100 ms; faster than human response time), the
proportion of homogeneous counterparties is significantly higher than we would
expect in a mixed market; whereas for markets containing robots AA-1 (ts =
1000 ms; a similar magnitude to human response time) and AA-10 (ts = 10,000 ms;
slower than human response time), the proportion of homogeneous counterparties
cannot be significantly differentiated from 50%. We present this as tentative first
evidence for a robot-phase transition in experimental markets with a boundary
between 100 ms and 1 s; although, in our experiments the effects of increasing robot
speed appear to give a progressive response rather than a step-change. However, we
feel obliged to caveat this result as non-conclusive proof until further experiments
have been run, and until our results have been independently replicated.
The careful reader may have noticed that the results presented have not demon-
strated fractures—ultrafast series of multiple sequential up-tick or down-tick trades
that cause market price to deviate rapidly from equilibrium and then just as
quickly return—phenomena that [35, 36] revealed in real market data. Since we
are constraining market participants to one role (as buyer, or seller) and strictly
controlling the flow of orders into the market and limit prices of trades, the simple
markets we have constructed do not have the capacity to demonstrate such fractures.
For this reason, we use the proportion of homogeneous counterparties as proxy
evidence for the robot phase transition.

6.2 Fast Agents and Market Efficiency

Results presented in Sect. 5.2 compare markets containing fast AA-0.1 robots to
markets containing slower AA-10 robots. It is shown that markets containing fast
AA-0.1 robots have higher α (Sect. 5.2.1), lower allocative efficiency (Sect. 5.2.2),
and higher profit dispersion (Sect. 5.2.3). Together, these facts suggest that when
agents act at super-human speeds, human performance suffers, causing an overall
reduction in the efficiency of the market. The reason for this could be, perhaps,
that the presence of very fast acting agents causes confusion in humans, resulting
in poorer efficiency. If an analogous effect occurs in real financial markets, it may
imply that high-frequency trading (HFT) can reduce market efficiency.
However, these findings largely contradict the findings presented in Sect. 5.1 and
discussed in Sect. 6.1; where market equilibration α (Sect. 5.1.2), market efficiency
(Sect. 5.1.3), and profit dispersion (Sect. 5.1.5) are shown to be unaffected by robot
speed. The reason for this disparity is primarily due to an unanticipated feature (a
bug) in the behaviour of AA agents used in the experiments of Sect. 5.2 that was
not discovered at the time (for details, see [9, pp. 25–26] and [50, p. 8]). These
AA agents included a spread jumping rule such that agents will execute against a
counterparty in the order-book if the relative spread width (the difference in price
between the highest bid and the lowest ask, divided by the mean of the highest bid
and lowest ask) is below a relative threshold of MaxSpread = 15%. This is a large,
62 J. Cartlidge and D. Cliff

unrealistic threshold, and it was reduced to MaxSpread = 1% for experiments

presented in Sect. 5.1.
It is reasonable to infer that the spread jumping behaviour of AA agents is directly
responsible for the higher α values presented in Sect. 5.2.1, compared with results
for non-spread jumping agents shown in Sect. 5.1.2.6 However, when considering
market efficiency (Sect. 5.2.2), the explanation is not quite so simple. Despite the
bug existing in the agent, efficiency for agents is largely unaffected when agent
speed is increased; whereas human efficiency drops from a level comparable with
agents when sleep-wake cycle time is 10 s to 6% lower than agents when agents
act with 0.1 s sleep-wake cycle time. Therefore, the effect of the bug on efficiency
only affects humans, and does so only when agents act at super-human speeds. We
also see a similar affect on the magnitude of profit dispersion (Sect. 5.2.3), such that
πdisp (humans) is 76% higher in AA-0.1 markets compared with AA-10 markets,
whereas πdisp (agents) is only 5% higher.
This slightly counter-intuitive result is perhaps again further evidence for the
RPT. When agents act at super-human speeds, fragmentation in the market means
that agents are more likely to trade with other agents. While agents that execute
a trade due to the spread jumping bug will lose out, the agent counterparty to
the trade will gain, thus cancelling out the negative efficiency effects for agents
overall. Human efficiency, however, is negatively affected by the resulting market
dynamics, as the market trades away from equilibrium. A similar phenomena
has also been observed in a recent pilot study performed at the University of
Nottingham Ningbo China (UNNC) in July 2016, using a different agent (ZIP)
and performed on a different experimental platform (ExPo2—details forthcoming
in future publication).7 In the pilot study, agents were allowed to submit loss-
making orders into the market (i.e., agent quote price was not constrained by
assignment limit price). Interestingly, when agents acted at human speeds (10 s
sleep-wake cycle), markets equilibrated as expected. However, when agents acted at
super-human speeds (0.1 s sleep-wake cycle), the market did not equilibrate to P0 .
This demonstrates that when agents act on human timescales, i.e., above the RPT,
the equilibration behaviour of humans can dampen idiosyncratic agent behaviour.
However, at super-human timescales (i.e., below the RPT), the cumulative effects
of agent behaviour dominate the market. Therefore, as we move from timescales
above the RPT to below the RPT, the market transitions from a more efficient to a
less efficient domain.
We see similar effects occur in real markets, for example, Knight Capital’s
fiasco, previously dubbed elsewhere as the Knightmare on Wall Street. On 1st
August 2012, Knight Capital—formerly the largest US equities trader by volume,

6 Some of the variation in α between results presented in Sects. 5.1.2 and 5.2.1 may be explained
by the different permit schedules used for the two experiments (compare Tables 1 and 2). However,
previous results from a direct comparison using an identical permit schedule to Table 2 show that
MaxSpread = 15% results in higher α than MaxSpread = 1% [9, Appendix B]. Although, a
more recent study [16] suggests the opposite result, so there is some uncertainty around this effect.
7 ExPo: the exchange portal: www.theexchangeportal.org.
Modelling Financial Markets Using Human–Agent Experiments 63

trading an average of 128,000 shares per second—started live trading their new
Retail Liquidity Provider (RLP) market making software on NYSE. Within 45 min,
RLP executed four million trades across 154 stocks, generating a pre-tax loss of
$440 million. The following day, Knight’s share price collapsed over 70%. Knight
subsequently went into administration, before being acquired by Getco, a smaller
rival, forming KCG Holdings (for further details, see [4]). It is widely accepted that
Knight’s failure was due to repurposing, and inadvertently releasing, deprecated test
code that began executing trades deliberately designed to move the market price.
In the live markets, and at high frequencies well above the RPT, this resulted in
Knight’s RLP effectively trading with itself, but at a loss on either side of the
trade. The parallel here with spread-jumping AA agents is clear; if RLP acted
at much lower frequencies, below the RPT, it is likely, perhaps, that the market
could have dampened the instability caused. Of further interest is that the market
perturbance caused by RLP percolated across a large number of stocks as other
automated trading systems reacted to the behaviour. This demonstrates how single
stock fractures below the RPT can have wider market impact over longer timescales.

6.3 Realism in Market Experiments: Artefacts or Evidence?

The cyclical-replenishment permit schedules presented in Sect. 4 approximate real-

world markets more poorly than random-replenishment permit schedules. In real
markets, demand and supply does not arrive in neat price-ordered cycles. For
that reason, where results from cyclical markets (presented in Sect. 5.1) show a
significant effect that is not also present in random markets, we interpret it as
an indication that introducing artificial constraints into experimental markets for
ease of analysis runs the risk of also introducing artefacts that, because they are
statistically significant, can be misleading.
The following relationships were all observed to be statistically significant in
cyclical markets and not statistically significant in random markets, providing
further support for the argument for realism in artificial-market experiment design,
previously advanced at length in [19]:
1. Cyclical-replenishment markets have significantly greater α in the first period
of trade (see Sect. 5.1.2). This is a direct consequence of cyclical-replenishment
allocating orders in a monotonically decreasing sequence from most profitable
to least profitable. As such, the first orders allocated into the market have limit
prices far from equilibrium. Since the market is empty, there is no mechanism
for price discovery other than trial-and-error exploration, leading to large α. In
random-replenishment markets, the initial orders entering the market are drawn
at random from the demand and supply schedules. This leads to lower bounds on
limit prices and hence lower α. Subsequently, price discovery is led by the order
book, resulting in lower α that is statistically similar in both cyclical and random
markets.
64 J. Cartlidge and D. Cliff

2. In cyclical-replenishment markets, the efficiency of AA-0.1 robots is signifi-

cantly higher than the efficiency of the other robot types (see Sect. 5.1.3). While
there is some evidence of an inverse relationship between robot sleep time and
robot efficiency across all markets, we infer that this difference is an artefact of
cyclical replenishment until further experimental trials can confirm otherwise.
3. In cyclical-replenishment markets, profit dispersion is significantly higher for
agents, humans, and the market as a whole (see Sect. 5.1.5). Since lower profit
dispersion is a desirable property of a market, this suggests that the relatively
high profit dispersion observed in previous cyclical-replenishment experiments
[11, 19] is an artefact of the experimental design.

7 Conclusion

We have presented a series of laboratory experiments between agent traders and

human traders in a controlled financial market. Results demonstrate that, despite the
simplicity of the market, when agents act on super-human timescales—i.e., when
the sleep-wake cycle of agents is 0.1 s—the market starts to fragment, such that
agents are more likely to trade with agents, and humans are more likely to trade with
humans. In contrast, when agents act on human timescales—i.e., when the sleep-
wake cycle of agents is 1 s, or above—the markets are well mixed, with agents and
humans equally likely to trade between themselves and between each other. This
transition to a fragmented market from a mixed market intriguingly appears to be
linked to market inefficiency, such that below the threshold of human reaction times
(i.e., at 0.1 s timescale) any idiosyncratic agent behaviours can adversely perturb
the market; whereas above the threshold (i.e., at timescales of 1 s and above) human
interactions help to dampen market perturbations, ensuring better equilibration and
efficiency.
This behaviour has parallels with the real financial markets, and in particular, we
present this as tantalising evidence for the robot phase transition (RPT), discovered
by Johnson et al. [35, 36]. In Johnson et al.’s words, “a remarkable new study by Cliff
and Cartlidge provides some additional support for our findings. In controlled lab
experiments, they found when machines operate on similar timescales to humans
(longer than 1 s), the ‘lab market’ exhibited an efficient phase (c.f. few extreme
price-change events in our case). By contrast, when machines operated on a
timescale faster than the human response time (100 ms) then the market exhibited
an inefficient phase (c.f. many extreme price-change events in our case)” [36].
In the final quarter of 2016, a new exchange node containing the first ever
intentional delay was introduced in the USA. To achieve a delay of 350 µs in
signal transmission, the exchange embedded a 38-mile coil of fibre optic cable.
The desired intention is to “level out highly asymmetric advantages available to
faster participants” in the market [34]. However, the impact this might have at the
system level is unknown. To address this, Johnson declares that more academic
studies need to focus on subsecond resolution data, and he identifies the work we
Modelling Financial Markets Using Human–Agent Experiments 65

have reported here as one of the few exceptions in the literature that attempts to
understand subsecond behaviours [34].
This work is presented as a demonstration of the utility of using experimental
human–agent laboratory controlled markets: (a) to better understand real-world
complex financial markets; and (b) to test novel market policies and structures
before implementing them in the real world. We hope that we are able to encourage
the wider scientific community to pursue more research endeavour using this
methodology.
Future Work
For results presented here, we used De Luca’s OpEx experimental trading software,
running on the Lab-in-a-box hardware, a self-contained wired-LAN containing
networked exchange server, netbooks for human participants, and an administrator’s
laptop. This platform is ideally suited for controlled real-time trading experiments,
but is designed for relatively small-scale, synchronous markets where participants
are physically co-located. If experiments are to be scaled up, to run for much longer
periods and to support large-scale human participation, an alternative platform
architecture is required. To this end, development began on ExPo—the Exchange
Portal—in 2011. ExPo has a Web service architecture, with humans participating
via interaction through a Web browser (see [50]). This enables users to connect to
the exchange via the Internet, and participate remotely. Immediately, ExPo negates
the requirement for specific hardware, and enables long-term and many-participant
experimentation, with users able to leave and return to a market via individual
account log-in. Currently, an updated version—ExPo2—is under development at
UNNC, in collaboration with Paul Dempster. As with OpEx and ExPo, ExPo2 will
be released open-source to encourage replication studies and engagement in the
wider scientific community.
In [8] a detailed proposal for future research studies is presented. In particular,
future work will concentrate on relaxing some experimental constraints, such as
enabling agents to trade on their own account, independent of permit schedules.
This relaxation—effectively changing the function of agents from an agency
trader (or “broker”) design to a proprietary “prop” trader design—should enable
the emergence of more realistic dynamics, such as Johnson et al.’s UEE price
swing fractures. If we are able to reproduce these dynamics in the lab, this will
provide compelling evidence for the RPT. Further, market structures and regulatory
mechanisms such as financial circuit breakers, intentional network delays, and
periodic (rather than real-time) order matching at the exchange will be tested to
understand the impact these have on market dynamics. In addition, preliminary
studies to monitor human emotional responses to market shocks, using EEG brain
data, are underway. Hopefully these studies can help us better understand how
emotional reactions can exacerbate market swings, and how regulatory mechanisms,
or trading interface designs, can be used to dampen such adverse dynamics.

Acknowledgements The experimental research presented in this chapter was conducted in 2011–
2012 at the University of Bristol, UK, in collaboration with colleagues Marco De Luca (the
66 J. Cartlidge and D. Cliff

developer of OpEx) and Charlotte Szostek. They both deserve a special thanks. Thanks also to
all the undergraduate students and summer interns (now graduated) that helped support related
work, in particular Steve Stotter and Tomas Gražys for work on the original ExPo platform.
Finally, thanks to Paul Dempster and the summer interns at UNNC for work on developing the
ExPo2 platform, and the pilot studies run during July 2016. Financial support for the studies at
Bristol was provided by EPSRC grants EP/H042644/1 and EP/F001096/1, and funding from the
UK Government Office for Science (Go-Science) Foresight Project on The Future of Computer
Trading in Financial Markets. Financial support for ExPo2 development at UNNC was provided
by FoSE summer internship funding.

References

1. Angel, J., Harris, L., & Spratt, C. (2010). Equity trading in the 21st century. Working Paper
FBE-09-10, Marshall School of Business, University of Southern California, February 2010.
Available via SSRN https://ssrn.com/abstract=1584026 Accessed 22.03.2017
2. Arthur, W. B. (2014). Complexity and the economy. Oxford: Oxford University Press.
3. Battiston, S., Farmer, J. D., Flache, A., Garlaschelli, D., Haldane, A. G., Heesterbeek, H., et al.
(2016). Complexity theory and financial regulation: Economic policy needs interdisciplinary
network analysis and behavioral modeling. Science, 351(6275), 818–819
4. Baxter, G., & Cartlidge, J. (2013). Flying by the seat of their pants: What can high frequency
trading learn from aviation? In G. Brat, E. Garcia, A. Moccia, P. Palanque, A. Pasquini, F.
J. Saez, & M. Winckler (Eds.), Proceedings of 3rd International Conference on Applied and
Theory of Automation in Command and Control System (ATACCS), Naples (pp. 64–73). New
York: ACM/IRIT Press, May 2013.
5. Berger, S. (Ed.), (2009). The foundations of non-equilibrium economics. New York: Routledge.
6. Bisias, D., Flood, M., Lo, A. W., & Valavanis, S. (2012). A survey of systemic risk analytics.
Annual Review of Financial Economics, 4, 255–296.
7. Bouchaud, J. P. (2008). Economics needs a scientific revolution. Nature, 455(7217), 1181.
8. Cartlidge, J. (2016). Towards adaptive ex ante circuit breakers in financial markets using
human-algorithmic market studies. In Proceedings of 28th International Conference on
Artificial Intelligence (ICAI), Las Vegas (pp. 77–80). CSREA Press, Athens, GA, USA. July
2016.
9. Cartlidge, J., & Cliff, D. (2012). Exploring the ‘robot phase transition’ in experimental
human-algorithmic markets. In Future of computer trading. Government Office for Science,
London, UK (October 2012) DR25. Available via GOV.UK https://www.gov.uk/government/
publications/computer-trading-robot-phase-transition-in-experimental-human-algorithmic-
markets Accessed 22.03.2017.
10. Cartlidge, J., & Cliff, D. (2013). Evidencing the robot phase transition in human-agent
experimental financial markets. In J. Filipe & A. Fred (Eds.), Proceedings of 5th International
Conference on Agents and Artificial Intelligence (ICAART), Barcelona (Vol. 1, pp. 345–352).
Setubal: SciTePress, February 2013.
11. Cartlidge, J., De Luca, M., Szostek, C., & Cliff, D. (2012). Too fast too furious: Faster financial-
market trading agents can give less efficient markets. In J. Filipe & A. Fred (Eds.), Proceedings
of 4th International Conference on Agents and Artificial Intelligent (ICAART), Vilamoura
(Vol. 2, pp. 126–135). Setubal: SciTePress, February 2012.
12. Chen, S. H., & Du, Y. R. (2015). Granularity in economic decision making: An interdisci-
plinary review. In W. Pedrycz & S. M. Chen (Eds.), Granular computing and decision-making:
Interactive and iterative approaches (pp. 47–72). Berlin: Springer (2015)
13. Cliff, D., & Bruten, J. (1997). Minimal-Intelligence Agents for Bargaining Behaviours in
Market-Based Environments. Technical Report HPL-97-91, Hewlett-Packard Labs., Bristol,
August 1997.
Modelling Financial Markets Using Human–Agent Experiments 67

14. Cliff, D., & Northrop, L. (2017). The global financial markets: An ultra-large-scale systems
perspective. In: Future of computer trading. Government Office for Science, London, UK
(September 2011) DR4. Available via GOV.UK https://www.gov.uk/government/publications/
computer-trading-global-financial-markets Accessed 22.03.2017
15. Das, R., Hanson, J., Kephart, J., & Tesauro, G. (2001) Agent-human interactions in the
continuous double auction. In Nebel, B. (Ed.), Proceedings of 17th International Conference
on Artificial Intelligence (IJCAI), Seattle (pp. 1169–1176). San Francisco: Morgan Kaufmann,
August 2001
16. De Luca, M. (2015). Why robots failed: Demonstrating the superiority of multiple-order
trading agents in experimental human-agent financial markets. In S. Loiseau, J. Filipe, B.
Duval, & J. van den Herik, (Eds.), Proceedings of 7th International Conference on Agents
and Artificial Intelligence (ICAART), Lisbon (Vol. 1, pp. 44–53). Setubal: SciTePress, January
2015.
17. De Luca, M., & Cliff, D. (2011). Agent-human interactions in the continuous double auction,
redux: Using the OpEx lab-in-a-box to explore ZIP and GDX. In J. Filipe, & A. Fred (Eds.),
Proceedings of 3rd International Conference on Agents and Artificial Intelligents (ICAART)
(Vol. 2, pp. 351–358) Setubal: SciTePress, January 2011.
18. De Luca, M., & Cliff, D. (2011). Human-agent auction interactions: Adaptive-aggressive
agents dominate. In Walsh, T. (Ed.), Proceedings of 22nd International Joint Conference on
Artificial Intelligence (IJCAI) (pp. 178–185). Menlo Park: AAAI Press, July 2011.
19. De Luca, M., Szostek, C., Cartlidge, J., & Cliff, D. (2011). Studies of interactions between
human traders and algorithmic trading systems. In: Future of Computer Trading. Government
Office for Science, London, September 2011, DR13. Available via GOV.UK https://www.
gov.uk/government/publications/computer-trading-interactions-between-human-traders-and-
algorithmic-trading-systems Accessed 22.03.17.
20. Easley, D., & Kleinberg, J. (2010). Networks, crowds, and markets: Reasoning about a highly
connected world. Cambridge: Cambridge University Press
21. Easley, D., Lopez de Prado, M., & O’Hara, M. (Winter 2011). The microstructure of the ‘flash
crash’: Flow toxicity, liquidity crashes and the probability of informed trading. Journal of
Portfolio Management, 37(2), 118–128
22. Farmer, J. D., & Foley, D. (2009). The economy needs agent-based modelling. Nature,
460(7256), 685–686
23. Farmer, J. D., & Skouras, S. (2011). An ecological perspective on the future of computer
trading. In: Future of Computer Trading. Government Office for Science, London, September
2011, DR6. Available via GOV.UK https://www.gov.uk/government/publications/computer-
trading-an-ecological-perspective Accessed 22.03.2017.
24. Feltovich, N. (2003). Nonparametric tests of differences in medians: Comparison of the
Wilcoxon-Mann-Whitney and Robust Rank-Order tests. Experimental Economics, 6, 273–297.
25. Foresight. (2012). The Future of Computer Trading in Financial Markets. Final
project report, The Government Office for Science, London, UK (October 2012).
Available via GOV.UK http://www.cftc.gov/idc/groups/public/@aboutcftc/documents/file/
tacfuturecomputertrading1012.pdf Accessed 22.03.17
26. Giles, J. (2012). Stock trading ‘fractures’ may warn of next crash. New Scientist (2852)
(February 2012). Available Online: http://www.newscientist.com/article/mg21328525.700-
stock-trading-fractures-may-warn-of-next-crash.html Accessed 22.03.17.
27. Gjerstad, S., & Dickhaut, J. (1998). Price formation in double auctions. Games and Economic
Behavior, 22(1), 1–29
28. Gode, D., & Sunder, S. (1993). Allocative efficiency of markets with zero-intelligence traders:
Markets as a partial substitute for individual rationality. Journal of Political Economy, 101(1),
119–137.
29. Gomber, P., Arndt, B., Lutat, M., & Uhle, T. (2011). High Frequency Trading. Technical report,
Goethe Universität, Frankfurt Am Main (2011). Commissioned by Deutsche Börse Group.
68 J. Cartlidge and D. Cliff

30. Grossklags, J., & Schmidt, C. (2003). Artificial software agents on thin double auction markets:
A human trader experiment. In J. Liu, B. Faltings, N. Zhong, R. Lu, & T. Nishida (Eds.),
Proceedings of IEEE/WIC Conference on Intelligent Agent and Technology (IAT), Halifax
(pp. 400–407). New York: IEEE Press.
31. Grossklags, J., & Schmidt, C. (2006). Software agents and market (in)efficiency: A human
trader experiment. IEEE Transactions on Systems, Man and Cybernetics, Part C (Applications
and Review) 36(1), 56–67.
32. Holt, C. A., & Roth, A. E. (2004). The Nash equilibrium: A perspective. Proceedings of the
National Academy of Sciences of the United States of America, 101(12), 3999–4002
33. Huber, J., Shubik, M., & Sunder, S. (2010). Three minimal market institutions with human and
algorithmic agents: Theory and experimental evidence. Games and Economic Behavior, 70(2),
403–424
34. Johnson, N. (2017). To slow or not? Challenges in subsecond networks. Science, 355(6327),
801–802.
35. Johnson, N., Zhao, G., Hunsader, E., Meng, J., Ravindar, A., Carran, S., et al. (2012).
Financial Black Swans Driven by Ultrafast Machine Ecology. Working paper published on
arXiv repository, Feb 2012.
36. Johnson, N., Zhao, G., Hunsader, E., Qi, H., Johnson, N., Meng, J., et al. (2013). Abrupt rise
of new machine ecology beyond human response time. Scientific Reports, 3(2627), 1–7 (2013)
37. Joint CFTC-SEC Advisory Committee on Emerging Regulatory Issues. (2010). Findings
Regarding the Market Events of May 6, 2010. Report, CTFC-SEC, Washington, DC, Septem-
ber 2010. Available via SEC https://www.sec.gov/news/studies/2010/marketevents-report.pdf
Accessed 22.03.2017.
38. Kahneman, D. (2011). Thinking, fast and slow. New York, NY: Farrar, Straus and Giroux.
39. Keim, B. (2012). Nanosecond trading could make markets go haywire. Wired (February 2012).
Available Online: http://www.wired.com/wiredscience/2012/02/high-speed-trading Accessed
22.03.2017.
40. Leinweber, D. (2009). Nerds on wall street. New York: Wiley.
41. May, R. M., Levin, S. A., & Sugihara, G. (2008) Complex systems: Ecology for bankers.
Nature, 451, 893–895
42. Nelson, R. H. (2001). Economics as religion: From Samuelson to Chicago and beyond.
University Park, PA: Penn State University Press.
43. Nelson, R. R., & Winter, S. G. (1982). An evolutionary theory of economic change. Harvard:
Harvard University Press.
44. Perez, E. (2011). The speed traders. New York: McGraw-Hill.
45. Price, M. (2012). New reports highlight HFT research divide. Financial News (February
2012). Available Online: https://www.fnlondon.com/articles/hft-reports-highlight-research-
divide-cornell-20120221 Accessed 22.03.2017.
46. Schweitzer, F., Fagiolo, G., Sornette, D., Vega-Redondo, F., & White, D. R. (2009). Economic
networks: what do we know and what do we need to know? Advances in Complex Systems,
12(04n05), 407–422
47. Smith, V. (1962). An experimental study of comparative market behavior. Journal of Political
Economy, 70, 111–137
48. Smith, V. (2006). Papers in experimental economics. Cambridge: Cambridge University Press.
49. Stotter, S., Cartlidge, J., & Cliff, D. (2013). Exploring assignment-adaptive (ASAD) trading
agents in financial market experiments. In J. Filipe & A. Fred (Eds.), Proceedings of 5th
International Conference on Agents and Artificial Intelligence (ICAART), Barcelona. Setubal:
SciTePress, February 2013.
50. Stotter, S., Cartlidge, J., & Cliff, D. (2014). Behavioural investigations of financial trading
agents using Exchange Portal (ExPo). In N. T. Nguyen, R. Kowalczyk, A. Fred, & F. Joaquim
(Eds.), Transactions on computational collective intelligence XVII. Lecture notes in computer
science (Vol. 8790, pp. 22–45). Berlin: Springer.
Modelling Financial Markets Using Human–Agent Experiments 69

51. Tesauro, G., & Bredin, J. (2002). Strategic sequential bidding in auctions using dynamic
programming. In C. Castelfranchi & W. L. Johnson (Eds.), Proceedings of 1st International
Conference on Autonomous Agents and Multiagent Systems (AAMAS), Bologna (pp. 591–598).
New York: ACM.
52. Tesauro, G., & Das, R. (2001). High-performance bidding agents for the continuous double
auction. In Proceedings of the ACM Conference on Electronic Commerce (EC), Tampa, FL
(pp. 206–209), October 2001.
53. Treanor, J. (2017). Pound’s flash crash ‘was amplified by inexperienced traders’. The Guardian,
January 2017. Available Online https://www.theguardian.com/business/2017/jan/13/pound-
flash-crash-traders-sterling-dollar Accessed 22.03.2017.
54. Vytelingum, P. (2006). The Structure and Behaviour of the Continuous Double Auction. PhD
thesis, School of Electronics and Computer Science, University of Southampton.
55. Vytelingum, P., Cliff, D., & Jennings, N. (2008). Strategic bidding in continuous double
auctions. Artificial Intelligence, 172, 1700–1729
56. Widrow, B., & Hoff, M. E. (1960). Adaptive switching circuits. Institute of Radio Engineers,
Western Electronic Show and Convention, Convention Record, Part 4 (pp. 96–104).
57. Zhang, S. S. (2013). High Frequency Trading in Financial Markets. PhD thesis, Karlsruher
Institut für Technologie (KIT).
Does High-Frequency Trading Matter?

Chia-Hsuan Yeh and Chun-Yi Yang

Abstract Over the past few decades, financial markets have undergone remark-
able reforms as a result of developments in computer technology and changing
regulations, which have dramatically altered the structures and the properties of
financial markets. The advances in technology have largely increased the speed of
communication and trading. This has given birth to the development of algorithmic
trading (AT) and high-frequency trading (HFT). The proliferation of AT and HFT
has raised many issues regarding their impacts on the market. This paper proposes
a framework characterized by an agent-based artificial stock market where market
phenomena result from the interaction between many heterogeneous non-HFTs and
HFTs. In comparison with the existing literature on the agent-based modeling of
HFT, the traders in our model adopt a genetic programming (GP) learning algorithm.
Since they are more adaptive and heuristic, they can form quite diverse trading
strategies, rather than zero-intelligence strategies or pre-specified fundamentalist or
chartist strategies. Based on this framework, this paper examines the effects of HFT
on price discovery, market stability, volume, and allocative efficiency loss.

Keywords High-frequency trading · Agent-based modeling · Artificial stock

market · Continuous double action · Genetic programming

C.-H. Yeh ()

Department of Information Management, Yuan Ze University, Taoyuan, Chungli, Taiwan
e-mail: imcyeh@saturn.yzu.edu.tw
C.-Y. Yang
Department of Computational and Data Sciences, College of Science, Krasnow Institute for
Advanced Study, George Mason University, Fairfax, VA, USA

© Springer Nature Switzerland AG 2018 71

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_4
72 C.-H. Yeh and C.-Y. Yang

1 Introduction

Over the past few decades, financial markets around the world have undergone a
remarkable transformation as a result of developments in computer technology and
regulatory changes, which have dramatically altered the structures as well as the
properties of financial markets. For example, in 1976, an electronic system was
introduced to the New York Stock Exchange for scheduling the sequence of market
orders and limit orders. In the 1980s, the order system of the NYSE was upgraded
to the SuperDot system, which made traders capable of trading whole portfolios
of stocks simultaneously, i.e. program trading. In 1998, the US Securities and
Exchange Commission (SEC) authorized the adoption of electronic communication
networks (ECNs), which was referred to the Regulation of Alternative Trading
Systems (Reg. ATS). In such a system, traders are allowed to trade assets outside
of stock exchanges and their orders are matched automatically without the interme-
diation of dealers or brokers. In 2005, the SEC employed the Regulation National
Market System (Reg. NMS) which was dedicated to the promotion of automatic
and immediate access to intermarket quotations. In addition, the advances in
communications technology and trading greatly increased the speed of information
dissemination and aggregation. The high-speed processing and communication
of data as well as computationally intensive mathematical models became more
important for the management of financial portfolios. Therefore, the competition for
transactions became increasingly fierce and the duration of arbitrage opportunities
correspondingly shorter.
All of these institutional and informational transitions have motivated financial
professionals to develop automated and intelligent trading algorithms to execute
orders via a computer network, giving birth to the development of algorithmic
trading (AT) and high-frequency trading (HFT). AT represents the implementation
of a set of trading algorithms implemented via computers, while HFT, a subset
of AT, refers to the execution of proprietary trading strategies through employing
advanced computing technology to automatically submit and manage orders at quite
high frequencies. Nowadays, HFT accounts for a large portion of transactions in
financial markets. According to [28], over 70% of dollar trading volume in the US
capital market is transacted via HFT. There is no doubt that the development of
computer technology has persistently had a profound influence on the market at
many levels and scales, and has affected the connectivity between market partici-
pants. Therefore, the importance of HFT will monotonically increase as computer
technology unceasingly fosters innovation and change in financial markets which
consist of human traders and program traders (software agents). The emergence of
AT and HFT trading has had a significant impact on the market and has permanently
changed the market’s financial structure and other properties. It has also increased
market complexity and resulted in new market dynamics, so that it will be harder
for regulators to understand and regulate the market.
The proliferation of AT and HFT raises many issues regarding their impacts on
the market, such as price discovery, market efficiency, market stability, liquidity,
Does High-Frequency Trading Matter? 73

risk, and financial regulations. In actual fact, the role of HFT is quite controversial.
The extent to which HFT may improve or undermine the functions of financial
markets as well as its impact on market volatility and risk are still being debated.
A consensus has not yet been reached. The proponents of HFT argue that HFT
provides liquidity to the market, improves price discovery, reduces transaction costs,
and lowers spreads and market volatility. By contrast, its opponents claim that
HFT increases the systematic risk associated with both financial instability and
management, and raises rather than lowers volatility. In addition, HFT may induce
market manipulation. In this regard, the Flash Crash of 6 May 2010 that occurred in
the Dow Jones stock index is arguably attributable to HFT. Therefore, researchers
in the past few years have paid much attention to its impact from both a theoretical
and empirical perspective. They have consistently investigated this uncharted hybrid
financial ecosystem of human and program traders, in the hope of preempting any
deceitful practice and providing adaptive guidelines for effective regulations that
ensure the efficiency and stability of financial markets and the welfare of market
participants.
To understand the channels through which and the extent to which HFT may
affect the market, a number of theoretical models as well as empirical studies have
been proposed to examine these issues, e.g. [5, 9, 11, 13, 17], and [4]. Although the
literature regarding this topic has experienced precipitous growth in the past few
years, many challenges remain, as mentioned in [21]:
From a regulatory perspective, HFT presents difficult and partially unresolved questions.
The difficulties stem partly from the fact that HFT encompasses a wide range of trading
strategies, and partly from a dearth of unambiguous empirical findings about HFT’s
effects on markets. . . . Overall, the empirical research does not demonstrate that HFT has
substantial social benefits justifying its clear risks. Regulatory measures including stronger
monitoring, order cancellation taxes, and resting rules deserve more urgent attention. (p. 1)

Subject to mathematical tractability, in order to derive analytical results, theoret-

ical approaches tend to restrict their analyses to some quite simplified frameworks
where traders’ heterogeneity as well as their interactions are represented in reduced
forms. However, heterogeneity is the main driving force in financial systems, which
fuels the market dynamics with emergent properties. Without taking into account
this characteristic, the results obtained in theoretical models could be biased.
Empirical studies also suffer from serious problems in that the analyses depend
on the availability of information regarding the quotes submitted by high-frequency
traders (HFTs) as well as the trading strategies employed by them. However, such
information are mostly proprietary. Without such information, empirical studies can
only be conducted based on the inconclusive evidence regarding the behavior and
order submissions of HFTs. As mentioned in [5]:
Research in this area is limited by the fact that the contents of the algorithms that define
HF trading are closely guarded secrets, and hence unavailable to academics. Without the
possibility of looking into the black boxes, empirical research is left with the task of
studying HF trading indirectly, via observed trading behavior. Some studies work with pure
trade data, others have been given tags that distinguish trades generated by one group of
firms, labeled HFTs, from those generated by the rest of the markets. (p. 1250014-2)
74 C.-H. Yeh and C.-Y. Yang

The conclusions reached through such an indirect channel could be problematic.

Therefore, under theoretical frameworks or empirical studies, it is rather difficult to
understand and analyze how and why HFT may affect the market when isolating and
tracing the effects of a specific trading strategy in a highly interactive system may
be unrealistic. In addition, its impact may also be affected by financial regulations
such as price limit rules or tick size rules, and the trading mechanism employed to
determine transaction prices. Moreover, as [22] point out:
The systematic danger lays in the possibility of cross-excitations to other markets causing
additional herding of low latency and/or fundamental traders. We believe that a complex
systems approach to future research is crucial in the attempt to capture this inter-dependent
and out-of-equilibrium nature of financial markets. Particularly relevant in assessing HFT
risk is to understand their behaviour in conjunction with other market participants across
various asset classes. Agent-based modeling (ABM hereafter) offers a key method to
improve our understanding of the systems’ dynamics. (p. 5)

According to research funded by the U.K. government, “Foresight: The Future of

Computer Trading in Financial Markets” of the Final Project Report of the Gov-
ernment Office for Science (2012), any regulation however effective it is should be
based on solid evidence and reliable analysis. Nevertheless, the regulators encounter
two particular challenges in the investigation of HFT. First, the development and
application of new technology together with the complexity of financial trading
has advanced rapidly. Such a situation makes understanding the nature of HFT,
developing associated policies, and executing regulations more difficult. Second,
the research regarding the effects of HFT often fails to maintain the same pace as
that of the advances in technology, and we are short of data that are accessible,
comprehensive, and consistent. These restrictions call for novel approaches, in
particular a computational one like agent-based modeling (ABM). The method of
heterogeneous agent models (HAM) is a promising endeavor that applies ABM to
the study of the forces driving the dynamics in financial markets. In addition to
serving as an attempt to uncover the origin of stylized facts observed in real financial
markets, it further serves as a tool to examine the effectiveness of financial policies.
The developments in this field are still a work in progress.1
However, the study of HFT by means of ABM remains in its infancy. To the
best of our knowledge, [8] is the first paper that adopts this technique to examine
the effects of HFT. So far, papers studying this issue have been few in number.
Gsell [8] examines the impacts of AT on price formation and volatility. Two types
of traders are employed in this model. One comprises the stylized traders who are
assumed to know the fundamental value of the asset. They are simulated along with
a special combination of informed traders, momentum traders and noise traders.
The other consists of algorithmic trading traders. The preliminary results show that
large volumes traded by algorithmic traders generate an increasingly large impact
on prices, while lower latency results in lower volatility. Hanson [10] adopts an

1 In[15] and [19], they provide a rather detailed review and description regarding a number of
different types of HAMs and show the importance of HAMs to the study of financial markets.
Does High-Frequency Trading Matter? 75

agent-based model where a continuous double auction (CDA) market is employed

to investigate the impacts of a simple HFT strategy on a group of zero-intelligence
(ZI) traders. Each of them is assigned a random reservation price. The HF traders
do not learn from their experience, but act as liquidity buyers at the beginning of the
simulation. However, when they engage in trading, they become potential sellers
based on new reservation prices calculated by adding a markup to the transaction
price. The results indicate that the population size of HFTs generates significant
impacts on market liquidity, efficiency, and the overall surplus. Market volatility
increases with the number of HFTs, and the profits of HFTs may be realized
at the cost of long-term investors. Walsh et al. [24] design a hybrid experiment
consisting of human traders and algorithmic traders. These traders adopt either
fundamental strategies or technical strategies. The difference between human traders
and algorithmic traders is that the latter behave like HFTs. Their findings indicate
that liquidity increases with the share of algorithmic traders, while there exist no
significant patterns regarding volatility. In a simple model of latency arbitrage,
[23] examine how latency arbitrage may affect allocative efficiency and liquidity
in fragmented financial markets. In a two-market model with a CDA mechanism,
two types of traders, i.e. the background traders and the latency arbitrageur (LA),
buy and sell a single security. The former adopt ZI strategies and each of them is
confined to a randomly assigned market without direct access to the information
outside its own market except, subject to some degree of latency, via an information
intermediating channel of the best price quoted over the two markets, while the
latter has an advantage of direct access to the latest prices quoted in both markets,
and posts his offer whenever a latency-induced arbitrage opportunity exists. Both
types of traders have no learning abilities. The simulation results show that market
fragmentation and the actions of a latency arbitrageur reduce market efficiency and
liquidity. However, when a periodic call market is employed, the opportunities for
latency arbitrage are removed, and the market efficiency is improved. Jacob Leal et
al. [16] discuss the interplay between low-frequency traders (LFTs) who according
to their relative profitability adopt a fundamentalist strategy or a chartist strategy,
and high-frequency traders (HFTs) who use directional strategies. Their activations
are based on the extent of price deviation. The simulation results indicate that the
behavior of HFTs can bring about flash crashes. In addition, the higher that their
rate of order cancellation is, the higher the frequency, and the shorter the duration
of the crashes.
At the risk of being too general, the traders employed in these papers have
no learning ability. In addition, the traders are either characterized as ZI traders
or pre-specified as fundamentalists or chartists. However, such a design seriously
restricts, if any, the emergent market dynamics as well as properties. Therefore,
the conclusions made in these papers will possibly be limited. On the contrary,
learning is an indispensable element of economic and financial behavior since in
practice traders continuously learn from experience. Without taking into account
this element, analyses would be biased. In [7], the authors point out the importance
of learning to market properties. A market populated with ZI traders is unable to
replicate the phenomena resulting from human traders. In [16], although the LFTs
76 C.-H. Yeh and C.-Y. Yang

are endowed with learning and they can determine which type of strategy they would
like to use, the HFTs employ pre-determined trading strategies without learning.
Therefore, the purpose of this paper is to provide a framework enabling us to carry
out a systemic investigation on the effects of HFT from a broader perspective. In
our model, not only LFTs but also HFTs can learn to develop new trading strategies
by means of a genetic programming (GP) algorithm. It is this hyperdimensional
flexibility in strategy formation disciplined by survival pressure, not some ad hoc
behavior rules as in the other heterogeneous agent models, that makes GP unique
with regard to the resulting model complexity, which is way beyond what could be
generated by the stochasticity of asset fundamentals, as agents collectively expand
the strategy space while continuously explore and exploit it. The advantages of
GP have been mentioned in [18] and [19]. The representation of GP not only can
incorporate the traders’ decision-making process and adaptive behavior, but also can
open up the possibility of innovative behavior. Under this more general framework
of an artificial stock market , we can have an in-depth discussion regarding the
functions and effects of HFT.
The remainder of this paper is organized as follows. Section 2 describes the
framework of the artificial stock market, traders’ learning behavior, and the trading
mechanisms. An analysis of the simulation outcomes and results is presented in
Sect. 3. Section 4 concludes.

2 The Model

2.1 The Market Model

The market model considered in this paper is quite similar to that of [3] and [20]. It
consists of traders whose preferences have the same constant absolute risk aversion
(CARA) utility function, i.e. U (Wi,t ) = −exp(−λWi,t ), where Wi,t is trader i’s
wealth in period t, and λ is the degree of absolute risk aversion. Each trader has two
assets for investment: a risky stock and riskless money. At the end of each period, the
stock pays dividends which follow a stochastic process (Dt ) unknown to traders.2
The riskless money is perfectly elastically supplied with a constant interest rate per
annum, r. Just as in the design proposed in [12], the interest rate can be measured
over different time horizons, i.e. rd = r/K (where K stands for the number of

2 While the dividend’s stochastic process is unknown to traders, they base their derivation of
the expected utility maximization on an Gaussian assumption. Under the CARA preference, the
uncertainty of the dividends affects the process of expected utility maximization by being part of
conditional variance.
Does High-Frequency Trading Matter? 77

transaction days over 1 year.3 ). Therefore, a gross return is equal to R = 1 + rd .

Trader i’s wealth in the next period can be described as

Wi,t+1 = RWi,t + (Pt+1 + Dt+1 − RPt )hi,t , (1)

where Pt is the price (i.e. the closing price) per share of the stock and hi,t is the
number of stock shares held by trader i at time t. The second term in Eq. (1) accounts
for the excess revenue from holding the stock over period t + 1 to earn the excess
capital gain, Rt+1 = Pt+1 + Dt+1 − RPt .
Each trader myopically maximizes his one-period expected utility based on the
wealth constraint shown in Eq. (1):

λ
max{Ei,t (Wt+1 ) − Vi,t (Wt+1 )}, (2)
h 2

where Ei,t (·) and Vi,t (·) are his one-period ahead forecasts regarding the conditional
expectation and variance at t + 1, respectively, on the basis of his information set,
Ii,t , updated up to period t. Therefore, we can derive

Ei,t (Wt+1 ) = RWi,t + Ei,t (Pt+1 + Dt+1 − RPt )hi,t

= RWi,t + Ei,t (Rt+1 )hi,t , (3)
Vi,t (Wt+1 ) = h2i,t Vi,t (Pt+1 + Dt+1 − RPt ) = h2i,t Vi,t (Rt+1 ), (4)

The optimal shares of the stock, h∗i,t , for trader i at the beginning of each period are
determined by solving Eq. (2),

Ei,t (Rt+1 ) Ei,t (Pt+1 + Dt+1 ) − RPt

h∗i,t = = . (5)
λVi,t (Rt+1 ) λVi,t (Rt+1 )

The difference between the optimal position and a trader’s current stockholding is
his demand for the stock:

di,t = h∗i,t − hi,t . (6)

Therefore, a trader acts as a buyer (seller) if di,t > 0 (di,t < 0).

3 For example, K = 1, 12, 52, 250 represents the number of trading periods measured by the units
of a year, month, week and day, respectively.
78 C.-H. Yeh and C.-Y. Yang

2.2 Expectations

Based on Eq. (5), it is clear that each trader’s optimal holding of the stock depends
on his own conditional expectation and variance. The functional form for his
expectation, Ei,t (·), is described as follows4 :

(Pt + Dt )[1 + θ0 (1 + sech(β ln(1 + fi,t )))] if fi,t ≥ 0.0,
Ei,t (Pt+1 + Dt+1 ) =
(Pt + Dt )[1 + θ0 (1 − sech(β ln(−1 + fi,t )))] if fi,t < 0.0,
(7)

where θ0 and β are coefficients of expectation transformation, and fi,t is the raw
forecast value generated by the GP algorithm. Pt is the last closing price for LFTs
and the last transaction price (P T ) for HFTs.
Each trader updates his active rule’s estimated conditional variance at the end of
2 denote V (R
each period. Let σi,t i,t t+1 ). Each trader’s conditional variance is updated
with constant weights θ1 , θ2 and θ3 by:
2
σi,t = (1−θ1 −θ2 )σi,t−1
2
+θ1 (Pt +Dt −ut−1 )2 +θ2 [(Pt +Dt )−Ei,t−1 (Pt +Dt )]2 ,
(8)
where

ut = (1 − θ1 )ut−1 + θ1 (Pt + Dt ). (9)

This variance update function is a variant of those used in [20] and [12]. In [12],
they argue that part of the autocorrelation features and the power-law pattern found
in financial markets may come from the geometric decay process of the sample
mean and variance. As one of many alternatives to variance formulation away
from rational expectations, our version is sufficient for representing the concept of
bounded rationality as well as replicating several critical stylized facts.
To maximize one-period expected utility, each trader has to maximize his excess
revenue by looking for a better forecast for the sum of the price and dividend of the
next period, while at the same time minimizing his perceived conditional variance
(uncertainty) of the excess return. In our model, each trader is endowed with a
population of strategies (i.e. NI models). Which strategy (or forecasting function)
will be used in the next period crucially depends on its relative performance

4 Our framework can incorporate other types of expectation functions; however, different specifi-
cations require different information available to traders, and the choice depends on the selected
degree of model validation. Unlike [1] and [20], we do not assume that traders have the information
regarding the fundamental value of the stock, i.e. the price in the equilibrium of homogeneous
rational expectations. The functional form used in this paper extends the designs employed in
[6, 20], and [26]. It allows for the possibility of a bet on the Martingale hypothesis if fi,t = 0. For
more details about applying the GP evolution to the function formation, we recommend the readers
to refer to Appendix A in [27].
Does High-Frequency Trading Matter? 79

compared with the others up until the current period. The performance (strength)
of each model is measured by its forecast accuracy, i.e. the negative value of its
conditional variance,5

si,j,t = −σi,j,t
2
, (10)

where si,j,t is the strength of trader i’s j th model, Fi,j,t , up to period t.

The strategies of each trader evolve over time. The evolutionary process operates
on the basis of the strengths of these strategies as follows. Each trader asyn-
chronously iterates the evolutionary process to update his own strategy pool every
NEC periods (an evolutionary cycle).6 At the beginning of the first period in each
evolutionary cycle, each trader randomly selects NT out of NI models. Among
these initial candidates, a trader selects the one with the highest strength to make
his forecast, fi,t as shown in Eq. (7), in each period of this cycle. At the end of
the last period in each cycle, each trader evaluates the strength of each model
in his own strategy pool. The strategy with the lowest strength is replaced by a
new model which is generated by one of the three operators: crossover, mutation,
and immigration. The crossover operator implies the concept of combining the
information existing in the strategy pool, while the mutation or the immigration
operator represents a kind of innovation.

2.3 Trading Process

This paper adopts a simplified continuous double auction (CDA) process designed
to mimic the intraday trading process. Unlike the existing literature applying ABM
to the study of HFT in which traders are characterized as ZI or pre-specified
fundamentalists or chartists, we consider two populations of uninformed and
bounded-rational GP-learning traders, LFTs and HFTs, that are ignorant of the
information regarding the fundamental value of the stock.7 The numbers of LFTs
and HFTs are NLF and NH F , respectively. The difference between LFTs and HFTs

5 Similardesigns were used in [1] and [20].

6 Although NEC , defining the time length of the periods between the occurrences of strategy
evolution, is a parameter common to all traders, each trader has his first strategy evolution occurred
randomly at one of the first NEC period, so that traders are heterogeneous in asynchronous strategy
learning. Because we focus on the effect of the speed difference between HFTs and LFTs, rather
than the relative speed difference among HFTs nor how market dynamics drive the participants
learn toward either kind of timing capability, on the market performance, a setting with endogenous
evolution cycles though intriguing may blur, if any, the possible speed-induced effect by entangling
the characteristics of the two trade-timing types.
7 One alternative is including informed HFTs so as to examine their impacts. However, in order

to focus on the effects resulting from high-frequency trading alone, both LFTs and HFTs are
uniformly assumed to be uninformed traders.
80 C.-H. Yeh and C.-Y. Yang

is the speed of incorporating market information into their expectations as well as

trading decisions. Although some instant market information, such as the best bid
and ask, are observed for all traders, the LFTs can only form their expectations
based on the information regarding the prices and dividends of the last k periods,
i.e. {Pt−1 , . . . , Pt−k , Dt−1 , . . . , Dt−k }, and then make their trading decisions.
By contrast, HFTs are assumed to be active speculators who can gain access to
and quickly process the instant market information to form their expectations and
trading decisions. To represent such a discrepancy, in addition to the information
available to LFTs, HFTs are also kept updated on the last l transaction prices (i.e.
T , ..., PT
{P T , P−1 T T T
−l+1 }) and the last l trading volumes (i.e. {V , V−1 , . . . V−l+1 }),
the highest l bids and the lowest l asks (i.e. {Bb , Bb,−1 , . . . Bb,−l+1 ; Ba , Ba,−1 ,. . . ,
Ba,−l+1 }),8 on the order book. Therefore, their expectations and trading decisions
may differ from one another as long as a new bid or ask is posted, even if they use
the same forecasting model. For both LFTs and HFTs, the factors of the market
information are incorporated into decision making by being part of the elements
of the information set that are fed into the leaf nodes of GP trees as input. With
the information input (leaf) nodes as the variables, a decision rule wrapped in
tree-like representation of the generative GP syntax then could be decoded into a
conventional mathematical equation.
At the beginning of each period, the orders in which both types of traders are
queued to take turns to submit quotes are randomly shuffled. A period starts with m
LFTs posting their orders, and all HFTs come after these LFTs to post their offers.
Each HFT trader recalculates his expectation based on the latest instant information
before submitting his order. After all HFTs have completed their submissions, the
next m LFTs enter the market.9 A trading period is finished when the last m LFTs
and the ensuing HFTs have posted their orders; then, a new period begins.10
Each trader regardless of types, based on the best ask (Ba ), the best bid (Bb )
and his demand for the stock (di,t ), determines his trading action, i.e. accepting (or
posting) a bid or an ask, according to the following rules:
• If di,t = 0, he has no desire to make a quote.
• If di,t > 0, he decides to buy,

8P T T
(V T ) is the last transaction price (trading volume), P−l+1 T
(V−l+1 ) is the last lth transaction
price (trading volume). Bb,−1 (Ba,−1 ) refers to the current second best bid (ask), and Bb,−l+1
(Ba,−l+1 ) is the current lth best bid (ask).
9 Although the way we design the “jump-ahead” behavior of HFTs is not information-driven, it

is neutral without subject interpretation of HFT behavior taking part in the assumptions. In light
of the fact that the motives and the exact underlying triggering mechanisms of the vast variety of
HFTs are still far from transparent, it would be prudent not to overly specify how HFTs would
react to what information and thus have the results reflecting the correlation induced by ad hoc
behavioral assumptions instead of the causality purely originated from the speed difference.
10 The relative speeds of information processing for different types of traders are exogenous. One

may vary the speed setting to investigate the effect of technology innovation in terms of the
improvement in traders’ reactions to market information.
Does High-Frequency Trading Matter? 81

– If Bb exists,
If Bb > Ei,t (·), he will post a buy order at a price uniformly distributed in
(Ei,t (·) − Si , Ei,t (·)), where Si = γ σi,t
2 .

If Bb ≤ Ei,t (·), he will post a buy order at a price uniformly distributed in

(Bb , min{Bb + Si , Ei,t (·)}).
– If Bb does not exist but Ba exists,
If Ba > Ei,t (·), he will post a buy order at a price uniformly distributed in
(Ei,t (·) − Si , Ei,t (·)).
If Ba ≤ Ei,t (·), he will post a market order, and buy at Ba .
– If neither Bb nor Ba exist,
If PLT > Ei,t (·), he will post a buy order at a price uniformly distributed
in (Ei,t (·) − Si , Ei,t (·)), where PLT is the previous transaction price right
before he submits his order.
If PLT ≤ Ei,t (·), he will post a buy order at a price uniformly distributed
in (PLT , PLT + Si ).
• If di,t < 0, he decides to sell,
– If Ba exists,
If Ba < Ei,t (·), he will post a sell order at a price uniformly distributed in
(Ei,t (·), Ei,t (·) + Si ).
If Ba ≥ Ei,t (·), he will post a sell order at a price uniformly distributed in
(max{Ba − Si , Ei,t (·)}, Ba ).
– If Ba does not exists but Bb exists,
If Bb < Ei,t (·), he will post an sell order at a price uniformly distributed in
(Ei,t (·), Ei,t (·) + Si ).
If Bb ≥ Ei,t (·), he will post a market order, and sell at Bb .
– If neither Bb nor Ba exist,
If PLT < Ei,t (·), he will post a sell order at a price uniformly distributed
in (Ei,t (·), Ei,t (·) + Si ).
If PLT ≥ Ei,t (·), he will post a sell order at a price uniformly distributed
in (PLT − Si , PLT ).
A transaction occurs when a trader accepts the existing best bid or ask, or when a
pair of bid and ask cross. All unfulfilled orders that are left at the end of each period
are canceled before the next period begins. The closing price of a period is defined
as the last transaction price.
82 C.-H. Yeh and C.-Y. Yang

3 Simulation Results

To have a better understanding of the impacts of HFT, besides the benchmark market
(BM) without HFTs, we experiment with a series of markets differing in the relative
portions of LFTs and HFTs and in the speed (activation frequency) of HFTs. The
activation frequency of HFTs is measured in the way that HFTs enter the market
after m LFTs have posted their orders. A higher activation frequency of HFTs
implies a lower value of m. Three different numbers of HFTs (i.e. NH F = 5, 10,
20) and three different activation frequencies (i.e. m = 40, 20, 10) are considered.
The scenarios of these markets are summarized in Table 1.
The model is calibrated to generate several prominent stylized facts observed
in real financial markets, e.g. the unconditional returns characteristic of excess
kurtosis, fat tails and insignificant autocorrelations as well as volatility clustering.
Because our focus is on the general impact of HFTs, the calibration does not extend
to the behavior of HFTs, i.e. the explicit trading tactics, and only aims at the daily
rather than intraday stylized facts. This modeling strategy lessens the difficulty in
isolating the pure effects of HFT activities, which would be very difficult to manage
if we were to consider all possible behavioral traits of HFTs speculated in real
financial markets.11 The difficulty would be compounded as the arguable features of
HFTs may vary with the evolutions of technology, communication techniques, and
trading methods.
The parameters are presented in Table 2. The net supply of the stock is set to
be zero. In addition, traders are allowed to sell short. Figure 1 presents the basic
time series properties of a typical BM run. The top two panels of Fig. 1 are the
price and return dynamics of the stock, respectively. The last two panels show
the autocorrelation features of the raw returns and absolute returns, respectively.

Table 1 The market scenario m

NH F 40 20 10
5 A1 A2 A3
10 B1 B2 B3
20 C1 C2 C3

Table 2 Stylized facts of the K α Hr H|r|

calibrated model
Minimum 13.98 2.49 0.49 0.94
Median 19.62 4.25 0.54 0.95
Average 22.83 4.31 0.54 0.95
Maximum 47.03 6.30 0.62 0.95

11 The real characteristics of HFTs deserve further examination, in the hope of uncovering the
causality between some specific trading behavior (not necessarily exclusive to HFTs) and market
phenomena.
Does High-Frequency Trading Matter? 83

200
Price

150

100

50

0 5000 10000 15000 20000

Period

6
Return (1/100)

4
2
0
−2
−4
−6
0 5000 10000 15000 20000
Period

250
200
Prob.density

150
100
50
0
−0.06 −0.04 −0.02 0.00 0.02 0.04 0.06
Return

0.2
Autocorrelation of r

0.1

0.0

−0.1

−0.2
0 20 40 60 80 100
Lag
Autocorrelation of lrl

0.4

0.3

0.2

0.1

0.0
0 20 40 60 80 100
Lag

Fig. 1 Time series properties of the calibrated model

84 C.-H. Yeh and C.-Y. Yang

Table 3 Parameters for simulations

The stock market
h0 0
M0 $100
(r, rd ) (0.05, 0.0002)
Dt N(D, σD2 )=N(0.02, 0.004)
NP 20000
Traders
(NLF , NH F ) (200, 5), (200, 10), (200, 20)
(NT , NI ) (5, 20)
(NEC , NSC ) (3, 3)
λ 0.5
θ0 0.5
β 0.05
(θ1 , θ2 ) (0.3, 0.3)
γ 0.5
Parameters of genetic programming
Function set {if-then-else;and,or,not;
≥,≤,=,+,-,×,%,}
sin,cos,abs,sqrt}
Terminal set for the LFTs {Pt−1 ,...,Pt−5 ,Dt−1 ,..., Dt−5 ,R}
Terminal set for the HFTs {Pt−1 ,...,Pt−5 ,Dt−1 ,..., Dt−5 ,
T ,V T ,...,V T ,
P T ,...,P−4 −4
Bb ,...,Bb,−4 ,Ba ,...,Ba,−4 ,R}
Probability of immigration (PI ) 0.1
Probability of crossover (PC ) 0.7
Probability of mutation (PM ) 0.2
R denotes the ephemeral random constant

Besides the first three lags, over most lag periods the autocorrelations of the raw
returns are insignificant at the 5% significance level, while the autocorrelations
of the absolute returns are significant for most lag periods. Table 3 reports some
statistics for the returns for 20 BM runs. The second column is the excess kurtosis
(K) of the raw returns. It is clear that the values of the excess kurtosis are larger
than zero, which is an indication of fat tails. The tail phenomena are also manifested
in the tail index α proposed in [14], which is a more appropriate statistic than the
kurtosis. The index value is calculated based on the largest 5% of observations. The
smaller the value is, the fatter is the tail. The empirical values usually range from
two to five. The third column shows that the average value of the tail index is 4.31,
which is consistent with the empirical results. The last two columns are the Hurst
exponents of the raw returns and absolute returns, respectively. The Hurst exponent
is used to measure the memory of a time series. Its value ranges between 0 and
1, and is 0.5 for a random series, above which a long memory process leading to
Does High-Frequency Trading Matter? 85

persistence exists in a time series. Empirically, the Hurst exponent is about 0.5 for
raw returns, and above that for absolute returns. It is evident that the values of the
Hurst exponents of the simulated raw returns are close to 0.5, while those of the
absolute returns are about 0.95. Therefore, these return series statistics indicate that
our model is capable of replicating some salient empirical features.
In conducting the simulation based on this calibrated model, we perform twenty
runs (each lasting 20,000 periods) with different random seeds for each market
scenario. The statistical analyses of HFT are then based on the average of the 20 run
averages for the whole period. First, we analyze the effects of HFT on the market
volatility, price discovery, and trading volume. Following the design of [25], the
market volatility (PV ) and price distortion (PD ) of a simulation run are measured
by

NP   NP  
100   Pt − Pt−1 
  , PD = 100
  Pt − Pf 
 ,
PV =  P   P  (11)
NP − 1 t−1 NP f
t=1 t=1

where Pf is the fundamental price of the stock under the assumption of homoge-
neous rational expectations.

1
Pf = (D − λσD2 h) (12)
R−1

Therefore, the price distortion measures the extent to which the price deviates from
the fundamental price. In our model, Pf = 100.0. Because the number of traders
across different market scenarios is not the same, we consider the trading volume
per trader for a fair comparison between the markets.
Figure 2 presents the relationship between the market volatility and the activation
frequency of HFT participation when different numbers of HFTs are considered.
The solid (black) line describes the average volatilities of 5 HFTs, while the
dashed (blue) line and the chain (red) line represent those of 10 and 20 HFTs,
respectively. When the markets have only 5 HFTs, the volatility pattern does not
exhibit any significant difference between the markets either with or without the
HFTs, regardless of the frequency of HFT participation. Basically, this phenomenon
does not change much when the number of HFTs is 10. The volatility is slightly
higher if the HFTs intervene for every 20 or 10 LFTs. However, such a property
disappears when the number of HFTs is 20. It is evident that overall the market
volatilities are greater for those markets with HFTs. In addition, more intense HFT
activity in terms of a higher activation frequency of participation by HFTs results
in greater volatility. The results reveal that, when the number of HFTs exceeds a
threshold, the activity of HFTs gives rise to the more volatile market dynamics.
The influence of HFT on price distortion is presented in Fig. 3. Compared with
the benchmark market, under the case of the number of HFTs being 5, the presence
of HFTs results in lower price distortions. Besides, the price distortion decreases
86 C.-H. Yeh and C.-Y. Yang

Fig. 2 Market volatility 1.2 A:5 HFTs

B:10 HFTs
1.1 C:20 HFTs

1.0

0.9

Volatility
0.8

0.7

0.6

0.5

0.4

0.3
BM 40 20 10
HFT Frequency

with an increase in the activation frequency of HFT participation. When the number
of HFTs increases from 5 to 10, such a pattern remains unchanged. Therefore, our
results indicate that the existence of HFTs may help to improve price discovery.
However, the increase in the number of HFTs still generates higher levels of price
distortion, implying a subtle discrepancy between the effect of HFT population size
(the scale) and that of HFT activation frequency (the intensity). When the number
of HFTs is further raised from 10 to 20, the results change dramatically. The price
distortions of the HFT markets become much larger than those in the BM except in
the case of low HFT activation intensity where the HFTs intervene every 40 LFTs.
Moreover, the price distortion increases with an increase in the activation frequency
of HFT participation. This phenomenon indicates that, as long as the markets are
dominated by the HFT activity, the HFT will hinder the price discovery process.
Figure 4 displays the patterns of volume. The trading volume largely increases
when the HFTs exist, regardless of their population size. When only 5 HFTs exist,
the trading volume does not significantly further increase as the HFT activity
intensifies. This phenomenon is the same in the case of 10 HFTs with the activation
frequencies for HFT participation at 10 and 20. However, when the number of
HFTs further increases from 10 to 20, the trading volume exhibits a relatively clear
upward trend as the activation frequency of HFT participation increases. Therefore,
our results indicate that the activity of HFTs helps to enhance market liquidity. In
addition, the impact of the number of HFTs seems to be greater than that of the
activation frequency of HFT participation.
Does High-Frequency Trading Matter? 87

Fig. 3 Price distortion 60 A:5 HFTs

B:10 HFTs
58 C:20 HFTs
56
54
52
50

Distortion
48
46
44
42
40
38
36

BM 40 20 10
HFT Frequency

Fig. 4 Trading volume

1.0 A:5 HFTs
B:10 HFTs
C:20 HFTs
0.9

0.8
Trading Volume

0.7

0.6

0.5

0.4

BM 40 20 10
HFT Frequency
88 C.-H. Yeh and C.-Y. Yang

Fig. 5 Allocative efficiency

A:5 HFTs
loss 0.955 B:10 HFTs
C:20 HFTs

Avg. Allocative Efficiency Loss

0.950

0.945

0.940

0.935

BM 40 20 10
HFT Frequency

To understand how HFT may affect market efficiency, we adopt the measure of
the allocative efficiency loss proposed in [2]. The allocative efficiency loss for trader
i in period t is defined as:

1
Li,t = 1 − , (13)
1 + |hi,t (Pt ) − hi,t |Pt

where hi,t (Pt ) is trader i’s desired quantity of the stock at price Pt , and hi,t is
the number of stock shares held at the end of period t. Therefore, this measure is
calculated based on the difference between a trader’s realized stockholding and his
optimal position at the end of each period. The results of the allocative efficiency
loss are presented in Fig. 5. It is clear that, when the markets have only 5 HFTs,
the level of allocative efficiency loss increases accompanied by an increase in the
activation frequency of HFT participation. The growth rate of this trend is lower
when the number of HFTs is 10. Such a pattern disappears when more HFTs, such
as 20, are added to the markets. In addition, the level of allocative efficiency loss
decreases as the number of HFTs increases. This implies that, when the number of
HFTs is below some threshold, the presence of HFTs results in a higher level of
allocative efficiency loss. However, as long as the number of HFTs is large enough,
the allocative efficiency losses in the markets with HFTs display no clear trend in
terms of the deviations from the loss in the BM case. As the chain (red) line shows,
except in the case where the activation frequency of HFT participation is 40, the
magnitudes of the losses do not differ much from that for the BM.
Does High-Frequency Trading Matter? 89

4 Conclusion

This paper proposes an agent-based artificial stock market to examine the effects
of HFT on the market. In this model, both HFTs and LFTs are adaptive, and are
able to form very diverse trading strategies. The differences between LFTs and
HFTs are the speeds of order submission and information processing capacity (i.e.
the ability to incorporate up-to-the-minute order book information in expectations).
Our analysis covers the market volatility, price distortion, trading volume, and
allocative efficiency loss as a result of the interplay between the scale (the size of
the population) and the intensity (the activation frequency of the participation) of the
HFT activity. The simulation results indicate that the presence of HFT may give rise
to both positive and negative effects on the markets. HFT does help in improving
market liquidity. However, its impacts on market volatility, price distortion, and
allocative efficiency loss vary with the number of HFTs. Therefore, evaluations
of the influence of HFT ought to take into consideration both the intensity of
participation by HFTs and the number of HFTs.

Acknowledgements Research support from MOST Grant no. 103-2410-H-155-004-MY2 is

gratefully acknowledged.

References

1. Arthur, W. B., Holland, J., LeBaron, B., Palmer, R., & Tayler, P. (1997). Asset pricing under
endogenous expectations in an artificial stock market. In W. B. Arthur, S. Durlauf, & D. Lane
(Eds.), The economy as an evolving complex system II (pp. 15-44). Reading, MA: Addison-
Wesley.
2. Bottazzi, G., Dosi, G., & Rebesco, I. (2005). Institutional architectures and behavioral
ecologies in the dynamics of financial markets. Journal of Mathematical Economics, 41(1–
2), 197–228.
3. Brock, W. A., & Hommes. C. H. (1998). Heterogeneous beliefs and routes to chaos in a simple
asset pricing model. Journal of Economic Dynamics and Control, 22(8–9), 1235–1274.
4. Carrion, A. (2013). Very fast money: High-frequency trading on the NASDAQ. Journal of
Financial Markets, 16(4), 680–711.
5. Cartea, Á., & Penalva, J. (2012). Where is the value in high frequency trading? Quarterly
Journal of Finance, 2(3), 1250014-1–1250014-46.
6. Chen, S.-H., & Yeh, C.-H. (2001). Evolving traders and the business school with genetic
programming: A new architecture of the agent-based artificial stock market. Journal of
Economic Dynamics and Control, 25(3–4), 363–393.
7. Cliff, D., & Bruten, J. (1997). Zero is Not Enough: On the Lower Limit of Agent Intelligence
for Continuous Double Auction Markets. HP Technical Report, HPL-97-141.
8. Gsell, M. (2008). Assessing the Impact of Algorithmic Trading on Markets: A Simulation
Approach. Working paper.
9. Hagströmer, B., & Lars Nordén, L. (2013). The diversity of high-frequency traders. Journal of
Financial Markets, 16(4), 741–770.
10. Hanson, T. A. (2011). The Effects of High frequency Traders in a Simulated Market. Working
paper.
90 C.-H. Yeh and C.-Y. Yang

11. Hasbrouck, J., & Saar, G. (2013). Low-latency trading. Journal of Financial Markets, 15(4),
646–679.
12. He, X.-Z., & Li, Y. (2007). Power-law behaviour, heterogeneity, and trend chasing. Journal of
Economic Dynamics and Control, 31(10), 3396–3426.
13. Hendersgott, T., Jones, C. M., & Menkveld, A. J. (2011). Does algorithmic trading improve
liquidity? The Journal of Finance, 66(1), 1–33.
14. Hill, B. M. (1975). A simple general approach to inference about the tail of a distribution.
Annals of Statistics, 3(5), 1163–1174.
15. Hommes, C. H. (2006) Heterogeneous agent models in economics and finance. In L. Tesfatsion
& K. L. Judd (Eds.), Handbook of computational economics (Vol. 2, Chap. 23, pp. 1109–1186).
Amsterdam: Elsevier.
16. Jacob Leal, S., Napoletano, M., Roventini, A., & Fagiolo, G. (2014). Rock around the clock:
An agent-based model of low- and high-frequency trading. In: Paper Presented at the 20th
International Conference on Computing in Economics (CEF’2014).
17. Jarrow, R. A., & Protter, P. (2012). A dysfunctional role of high frequency trading in
electronic markets. International Journal of Theoretical and Applied Finance, 15(3), 1250022-
1–1250022-15.
18. Kirman, A. P. (2006). Heterogeneity in economics. Journal of Economic Interaction and
Coordination, 1(1), 89–117.
19. LeBaron, B. (2006). Agent-based computational finance. In L. Tesfatsion & K. L. Judd
(Eds.), Handbook of computational economics (Vol. 2, Chap. 24, pp. 1187–1233). Amsterdam:
Elsevier.
20. LeBaron, B., Arthur, W. B., & Palmer, R. (1999). Time series properties of an artificial stock
market. Journal of Economic Dynamics and Control, 23(9–10), 1487–1516.
21. Prewitt, M. (2012). High-frequency trading: should regulators do more? Michigan Telecommu-
nications and Technology Law Review, 19(1), 1–31.
22. Sornette, D., & von der Becke, S. (2011). Crashes and High Frequency Trading. Swiss Finance
Institute Research Paper No. 11–63.
23. Wah, E., & Wellman, M. P. (2013). Latency arbitrage, market fragmentation, and efficiency:
A two-market model. In Proceedings of the Fourteenth ACM Conference on Electronic
Commerce (pp. 855–872). New York: ACM.
24. Walsh, T., Xiong, B., & Chung, C. (2012). The Impact of Algorithmic Trading in a Simulated
Asset Market. Working paper.
25. Westerhoff, F. (2003). Speculative markets and the effectiveness of price limits. Journal of
Economic Dynamics and Control, 28(3), 493–508.
26. Yeh, C.-H. (2008). The effects of intelligence on price discovery and market efficiency. Journal
of Economic Behavior and Organization, 68(3-4), 613–625.
27. Yeh, C.-H., & Yang, C.-Y. (2010). Examining the effectiveness of price limits in an artificial
stock market. Journal of Economic Dynamics and Control, 34(10), 2089–2108.
28. Zhang, X. F. (2010). High-Frequency Trading, Stock Volatility, and Price Discovery. Working
paper.
Modelling Price Discovery in an Agent
Based Model for Agriculture in
Luxembourg

Sameer Rege, Tomás Navarrete Gutiérrez, Antonino Marvuglia,

Enrico Benetto, and Didier Stilmant

Abstract We build an ABM for simulation of incentives for maize to produce

bio-fuels in Luxembourg with an aim to conduct life cycle assessment of the
additional maize and the consequent displacement of other crops in Luxembourg.
This paper focuses on the discovery of market price for crops. On the supply side
we have farmers who are willing to sell their produce based on their actual incurred
costs and an expected markup over costs. On the demand side, we have buyers or
middlemen who are responsible for quoting prices and buying the output based on
their expectation of the market price and quantity. We have N buyers who participate
in the market over R rounds. Each buyer has a correct expectation of the total
number of buyers in each market. Thus in each round, the buyer bids for a quantity
Qeb
qbr = N ×R , where Qeb is the expected total output of a crop. The buyer at each
round buys min(qbr , Str ), the minimum of the planned purchase at each round r and
the total supply Str by farmers in the round at a price pbr . The market clears over
multiple rounds. At each round, the buyers are sorted by descending order of price
quotes and the highest bidder gets buying priority. Similarly the farmers are sorted
according to the ascending order of quotes. At the end of each round, the clearance
prices are visible to all agents and the agents have an option of modifying their bids
in the forthcoming rounds. The buyers and sellers may face a shortfall which is the
difference between the target sale or purchase in each round and the actual realised
sale. The shortfall is then covered by smoothing it over future rounds (1–4). The
more aggressive behaviour is to cover the entire shortfall in the next round, while a
more calm behaviour leads to smoothing over multiple (4) rounds. We find that there

S. Rege ()
ModlEcon S.ã.r.l-S, Esch-sur-Alzette, Luxembourg, Luxembourg
T. Navarrete Gutiérrez · A. Marvuglia · E. Benetto
Luxembourg Institute of Science and Technology (LIST), Belvaux, Luxembourg
e-mail: tomas.navarrete@list.lu; antonino.marvuglia@list.lu; enrico.benetto@list.lu
D. Stilmant
Centre Wallon de Recherches Agronomiques (CRA-W), Libramont, Belgium
e-mail: d.stilmant@cra.wallonie.be

© Springer Nature Switzerland AG 2018 91

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_5
92 S. Rege et al.

is a statistically distinct distribution of prices and shortfall over smoothing rounds

and has an impact on the price discovery.

Keywords Agent based models (ABM) · Agriculture · Biofuels · Price

discovery · Life cycle assessment (LCA) · Luxembourg

1 Introduction

Life cycle assessment (LCA) is a standardised methodology used to quantify the

environmental impacts of products across their whole life cycle [6]. LCA is nowa-
days an assessment methodology recognised worldwide, although with some per-
sisting limitations and certain barriers, as witnessed by the survey of [3]. It has been
used at the fine grain level (at the product level) or at a more global level (policy).
The MUSA (MUlti-agent Simulation for consequential LCA of Agro-systems)
project http://musa.tudor.lu, aims to simulate the future possible evolution of the
Luxembourgish farming system, accounting for more factors than just the economy
oriented drivers in farmers’ decision making processes. We are interested in the
behavioural aspects of the agricultural systems, including the green conscience of
the farmers. The challenges facing the farming system are manifold. Dairy and
meat production have been the financial base of the national agricultural landscape
with production of cereals and other crops as a support to the husbandry. This
sector is fraught with multiple rules and regulations including restrictions on milk
production via quotas as dictated by the Common Agricultural Policy of the EU.
There is also a complex set of subsidies in place to enable the farmers to be more
competitive. The quotas are set to disappear adding to increased pressure on the
bottom line for dairy farmers. Just as a chain is as strong as its weakest link
any model is as robust as its weakest assumption. Model building is a complex
exercise but modelling behaviour is far more complex. Statistical and optimisation
models fail to account for vagaries of human behaviour and have limited granularity.
Preceding the MUSA project, in [9] we have built a partial equilibrium model
for Luxembourg to conduct a consequential LCA of maize production for energy
purposes (dealing with an estimated additional production of 80,000 tons of maize)
using non-linear programming (NLP) and positive mathematical programming
(PMP) approaches. PMP methodology [5] has been the mainstay of modelling
methodology for agriculture models relating to cropping patterns based on economic
fundamentals. This approach converts a traditional linear programming (LP) into an
NLP problem by formulating the objective function as a non-linear cost function
to be minimised. The objective function parameters are calibrated to replicate the
base case crop outputs. This approach is useful as a macro level such as countries
or regions where one observes the entire gamut of crops planted at a regional level.
To increase the granularity to investigate the impacts of policy on size of farms
is still possible provided each class of farms based on size exhibits plantation of
all crops in the system. When the granularity increases to the farm level, the crop
rotation takes priority for the farmer and then one observes only a subset of the
Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg 93

entire list of crops in the system. It is at this level that the PMP approach fails
as the objective function is calibrated to the crops observed at a specific point in
time. This limitation is overcome by the agent based model (ABM) approach. To
investigate the possible set of outcomes due to human responses to financial and
natural challenges, ABM are a formal mechanism to validate the range of outcomes
due to behavioural differences.
In the work we report here, we investigate the different price discovery mech-
anisms based on agent behaviour to assess convergence of prices in thin markets
which are typically agriculture markets.
In Sect. 2 we present a brief tour of different agent-based models dealing with
agricultural systems as well as LCA research done using agent-based models. Then,
in Sect. 3 we describe our model. We present the initial results of the simulations in
Sect. 4 and present our conclusions in Sect. 6.

2 Literature Survey

Happe et al. [4] use an ABM called AgriPoliS, for simulation of agricultural
policies. The focus is on simulating the behaviour of farms by combining traditional
optimisation models in farm economics with agent based approach for the region
of Hohenlohe in southwest Germany. The paper classifies farms into size classes
and assumes all farms in a specific class size to have the same area. The variation
in the farms is on account of individual asset differences (both financial and
physical). Farms are classified only as grasslands or arable lands with no further
classification by crops or rotation schemes. Berger [1] is another application on the
lines of [4] with an integration of farm based linear programming models under
a cellular automata framework, applied to Chile. The model explicitly covers the
spatial dimension and its links to hydrology with a policy issue of investigating the
potential benefits of joining the Mercosur agreement. Le et al. [8] is an application of
agent based modelling framework to assess the socio-ecological impacts of land-use
policy in the Hong Ha watershed in Vietnam. The model uses in addition to farms as
agents, the landscape agents that encapsulate the land characteristics. Additional sub
models deal with farmland choice, forest choice, agricultural yield dynamics, forest
yield dynamics, agent categoriser that classifies the households into a group and
natural transition to enable transformation between vegetation types. Berta et al. [2]
is a similar study to simulate structural and land use changes in the Argentinian
pampas. The model is driven by agent behaviour that aims to match the aspiration
level of the farmer to the wealth level. The farmer will change behaviour until the
two are close. In addition to 6 equal sized plots in the farm, the farmer has a choice
of planting 2 varieties of wheat, maize and soybean. The model studies the potential
penetration of a particular crop and find that soybean is being cultivated on a much
larger scale.
Despite most of the previously cited related works include a high number of
economical aspects in their models, they lack the inclusion of aspects related to
price discovery.
94 S. Rege et al.

3 Data and Model Structure

In the model, the farmers are represented via entities called “agents” who take
decisions based on their individual profiles and on a set of decision (behavioural)
rules defined by the researchers on the basis of the observation of real world (e.g.
the results of a survey) and on the interactions with other entities.
Luxembourg provides statistics about the economy of the agricultural sector
through STATEC [11] and Service d’Economie Rural [10]. The statistics deal with
the area under crops over time, farm sizes and types of farms by output classification,
use of fertilisers, number of animals by type (bovines, pigs, poultry, horses) and use
(meat, milk). The latest year for which one obtains a consistent data set across all
the model variables was 2009.1 In 2009 there were 2242 farms with an area of
130,762 ha under cultivation that included vineyards and fruit trees and pastures
amongst other cereal and leaf crops.
Table 1 shows the distribution of farm area by size of farms in Luxembourg for
the year 2009. We also know the total area of crops planted under each farm type.
Figure 1 shows the initial proportions of each crop present in 2009. From [7] we
know the different rotation schemes for crops. Rotation schemes are used by farmers
to maintain the health of the soil and rotate cereals and leaves on the same field in
a specified manner. We randomly assign a rotation scheme to each farm and then
randomly choose a crop from cereals or leaves. As to crops that are neither cereals
nor leaves, they are a permanent cultivation such as vineyards, fruits, meadows or
pastures. Once the random allocation of crops has been made to the farms, we scale
the areas so as to match the total area for the crop under a specific farm type.

3.1 Model Calibration

In the absence of real data on field ownership by farmers and the crops planted on
those fields, one has to find proxy methods that can be used to allocate the farms
to farmers such that the allocation of land to each crop for each farm type (from
A to I) matches the data in 2009. In reality farmers plant crops on fields based

Table 1 Distribution of farms by size (ha) in Luxembourg in 2009

Farm type
A B C D E F G H I
Total <2 2–4.9 5–9.9 10–19.9 20–29.9 30–49.9 50–60.9 70–99.9 100+
Number 2242 230 165 217 186 116 246 263 398 421
Area (ha) 130,762 131 598 1533 2667 2890 9956 15,743 33,583 63,661
Average size 58.32 0.57 3.62 7.06 14.34 24.91 40.47 59.86 84.38 151.21

1 The details of the data are available in [9].

Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg 95

Fig. 1 Initial proportions of crops planted in 2009

on seasons and in many cases plant more than one crop on the same field in a
calendar year. This leads to area under crops being greater than the actual amount
of land area available for cultivation. Fortunately for the case of Luxembourg, the
winter (September/October to July/August) and summer or spring (March/April to
August/September) crops are not amenable for planting multiple times on the same
plot of land as the winter crops are still growing when the summer crops need to be
planted. Another important aspect related to the calibration is data collection by the
statistical services. In the case of Luxembourg, the annual data on the crops planted
come from a survey administered once annually. Since the number of farmers and
farms is small, it covers very well the population. Farmers involved in rearing
animals also grow crops for feed but harvest them before maturity for conversion
into silage. In such instances one can encounter multiple crops on the same plot
of land where in the winter crop is harvested before the summer season begins.
The occurrence of such instances is quite low, but there is no statistical data of this
phenomenon in the case of Luxembourg.
The objective of the calibration is to assign farm area to each of the 2242 farmers
based on the farm type they belong and assign crops to the farms such that they
match the randomly drawn rotation schemes for each farm and also match the
aggregate area under cultivation for each crop. The ideal situation would be to have
information on each farmer’s land holdings by field type and choice of a rotation
scheme so as to proceed with the actual modelling of behaviour in a more realistic
96 S. Rege et al.

and inclusive manner, but as previously mentioned, this is not the case. In absence
of this information we proceed with the calibration as follows:
1. Randomly allocate farm size to each farmer between the minimum and maximum
based on the farm type (A to I) as shown in Table 1
2. Each farmer has a rotation scheme associated with a field. We assume that the
farmer consistently uses the same rotation scheme across all the fields (owned
or rented) under cultivation. There are six rotation schemes in use, LCC, LCCC,
LCCLC, LC, LLCC, LLLLC, where L stands for leaf crops and C for cereals. So
a rotation scheme LLCC means that the field has a leaf crop, which is substituted
by another leaf crop, which is then substituted by a cereal and finally by another
cereal preferably different from the earlier cereal crop. The field then returns to
a leaf crop to begin the cycle.
A uniform random number between 1 and 6 (inclusive) is chosen to assign a
rotation scheme to each farmer. Once a scheme has been allotted to a farmer, we
assume that it remains in place till the end of the simulation over the necessary
time periods.
3. (a) We initially use the naïve assumption that the number of farms planting a
crop will be in proportion to the share of that crop’s area in the total area for
that crop’s type. So to compute the number of farms planting triticale winter
for farm type I, we take the area under triticale winter which is a cereal and
divide it by the total area under all cereals for a specific farm type (in this case
I) and use that share as a proportion of the total number of farms (here 421
as witnessed by Table 1) for that farm type (here I) to arrive at a preliminary
number of farms planting that crop.
With this procedure one could in principle land up with a very small
number of farms that are planting a specific crop. In order to overcome this
we use the following modification to ensure a substantial number of farms
planting a crop.

if selectNumberOf F arms ≤ 20 then

selectNumberOf F arms =
max(f armI D.size()/2, selectNumberOf F arms);
else
selectNumberOf F arms =
min(f armI D.size(), (int)(selectN umberOf F arms × 1.5));
end
Algorithm 1: Selecting number of farms planting crops
where f armI D.size() is the total number of farms in each farm type and
selectNumberOfFarms is the initial number of farms to plant a given crop.
(b) After choosing the number of farms, we randomly choose unique farms from
the list of farms for the specific farm type (A to I). This then completes the
list of farms planting the chosen crop and same procedure is repeated for all
crops.
(c) To allocate the area for each crop under each farm, multiply the total crop
area under each farm type by the fraction of the total farm area of selected
farms.
Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg 97

Table 2 Area (ha) under crop by farm type (A to I)

cropName Type A B C D E F G H I
wheat_winter C 0 4 25 81 124 355 684 1590 3712
wheat_summer C 0 4 27 85 129 370 715 1660 3877
Spelt C 0 0 2 5 8 22 42 97 226
rye_winter C 0 1 7 31 27 90 141 221 585
barley_winter C 0 8 39 118 200 573 792 1375 2759
barley_spring C 0 4 23 70 119 343 474 823 1651
Oats C 0 3 5 36 65 169 253 314 539
mixed_grain_winter C 0 1 2 3 3 9 15 32 59
mixed_grain_spring C 0 0 1 2 2 9 14 31 58
grain_maize L 0 2 5 9 9 31 49 106 198
triticale_winter C 0 5 29 108 124 450 580 1047 1712
other_forage_crops L 0 1 7 29 59 388 1196 3148 6515
maize_dry_matter _BG L 0 1 8 32 66 434 1339 3526 7296
dried_pulses L 0 1 4 7 6 23 37 79 148
Beans L 0 0 1 2 2 6 9 20 37
Potatoes L 0 3 15 12 5 66 37 138 328
Rapeseed L 0 0 3 11 23 153 473 1246 2577
clover_grass_mix L 2 9 24 40 39 138 221 481 896
Meadows O 1 38 90 206 196 729 1113 2445 4207
Pastures O 10 249 584 1329 1268 4714 7194 15,804 27,191
Vineyards O 106 203 564 293 47 10 1 16 60
crops_NES O 0 2 4 7 7 24 38 83 155
crops_NES are crops not specified

(d) Finally we update the area of each farm as the sum of the area under
each crop under plantation. This completes the crop allocation aspect of the
model.
We have to resort to the Algorithm 1 due to the unavailability of actual data
and the ground realities of cropping activities in the agriculture sector. The ideal
situation would have been complete information on the field size by farm, the
crop planted on each field and the rotation scheme for each of the fields. Also
the data released in public statistics is an outcome of an annual survey conducted
by the government based on a sample of farms with information reported on
a specific day in the year. In case we did not use the heuristics as outlined
in Algorithm 1 we would have had to contend with a larger than normal area
allocated to a particular crop in randomly chosen farms. To be more precise,
from Table 2, the area under cropping for winter barley for farm type C is 39
ha. There are 217 farms totaling 1533 ha for this type C. When we calibrate, and
have a very small number of farms, the average area allocated to winter barley
per farm is larger than should be, even though the calibration in terms of area and
crop allocation is correct, the distribution is biased in favour of larger farm sizes.
This implicitly increases the risk associated with a particular crop for the farmer
and has potential implications for the crop rotation scheme. The larger the area
98 S. Rege et al.

devoted to each crop, the greater the swings in cropping patterns and hence larger
volatility of output. This volatility is only on account of calibration. In order to
mitigate this problem, we artificially increase the number of farms planting the
crops, from the naïve value. The other option would have been to allocate every
single crop to every farm but that again would bias the estimates to smaller sizes.
in absence of information, the heuristic mentioned above was the best possible
and the approach is flexible enough to cover various degrees of farm sizes.
4. Table 3 shows the average data on yield (t/ha), price (e /t), output (t), various
costs in e /ha for seeds, crop protection, miscellaneous others, rentals, machine,
labour, area and building and data on the kg/ha of fertiliser of N, P and K.
To initialise the data across all farms we take these numbers as the mean for all
farms and use a standard deviation of 15% to allocate the different heads across
all farms. The cost of fertilisers is expected to be the same for all farmers, while
the change is on the amount applied.
The summation of all costs leads to the cost of production for each crop
planted for each farmer. The markup over cost is a random number between 5
and 15% and each farmer is assigned a random markup over cost. Sorting the
farmers’ supply price in an ascending order generates the supply curve for each
crop.

3.2 Price Discovery in Rounds

We have N buyers who participate in the market over R rounds. Each buyer has a
correct expectation of the total number of buyers in each market. In each round, the
Qeb
buyer bids for a quantity qbr = N ×R , where Qeb is the expected total output of a crop.
The buyer at each round buys min(qbr , Str ), the minimum of the planned purchase
at each round r and the total supply Str by farmers in the round at a price pbr . The
market clears over multiple rounds, following the six steps described next.
1. Initialise buyer and seller (farmer) rounds
2. Buyer b enters the quantity (qbrb ) and price (pbrb ) for quantity qb in buyer round
rb . For example, say for buyer b = 5, in buyer round rb = 3 out of a maximum
buyer rounds 5, qbrb = 20, where 20 tons is qb in round rb = 3
3. The market maker sorts the buyer bids in descending order with the highest
bidder getting the right to purchase first.
4. Seller (farmer) s, enters the quantity (qsrs ) and price (psrs ) for quantity qs in round
rs . For example, say for farmer (seller) s = 1234, in seller round rs = 2 out of a
maximum seller rounds 5, qsrs = 3, where 3 tons is qs in round rs = 2
5. The market maker sorts the seller bids in ascending order with the lowest priced
seller getting the chance to sell first.
6. Once the buyers and sellers are sorted, the first buyer gets a chance to purchase
the quantity desired. Two things can happen for the buyer
(a) The entire quantity bid is available and the market clears at price pbrb
Table 3 Crop details
cropName T S E Season Yield Price Output Seed Protection Other Rentals Machine Labour Area Building N P K
wheat C 10 8 Winter 6.6557 145.74 43,761 48.44 41.52 55.36 55.36 55.36 55.36 13.84 13.84 147.9 11.5 11.3
_winter
wheat C 2 8 Summer 6.6187 105.76 45,451 51.44 30.86 41.15 41.15 41.15 61.73 10.29 10.29 147.9 11.5 11.3
_summer
Spelt C 10 8 Winter 4.6425 208.94 1866 69.20 69.20 69.20 69.20 69.20 69.20 13.84 13.84 147.9 11.5 11.3
rye_winter C 10 8 Winter 6.2888 80.30 6937 35.11 35.11 35.11 35.11 35.11 35.11 7.02 7.02 103.1 9.5 8.5
barley C 10 7 Winter 6.1477 87.02 36,050 34.86 34.86 34.86 34.86 34.86 34.86 6.97 6.97 134.5 11.5 9.0
_winter
barley C 3 8 Summer 5.2335 90.76 18,354 30.97 30.97 30.97 30.97 30.97 30.97 6.19 6.19 134.5 11.5 9.0
_spring
Oats C 3 8 Summer 5.2001 87.68 7197 67.01 33.50 33.50 100.51 33.50 33.50 6.70 6.70 103.1 9.5 8.5
mixed_grain C 10 8 Winter 5.2562 87.68 652 70.72 35.36 35.36 106.09 35.36 35.36 7.07 7.07 103.1 9.5 8.5
_winter
mixed_grain C 3 8 Summer 5.2562 87.68 615 67.73 33.87 33.87 101.60 33.87 33.87 6.77 6.77 103.1 9.5 8.5
_spring
grain_maize L 4 11 Summer 5.9976 134.12 2453 96.91 48.46 48.46 145.37 48.46 48.46 9.69 9.69 134.5 19.1 25.4
triticale C 10 8 Winter 6.2676 86.16 25,415 82.87 41.44 41.44 124.31 41.44 41.44 8.29 8.29 103.1 9.5 8.5
_winter
other_forage L 4 10 Summer 13.6743 98.57 155,108 133.70 66.85 66.85 200.55 66.85 66.85 13.37 13.37 88.2 4.5 4.0
_crops
maize_dry L 4 10 Summer 13.6743 98.57 173,691 188.02 94.01 94.01 282.02 94.01 94.01 18.80 18.80 134.5 19.1 25.4
_matter_BG
Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg

dried_pulses L 3 8 Summer 3.9541 25.29 1206 9.96 4.98 4.98 14.94 4.98 4.98 1.00 1.00 22.4 11.5 12.7
(continued)
99
Table 3 (continued)
100

cropName T S E Season Yield Price Output Seed Protection Other Rentals Machine Labour Area Building N P K
Beans L 1 1 All 3.5195 125.00 271 43.81 21.91 21.91 65.72 21.91 21.91 4.38 4.38 22.4 11.5 12.7
Potatoes L 4 10 Summer 33.1854 179.14 20,044 625.46 312.73 312.73 938.18 312.73 312.73 62.55 62.55 12.5 26.7 48.1
Rapeseed L 9 7 Winter 3.9170 259.84 17,572 141.97 70.99 70.99 212.96 70.99 70.99 14.20 14.20 67.2 7.6 4.2
clover_grass L 9 7 Winter 53.1337 29.26 98,297 137.45 68.72 68.72 206.17 68.72 68.72 13.74 13.74 22.4 11.5 12.7
_mix
Meadows O 1 1 All 8.2248 163.53 74,229 118.91 59.46 59.46 178.37 59.46 59.46 11.89 11.89 88.2 4.5 4.0
Pastures O 1 1 All 8.2251 222.87 479,877 162.07 81.03 81.03 243.10 81.03 81.03 16.21 16.21 88.2 4.5 4.0
Vineyards O 1 1 All 10851.3740 1.97 14,106,786 2042.74 1021.37 1021.37 5106.85 3064.11 3064.11 204.27 204.27 22.4 11.5 12.7
crops_NES O 1 1 All 6.2023 330.04 1985 226.24 113.12 113.12 339.36 113.12 113.12 22.62 22.62 22.4 11.5 12.7
S. Rege et al.
Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg 101

140
Buy
Sell
130

120

110

100

90

80

70
0 500 1000 1500 2000 2500 3000 3500 4000 4500

Fig. 2 Market clearance. In x axis, the quantity in tons for a crop, in y axis the price in e per ton

(b) A partial quantity is available for sale and the market clears at price pbrb but
only with lesser quantity qarb . There is a shortfall of qbrb − qarb > 0, which is
then added to the remaining rounds.
For the seller something similar
(a) The entire quantity offered by the farmer (seller) is sold at price pbrb
(b) Only a partial quantity is sold qars and there is a shortfall of qsrs − qars > 0,
which is then added to the remaining rounds.
The market clearance takes place as illustrated in Fig. 2. The farmers are sorted
in the ascending order of their price bids and a supply curve is generated. For the
demand curve each buyer at each round is willing to buy a quantity qbrb at price pbrb .
The demand curve is a horizontal line at price pbrb which is qbrb wide.
The behaviour aspects enter the process after each round. Each agent (buyer or
seller individual) evaluates his/her inventory of the unsold or non procured stock.
Based on their behaviour, each agent is willing to modify their quantity and price
quotes.
We follow one strict procedure as long as all buyers are not done with the
purchase in a specific round. Modifications to the pricing and quantity behaviour
by previously successful buyers in the same round is disallowed. For example, in
buyer round 2, if buyer 4 was the highest bidder to procure 10 tons of maize at
300e/t, buyer 4 is not allowed to purchase additional quantity of maize even though
the revised bid may be the highest, until all buyers have finished their purchase for
that round. All the remaining buyers can be unsuccessful.
102 S. Rege et al.

Similar procedure is adopted for farmers with a round being complete before
moving to the next.
The price convergence dynamics enters the market via the behaviour when buyers
and sellers panic or are cool and modify their price and quantity quotes.
The behaviour part is divided into impatient and patient buyers and sellers.
Impatience can be applied to price or quantity or both.
1. For an impatient buyer b, if shortfall is sbrb , then the modified price is pn =
 sb
r
po × 1 + brb . Increase the bid price in the next round by the extent of shortfall.
qb
2. For a buyer b, with quantity impatience, if shortfall is sbrb , then the amount to be
purchased in the next round r b + 1 is sum of the existing shortfall sbrb and the
r rn ro
original bid in round r b + 1: qbb+1 . Hence qbb+1 = qbb+1 + sbrb .
3. For a medium level of quantity impatience, the shortfall could be spread over the
remaining rounds in an equal fashion leading to a smooth buying profile. If there
rb
s
are k rounds left, then the shortfall for the each of the remaining round is bk ,
which is added to the bid quantity of the future rounds.
4. For a patient buyer, the bid price hike could be in steps of the remaining rounds
 sb 
r
with pn = po × 1 + b rb .
k×qb

Similar behaviour applies to the sellers but in a reverse manner with lower quotes on
prices in case they are unable to sell their output at the quoted price. The impatience
on the quantity however would remain similar as both buyers and sellers want an
increasing amount of quantity traded over successive rounds.

3.3 Scheduling

The buyers are sorted according to the descending order of prices quoted, with the
highest bidder having priority. At each round, the buyer with the highest price
completes the purchase before another buyer is permitted to begin buying. The
buyers are free to modify their purchase price and quantities during future rounds.
Similarly the sellers or farmers are sorted according to the ascending order of price
with the farmer quoting the lowest getting the priority to sell. The farmers are also
allowed to modify their markups and quantities for sale.

4 Experiments and Results

The agriculture system in Luxembourg does not lend itself to open access to
information on price discovery of farm products. One can confidently assume that
the markets are thin, in terms of volume turnover, and the amount of trade in any
Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg 103

commodity is sparse. Most of the crops grown are used as feed for animals and
may not be traded in the open market but carries a shadow price. This is determined
by the opportunity cost of planting or not planting a particular crop at any season.
To compound the lack of data, the main stay of the sector is animal products like
milk and meat, with milk being the dominant commodity. Milk is procured from
the farmers by the co-operative of farmers themselves and thus price setting is
opaque to the outside world. In such a scenario, we proposed a generic tool to
discover the price in any market wherein buyers and sellers engage in trading over
multiple rounds. We assume that the last traded price is available to all the market
participants.2
From a perspective broader than the work we present here, the ultimate goal is to
study the evolution of cropping patters when maize is given some sort of incentive to
be used as an input in biogas plants. This warrants a dynamic model wherein the past
behavioural responses determine the future evolution of the agricultural landscape.
One could in principle use two approaches.
1. Use an exogenous model for price evolution of all crops based on time series
methodology
2. Determine prices endogenously based on price discovery and enable the intro-
duction of a wide genre of behaviours ranging from using own and cross price
elasticity of supply for crop by farmers to risk appetite modelled via impatience
exhibited in maintaining the minimum possible stock for each crop.
We use the latter approach in the current study as it enables endogenous modelling
of behaviour that is consistent over time and also helps in determining the impact of
agricultural practices such as crop rotation on the price of crops.

4.1 Modelling Behaviour

Individuals are different and exhibit a variety of responses to the same stimuli. This
is true of markets too. Some have a low capacity to risk while others are high risk
takers. When the information dimension is expanded to include wealth, one could
then further classify rich and poor traders who are risk averse and who exhibit a high
appetite for risk. From our model perspective we have kept the risk behaviour uni-
dimensional in that we do not differentiate based on wealth of agents. Agents only
decide the extent to which they would like to smooth the shortfall that they incur
in trading at each round. The shortfall is determined by difference between their
expected purchase or sale quantity and the actual amount of quantity purchased or

2 This assumption is violated in many markets wherein each buyer and seller negotiate a price in
private and is unknown to all, but to the buyer and seller. Each buyer and seller goes through a
protracted buying and selling process with multiple individuals before forming an opinion of the
supply and demand in the market.
104 S. Rege et al.

sold. A shortfall for the buyer implies that the quoted price was low while that for a
seller implies that the quoted price was on the higher side. Based on the proclivity
of the buyer or seller towards risk, a buyer or seller might want to modify the price
quote for the next round besides modifying the target quantity for sale or purchase.
The approach to smooth the shortfall over the remaining rounds implicitly defines
the impatience of the agents. We arbitrarily set the maximum rounds to smooth to
shortfall to 4, but this could be easily extended to an arbitrary number Sr . We run
simulations for buyers and sellers changing the number of rounds to smooth the
shortfall from 1 to 4. If the response to shortfall made no difference then we would
observe similar distributions of shortfalls over the different smoothing responses.
The moot point is that the price quoted by buyers is based on some previous price
that is available in the base data but there is no additional source of information that
will enable the buyers to deviate substantially from that price. The same applies
to the sellers. This is a crucial aspect in the time evolution of the model as the
lack of formal methods to identify a price trend or speculation does not lead to
substantial swings in pricing. 2009 was a bad year in which prices were almost at
an all time low and they picked up in 2010. Given the lack of information on supply
and demand in other markets available to both buyers and sellers, there is no a priori
case to arbitrarily set a price different from the existing one. From a pure academic
perspective, one could set extreme prices (high and low) that would only lead to a
particular buyer either meeting the target earlier at a higher cost or not being able to
meet the target on account of lower prices. From the farmers’ perspective as long as
the prices cover the cost of cultivation, the only impact is on the profits. Agriculture
markets are predominantly buyer markets with prices set by buyers or middlemen.
The farmers have little or no market power and are left to the vagaries of not only
nature but also market elements. In such a situation where there is no alternative of
access to other markets or limited ability to store output across seasons, there is no
choice but to sell at the prevailing price.3

4.2 Experiments

We have 2242 farmers, 10 buyers (randomly chosen number), 4 rounds for farmers
to choose smoothing (1–4), 4 rounds for buyers to choose smoothing (1–4) for 41
simulations for 22 crops. In sum, there are an arbitrary maximum of 19 trading
rounds for the buyers, times 2242 sellers, times 10 buyers, times 16 smoothing
options (4 for sellers, 4 for buyers), times 22 crops = 19 × 2242 × 10 × 16 × 22 =
149,944,960. All this repeated 41 times, yielding 6,147,743,360 market clearance
rounds.

3 Thedownward pricing pressure on the French farmers after the tit-for-tat sanctions imposed by
Russia on agriculture produce from NATO countries is a classic case of the above.
Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg 105

In the following experiments we have restricted the simulations to only 1

time period as the objective was to study the impact of various behaviours on
price discovery. The model presented in Sect. 3 was coded in java and the full
simulation was run in an average PC with 4 GigaBytes of ram and a Centrino
I I 2010 microprocessor. The full simulation took around 48 h. No specific agent
programming libraries were used, only code developed explicitly for the study and
libraries for reading and writing files.

4.3 Results

We focus on the price discovery for maize_dry_matter_BG (BG for biogas) in

the figures. The simulations for price discovery are run for all crops. Figure 3
shows the box plots of market clearing price (PE) for each buyer (0–9) for
different smoothing options (1–4). Figure 4 shows the box plots of shortfall
for all buyers across different smoothing options. The mean of the shortfall for
smoothing rounds 1–4 are μs = [153.66, 152.90, 169.88, 189.45] with a standard
deviation of σs = [77.62, 71.61, 77.33, 73.79], respectively. The mean and standard
deviation for prices is as follows μpe = [81.61, 81.46, 80.30, 80.38] and σpe =
[6.85, 6.84, 6.50, 6.18]. We find that the prices are similar due to the fact that the
buyers set the price and these prices do not lead to a loss for the farmers. In addition
the choice available to the farmers is to vary the quantity supplied and implicitly
accept a markup over costs that does not result in losses.
Figures 5 and 6 show the cumulative distribution functions of market clearing
price (PE) and shortfall, respectively, for each buyer (0–9) for different smoothing
options (1–4).
Tables 4 and 5 show the computed values for the Anderson-Darling (AD)
and Kolmogorov-Smirnov (KS) test to compare the similarities between the
distributions of the prices and shortfalls across different smoothing options (1–4).
What we find is that barring smoothing choice over 1 or 2 rounds, the cumulative
distribution of shortfall and prices do not statistically exhibit similarity as shown by
the AD and KS test. The null hypothesis that all smoothing over rounds makes no
difference to the overall market outcome does not hold. We find that both quantities
and prices show different distributions.

4.3.1 Price Convergence

In theory we can plot 16 graphs of variations in the selling prices of sellers for
4 smoothing rounds of buyers and sellers with variations across rounds. Figure 7
shows the variation in quoted selling price by sellers aggregated across all rounds
for different seller smoothing behaviours. One of this variation for seller smoothing
round and buyer smoothing round both equaling 1 is shown in Fig. 8. We largely
observe a smaller variation across price quotes as the smoothing rounds increases
106 S. Rege et al.

Fig. 3 Boxplot of prices for different smoothing rounds of buyers

from b = 1 to b = 4. The choice of smoothing has a bearing on the quantity

procured by each buyer in the subsequent rounds and smoothing over a single round
exhibits aggressive behaviour to cover the shortfall at the earliest possible time, with
an aggressive price bid. This effect is countered by the price quote mechanism used
in the model that uses an arithmetic mean of the average of previous market prices
and the price quote of the buyer.

5 Discussion

Globally farmers face risk on multiple fronts. There is little they can do to mitigate
environmental risks and risks attributed to weather patterns. The possibility of
market risk can be hedged to a certain extent by planting multiple crops and having
access to the market.
The market maker in these markets is the middleman and larger the number of
middlemen in the chain from the farm gate to food plate, the greater is the gap
between the realised price for the farmer and the market price paid by the consumer.
Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg 107

Fig. 4 Boxplot of shortfall for different smoothing rounds of buyers

1.00

0.75
y

0.50

0.25

0.00

60 70 80 90 100
x
buyer_smoothing 1 2 3 4

Fig. 5 Cumulative distribution function of prices for different smoothing rounds of buyers
108 S. Rege et al.

1.00

0.75
y

0.50

0.25

0.00

0 200 400 600

x

buyer_smoothing 1 2 3 4

Fig. 6 Cumulative distribution function of shortfall for different smoothing rounds of buyers

Table 4 AD & KS test for CDF AD test AD p value KS test KS p value

difference in price
convergence across s1:s2 1.0013 0.35617 0.042262 0.0995
smoothing s1:s3 20.69 1.3652e–11 0.125 7.958e–12
s1:s4 33.688 9.8185e–19 0.18274 2.2e–16
s2:s3 16.252 3.7518e–09 0.10595 1.29e–08
s2:s4 28.329 8.6558e–16 0.14405 1.443e–15
s3:s4 6.1219 0.00081821 0.078571 6.263e–05

Table 5 AD & KS test for CDF AD test AD p value KS test KS p value

difference in shortfall across
smoothing s1:s2 0.6523 0.59941 0.02381 0.7277
s1:s3 22.796 9.4979e–13 0.12381 1.309e–11
s1:s4 153.39 1.6255e–84 0.38988 2.2e–16
s2:s3 22.892 8.4196e–13 0.11786 1.467e–10
s2:s4 152.8 3.4214e–84 0.38452 2.2e–16
s3:s4 72.966 2.5516e–40 0.31964 2.2e–16

Buyers and sellers implement strategies to get the maximum return on their
investment. The buyers want to procure the output at the least possible price as
opposed to the farmers who want the maximum. In absence of information these
markets are supposed to be operating in perfect competition as far as farmers are
concerned.
Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg 109

Fig. 7 Boxplot of prices for different smoothing rounds of sellers aggregated across all rounds.
(a) Seller smoothing round = 1, (b) Seller smoothing round = 2, (c) Seller smoothing round = 3,
(d) Seller smoothing round = 4

Fig. 8 Boxplot of prices for different smoothing rounds of buyers

110 S. Rege et al.

Our model is an agent based model which though classified as a bottom-up

model,4 the driver of the model is still top-down.5 In standard economic models
such as computable general equilibrium or partial equilibrium models, there is the
law of a single price. However in reality, markets operate through a double auction
process and there is a variation in the price and one normally assumes that the market
clearing price is the mean of the prices observed during the day.
What we have embarked upon is to discover if there is a strategy in smoothing
inventory of unsold stock in perishable commodities as exhibited in the agriculture
sector. Quoting lower price would mean lower returns but possible lower or zero
inventory, while higher price would mean the opposite. In our model there is no
way to control the initial opening quotes of the buyers and it is possible that the
opening quote can be lower than the production cost of the farmer, thus causing a
loss to the farmer. This is a phenomenon not unheard of in agriculture markets.
We find that price does converge to a “mean” value as exhibited by a lower
variation in the box plots of prices over buyer and seller rounds. The case where
both the buyers and sellers are smoothing over multiple rounds enables the price
discovery process to exhibit lower volatility captured by the lower extreme values
in the box plots.
We are assuming that the buyers are unable to exhibit market power or collude in
order to dictate the price and that the initial quote is above the cost of production of
the farmer for a specific crop.
Empirical validation is the best possible test for any agent based model. However
in Luxembourg, given the legislation on data confidentiality, it is impossible to
obtain information on the price and profitability of the agriculture operations. We
have carried out a survey of farmers with limited success and it was difficult
to obtain information across all data fields to generate meaningful statistical
distribution to estimate parameters for the agent based model. Despite this setback,
the model has been used to generate numbers that are close to reality. Given the
vagaries in the agriculture sector it is difficult if not impossible to use agent based
models as a forecasting tool for price evolution and use these forecasts to evaluate
the potential cropping pattern changes induced by policy interventions.

6 Conclusions

We have developed a generic model for any agriculture system with pure cropping
and endogenous price discovery of crops. In absence of market information, lack
of time series for exogenously forecasting prices, this approach has the benefit of
endogenous price discovery. The behaviour to smooth buying and selling across
multiple agents introduced the element of patience as far as purchasing or selling

4 Decisions are made by individual agents.

5 The price quotes of the buyers are exogenously given before each simulation.
Modelling Price Discovery in an Agent Based Model for Agriculture in Luxembourg 111

decisions are concerned. In the current simulation we have used a single layer of
buyers interacting directly with farmers. We believe that given a sufficient number
of layers of buyers and sellers, where within each layer the buyer has market power,
one can show evidence of the price gap between food plate and farm gate, leading
to low realised prices for farmers.
From an architecture perspective for agent based models, the model is in a
standard reactive framework as far as price discovery is concerned. A cognitive
architecture that would take into account memory and different beliefs and inten-
tions of agents, for example, could be considered in future work. For the current
state, given the limited time period for simulation, the changes in the agriculture
system in Europe, and lack of more precise data, make it difficult if not impossible
to implement.
From an economic price discovery perspective, this is an empirical demonstration
of a bottom-up model using agent based models. The limitations of all bottom-up
models are relevant for this approach too. The information asymmetry built into
bottom-up models as opposed to top-down models prevents the agents from setting
prices that are not generated from the model simulations.
Finally smoothing behaviour of covering shortfall for buyers and sellers used
as a proxy for impatience does exhibit statistically significant difference as far as
distribution of prices is concerned over different smoothing rounds.

Acknowledgements This work was done under the project MUSA (C12/SR/4011535) funded
by the Fonds National de la Recherche (FNR), Luxembourg. We thank Romain Reding and
Rocco Lioy from CONVIS (4 Zone Artisanale Et Commerciale, 9085 Ettelbruck, Grand-duchy of
Luxembourg) for their valuable insight and for the participation to the discussions for the definition
of the project MUSA, and of the data collection survey. We thank professors Shu-Heng Chen, Ye-
Rong Du, Ragu Pathy, Selda Kao and an anonymous referee for their valuable comments. This
paper was presented at the 21st International Conference on Computing in Economics and Finance
June 20–22, 2015, Taipei, Taiwan. Sameer Rege gratefully acknowledges the FNR funding for the
conference participation.

References

1. Berger, T. (2001). Agent-based spatial models applied to agriculture: A simulation tool for
technology diffusion, resource use changes and policy analysis. Agricultural Economics, 25,
245–260.
2. Berta, F. E., Podestá, G. P., Rovere, S. L., Menéndez, A. N., North, M., Tatarad, E., et al. (2011).
An agent based model to simulate structural and land use changes in agricultural systems of
the argentine pampas. Ecological Modelling, 222, 3486–3499.
3. Cooper, J. S., & Fava, J. A. (2006). Life-cycle assessment practitioner survey: Summary of
results. Journal of Industrial Ecology, 10(4), 12–14.
4. Happe, K., Kellermann, K., & Balmann, A. (2006). Agent-based analysis of agricultural poli-
cies: An illustration of the agricultural policy simulator AgriPoliS, its adaptation, and behavior.
Ecology and Society, 11(1), 329–342. Available from: http://www.ecologyandsociety.org/
vol11/iss1/art49/.
112 S. Rege et al.

5. Howitt, R. E. (1995). Positive mathematical programming. American Journal of Agricultural

Economics, 77, 329–342.
6. ISO. (2010). Environmental management — Life cycle assessment — Principles and frame-
work. Geneva, Switzerland: International Organization for Standardization. 14040:2006.
7. KTBL. (2006). Faustzahlen für die Landwirtschaft (in German). Darmstadt, Germany: Kura-
torium für Technik und Bauwesen in der Landwirtschaft.
8. Le, Q. B., Park, S. J., Vlek, P. L. G., & Cremers, A. B. (2010). Land-use dynamic simulator
(LUDAS): A multi-agent system model for simulating spatio-temporal dynamics of coupled
human–landscape system. I. Structure and theoretical specification. Ecological Informatics, 5,
203–221.
9. Rege, S., Arenz, M., Marvuglia, A., Vázquez-Rowe, I., Benetto, E., Igos, E., et al. (2015).
Quantification of agricultural land use changes in consequential Life Cycle Assessment using
mathematical programming models following a partial equilibrium approach. Journal of
Environmental Informatics. https://doi.org/10.3808/jei.201500304
10. SER. Available from: http://www.ser.public.lu/
11. STATEC. Available from: http://www.statistiques.public.lu/en/actors/statec/index.htm
Heterogeneity, Price Discovery and
Inequality in an Agent-Based Scarf
Economy

Shu-Heng Chen, Bin-Tzong Chie, Ying-Fang Kao, Wolfgang Magerl,

and Ragupathy Venkatachalam

Abstract In this chapter, we develop an agent-based Scarf economy with het-

erogeneous agents, who have private prices and adaptively learn from their own
experiences and those of others through a meta-learning model. We study the
factors affecting the efficacy of price discovery and coordination to the Walrasian
Equilibrium. We also find that payoff inequality emerges endogenously over time
among the agents and this is traced back to intensity of choice (a behavioural
parameter) and the associated strategy choices. Agents with high intensities of
choice suffer lower payoffs if they do not explore and learn from other agents.

Keywords Non-tâtonnement processes · Coordination · Learning · Agent-based

modeling · Walrasian general equilibrium · Heterogeneous agents

1 Introduction

How do market economies achieve coordination, even if only imperfectly, among

millions of actors performing a multitude of actions without centralization?
Research in General Equilibrium theory over many decades has provided many
interesting insights. The notion of coherence is in this literature interpreted in terms

S.-H. Chen () · Y.-F. Kao

AI-ECON Research Center, Department of Economics, National Chengchi University, Taipei,
Taiwan
B.-T. Chie
Tamkang University, Tamsui, Taipei, Taiwan
W. Magerl
Vienna University of Technology, Vienna, Austria
AI-ECON Research Center, Department of Economics, National Chengchi University, Taipei,
Taiwan
R. Venkatachalam
Institute of Management Studies, Goldsmiths, University of London, London, UK

© Springer Nature Switzerland AG 2018 113

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_6
114 Shu-Heng Chen et al.

of an equilibrium phenomenon. Furthermore, the discovery of equilibrium prices

that balance the demands and supplies of various actors in the aggregate is assumed
to be mediated through a fictional, centralized authority. This authority, known as
the Walrasian auctioneer, supposedly achieves this discovery through a process of
trial and error (tâtonnement). In this framework, trading happens, if it at all happens,
only in equilibrium.
What happens when agents are out of this equilibrium configuration? Whether
and how these agents manage to coordinate back to equilibrium over time through
decentralized exchanges is an important question. The tâtonnement narrative is
highly divorced from how agents go about achieving coordination by searching and
learning to discover efficient prices in reality. Therefore, it becomes necessary to
go beyond the tâtonnement process and find plausible descriptions of behaviour that
respect the cognitive and informational limitations that human agents face. Research
on non-tâtonnement processes [8, 10, 25] that permit disequilibrium trading have
been developed over the years. Agent-based models provide a conducive setup to
understand decentralized, disequilibrium dynamics of market economies.
Many aggregate economic outcomes such as growth and fluctuations are often
explained as being driven by the expectations of agents. There is often a diversity
of expectations among agents and in the face of this heterogeneity, the aggregate
outcomes that result from the interaction among agents are complex to understand.
But how these expectations come into being and evolve over time has not yet been
completely understood. Similarly, their psychological underpinnings and relation
to agent-level characteristics have not been sufficiently clarified. The mechanics
of expectation formation are intimately linked to how agents learn from their
experiences and the environment. Hence, it may not be sufficient to just point to
the existence of a diversity of expectations, but it may be necessary to go deeper
into understanding their origins and dynamics.
Even in a simple decentralized economy with barter, there is no a priori reason to
believe that heterogeneous agents can successfully learn to align their expectations
regarding future prices. There is no guarantee that they will eventually discover the
equilibrium prices, where all mutually beneficial exchanges will be exhausted. Even
if we suppose that they do, how the gains from these exchanges will be distributed
among these agents and their relation to the structure of expectations is not
obvious.
In this chapter, we investigate these issues through an agent-based, Scarf
economy, where the agents are characterized by Leontief payoff functions. Agents
attempt to maximize their payoffs, and engage in bilateral trade with others based
on their initial endowment and a subjective perception of “fair” prices. Trade
failures, either in the form of unsatisfied demand or involuntary inventory, signal
misperceptions in the economy. No one person has complete knowledge of the
entire system and the economy is best viewed as a complex adaptive process that
dynamically evolves. In order to successfully coordinate in such a system, agents
may need to learn both from their own experience (individual learning) and from
each other (social learning). In our model, agents consciously decide when to engage
in social and individual. This choice is formulated as meta-learning and we apply
reinforcement learning to model this meta-learning. Heterogeneity among agents
Heterogeneity, Price Discovery and Inequality 115

is in terms of differences in their learning process, more specifically, the intensity

of choice parameter of the meta-learning model. Within this set-up, we ask the
following questions:
• Can heterogeneous agents successfully discover the (unique) equilibrium prices
of the model?
• How do current and accumulated payoffs vary among agents who are identical in
all respects except for their intensity of choice?
• If there is a substantial variation in payoffs, what are the factors that drive this
inequality?
The rest of this chapter is organized as follows: Sect. 2 develops the agent-
based Scarf economy, and Sect. 3 describes the individual, social and meta-learning
mechanisms that the agents in the model employ. The simulation design is explained
in Sects. 4 and 5 presents the results. Section 6 provides a discussion of the results.

2 The Scarf Economy and the Non-Tâtonnement Process

Herbert Scarf [20] demonstrated an economy in which prices may not converge
to the Walrasian general equilibrium under the tâtonnement process, even when the
equilibrium is unique. This paper has since led to a lot of research on the issue of the
stability of the general equilibrium. Recently, [9] developed an agent-based model
and showed that when agents hold private prices and imitate other agents who have
been successful, prices converge to the unique general equilibrium of the Scarf-like
economy. The Scarf economy has a simple structure that is amenable to investigation
and has been widely studied. Previous studies in this area have not tackled the issue
of heterogeneity among agents in a Scarf-like set-up. In the remainder of the section,
we will develop a heterogeneous, agent-based model of the Scarf economy.

2.1 An Agent-Based Model of the Scarf Economy

Following [20], we consider a pure exchange economy composed of N agents. This

economy has three goods, denoted by j = 1, 2, 3, and correspondingly, N agents are
grouped into three different ‘types’, τj , j = 1, 2, 3, where τ1 ≡ {1, . . . , N1 }; τ2 ≡
{N1 + 1, . . . , N1 + N2 }; τ3 ≡ {N1 + N2 , . . . , N }. Agents who belong to τj (type-j )
are initially endowed with wj units of good j , and zero units of the other two goods.
Let Wi be the endowment vector of agent i:
⎧
⎨ (w1 , 0, 0), i ∈ τ1 ,
Wi = (0, w2 , 0), i ∈ τ2 , (1)
⎩
(0, 0, w3 ), i ∈ τ3 .
116 Shu-Heng Chen et al.

All agents are assumed to have a Leontief-type payoff function.

⎧ x2 x3
⎪
⎨ min{ w2 , w3 }, i ∈ τ1 ,
Ui (x1 , x2 , x3 ) = min{ wx11 , wx33 }, i ∈ τ2 , (2)
⎪
⎩ min{ x1 , x2 }, i ∈ τ .
w1 w2 3

Given the complementarity feature of this payoff function, we populate equal

numbers of agents in each type. This ensures that the economy is balanced and
that no free goods appear.
We assume that agents have their own subjective expectations of the prices of
different goods,

Pei (t) = (Pi,1

e e
(t), Pi,2 e
(t), Pi,3 (t)), i = 1, . . . , N. (3)

e (t) is agent i’s price expectation of good j at time t.1

where Pi,j
Given a vector of the subjective prices (private prices), the optimal demand
vector, X∗ = Ψ ∗ (W) can be derived by maximizing payoffs with respect to the
budget constraint.
⎧
⎨ (0, ψi∗ w2 , ψi∗ w3 ), i = 1, . . . , N1 ,
∗ ∗ ∗ ∗
Xi = (xi,1 , xi,2 , xi,3 ) = (ψi∗ w1 , 0, ψi∗ w3 ), i = N1 + 1, . . . , N1 + N2 , (4)
⎩ ∗
(ψi w1 , ψi∗ w2 , 0), i = N1 + N2 + 1, . . . , N.

where the multiplier

⎧
⎪ e
⎨ (Pi,1 w1 )/(
e
j =2,3 Pi,j wj ), i = 1, . . . , N1
ψi∗ = (Pi,2
e w )/(
2
e
j =1,3 Pi,j wj ), i = N1 + 1, . . . , N1 + N2 , (5)
⎪
⎩ (P e w )/(
i,3 3
e
j =1,2 Pi,j wj ), i = N1 + N2 + 1, . . . , N.

Note that the prices in the budget constraint are ‘private’ prices. Rather than
restricting the prices within a certain neighbourhood (for instance, a unit sphere in
[20]), we follow [2] and set one of the prices (P3 ) as the numéraire. The Walrasian
Equilibrium (WE) for this system is P1∗ = P2∗ = P3∗ = 1, when the endowments
for each type are equal and symmetric.2 However, in this model agents may have
own price expectations that may be very different from this competitive equilibrium
price, and may base their consumption decision on their private price expectations
e . To facilitate exchange, we randomly match agents in the model with each other
Pi,j
and they are allowed to trade amongst each other if they find it to be beneficial.

1 More specifically, t refers to the whole market day t, i.e., the interval [t, t − 1).
2 Given this symmetry, the type-wise grouping of agents does not qualify as a source of real
heterogeneity in terms of characteristics. This, as we will see in the later sections, will be
characterized in terms of the agents’ learning mechanisms.
Heterogeneity, Price Discovery and Inequality 117

For this, we need to specify a precise bilateral trading protocol and the procedures
concerning how agents dynamically revise their subjective prices.

2.2 Trading Protocol

We randomly match a pair of agents, say, i and i . Let i be the proposer, and i be the
responder. Agent i will initiate the trade and set the price, and agent i can accept
or decline the offer. We check for the double coincidence of wants, i.e., whether
they belong to the same type. If they do not, they will be rematched. If not, we will
then check whether the agents have a positive amount of endowment in order for
them to engage in trade. Let mi be the commodity that i is endowed with (mi =
1, 2, 3). Agent i, based on his subjective price expectations, proposes an exchange
to agent i .
e x
Pi,m
∗ i i ,mi
xi,m = e , (6)
i Pi,m
i

Here, the proposer (i) makes an offer to satisfy the need of agent i up to xi ,mi in
∗
exchange for his own need xi,m .
i
Agent i will evaluate the ‘fairness’ of the proposal using his private, subjective
expectations and will consider the proposal to be interesting provided that
∗
Pie,mi xi ,mi ≥ Pie,m xi,mi
; (7)
i

otherwise, he will decline the offer. Since the offer is in the form of take-it-or-leave-
it (no bargaining), this will mark the end of trade.
Agent i will accept the proposal if the above inequality (7) is satisfied, and
if xi ,mi ≤ xi∗ ,mi . This saturation condition ensures that he has enough goods to
trade with i and meet his demand. However, only if the saturation condition is not
satisfied, will the proposal still be accepted, but the trading volume will be adjusted
downward to xi,mi < xi,m ∗ . Agents update their (individual) excess demand and
i
as long as they have goods to sell, they can continue to trade with other agents.
The agents are rematched many times to ensure that the opportunities to trade
are exhausted. Once the bilateral exchange is completed, the economy enters the
consumption stage and the payoff of each agent Ui (Xi (t)), i = 1, 2, . . . , N, is
determined. Note that Xi (t), the realized amount after the trading process may not
be the same as the planned level X∗i (t). This may be due to misperceived private
prices and a sequence of ‘bad luck’, such as running out of search time (number of
trials) before making a deal, etc. Based on the difference between Xi (t) and X∗i (t),
each agent i in our model will adaptively revise his or her private prices through a
process of learning.
118 Shu-Heng Chen et al.

3 Learning

Agents have to learn to coordinate, especially in dynamic, disequilibrium environ-

ments and non-convergence to the most efficient configuration of the economy can
be interpreted as a coordination failure. In this section, we introduce two modes of
learning that are frequently used in agent-based computational economics, namely,
individual learning and social learning. In the individual learning mode, agents
learn from their own past experiences. In social learning, agents learn from the
experiences of other agents with whom they interact. These two modes of learning
have been characterized in different forms in the literature and have been extensively
analysed (see: [3, 14, 21, 22]). The impact of individual and social learning on
evolutionary dynamics has been analysed in [18, 26] and more recently in [5]. We
describe the learning procedures in more detail below.

3.1 Individual Learning

Under individual learning, an agent typically learns from analysing his own
experience concerning the past outcomes and strategies. The mechanism that we
employ can be thought of as a modified agent-level version of the Walrasian price
adjustment equation. Let the optimal and actual consumption bundles be denoted
by the vectors X∗i (t) and Xi (t), respectively. Agent i can check his excess supply
and excess demand by comparing these two vectors component-wise. Agents then
reflect on how well their strategy (private price expectations) has performed in the
previous trading round and adjust their strategies based on their own experience of
excess demand and supply. We employ a gradient descent approach to characterize
individual learning and in a generic form it can be written as:

Pei (t + 1) = Pei (t) + ΔPi (Xi (t), X∗i (t)), i = 1, 2, . . . , N (8)

  
gradient descent

We shall detail its specific operation as follows.

y
Let mi and mci denote the production set and consumption set of agent i. In
y
the Scarf economy, mi ∩ mci = ∅ and in this specific 3-good Scarf economy,
y
mi = {mi }. At the end of each market period, agent i will review his expectations
for all commodities, Pi,je (t), ∀j . For the good that the agent ‘produces’,3 j ∈ my ),
i
e (t) will be adjusted downward if m is not completely sold
the price expectations Pi,j i
out (i.e., when there is excess supply). Nonetheless, even if mi has been completely
sold out, it does not mean that the original price expectation will be sustained.
e (t) may be
In fact, under these circumstances, there is still a probability that Pi,j
adjusted upward. This is to allow agent i to explore or experiment with whether

3 More precisely, the good that he is endowed with.

Heterogeneity, Price Discovery and Inequality 119

his produced commodity might deserve a better price. However, so as not to make
our agents over-sensitive to zero-inventory, we assume that such a tendency for
them to be changing their prices declines with the passage of time. That is to say,
when agents constantly learn from their experiences, they gradually gain confidence
in their price expectations associated with zero-inventory. Specifically, the time-
decay function applied in our model is exponential, which means that this kind of
exploitation quickly disappears with time. For those goods in the vector Xi that are
a part of the consumption set of the agent, i.e., j ∈ mci , the mechanism will operate
in exactly the opposite manner by increasing the price expectations if there is excess
demand. The individual learning protocol is summarized below.

3.1.1 Protocol: Individual Learning

1. At the end of the trading day, agent i examines the extent to which his planned
demand has been satisfied. Let

∗ (t) − x (t), if j ∈ mc ,
xi,j i,j
Δxi,j (t) = y
i (9)
0 − xi,j (t), if j ∈ mi

2. The subjective prices Pi,j e of all three goods will be adjusted depending on

| Δxi,j (t) |.
3. If | Δxi,j (t) |> 0 (i.e., | Δxi,j (t) |= 0),

(1 + α(| Δxi,j (t) |))Pi,j

e (t), if j ∈ mc .
e
Pi,j (t + 1) = i
e (t), if j ∈ my . (10)
(1 − α(| Δxi,j (t) |))Pi,j i

where α(.) is a hyperbolic tangent function, given by:

e(ϕ|Δxi,j (t)|) − e(−ϕ|Δxi,j (t)|)

α(| Δxi,j (t) |) = tanh(ϕ | Δxi,j (t) |) = (11)
e(ϕ|Δxi,j (t)|) + e(−ϕ|Δxi,j (t)|)

4. If | Δxi,j (t) |= 0,

(1 − β(t))Pi,j
e (t), if j ∈ mc .
e
Pi,j (t + 1) = i
e (t), if j ∈ my . (12)
(1 + β(t))Pi,j i

where β is a random variable, and is a function of time.

−t
β = θ1 exp , θ1 ∼ U [0, 0.1], (13)
θ2

where θ2 is a time scaling constant.

120 Shu-Heng Chen et al.

3.2 Social Learning

Social learning broadly involves observing the actions of others (peers, strangers or
even members of different species) and acting upon this observation. This process
of acquiring relevant information from other agents can be decisive for effective
adaptation and evolutionary survival. Imitation is one of the most commonly
invoked, simplest forms of social learning. Players imitate for a variety of reasons
and the advantages can be in the form of lower information-gathering costs and
information-processing costs, and it may also act as a coordination device in games
[1]. Although the idea of imitation is fairly intuitive, there are many different forms
in which agents and organisms can exhibit this behaviour ([11], section G, 115–
120.).
We adopt a fairly basic version of imitation behaviour, where agents exchange
their experiences regarding payoffs with other agents, with whom they are randomly
matched. This can be thought of as a conversation that takes place with friends
in a café or a pub at the end of a trading day, where they share their experiences
regarding their amount of and pleasure associated with consumption as well as their
price expectations. An agent with a lower payoff can, based on observing others,
replace his own price expectations with those of an agent with a higher payoff. This
is assumed to be done in the hope that the other agent’s strategy can perform better.
If they both have the same payoffs, then the agent chooses between them randomly.
If the agent ends up meeting someone who has performed worse than him, he does
not imitate and retains his original price expectations.

3.2.1 Protocol: Social Learning

1. At the end of each day, each agent consumes the bundle of goods that he has
obtained after trading, and derives pleasure from his consumption Ui (t) (i =
1, . . . , N ).
2. Agents are matched randomly, either with other agents of the same type or with
agents who are of different types. This is achieved by randomly picking up a pair
of agents (i, i ) without replacement and they are given a chance to interact.
3. Their payoffs are ranked, and the price expectations are modified as follows:
⎧ e
⎨ Pi (t), if Ui (t) < Ui (t),
Pei (t + 1) = Pei (t), if Ui (t) > Ui (t), (14)
⎩
Random(Pei (t), Pei (t)), if Ui (t) = Ui (t).

The protocol makes it easier for the agents to meet locally and enables them to
exchange information. An agent who has performed well can influence someone
who hasn’t performed as well by modifying his perception of the economy (i.e.,
price expectations).
Heterogeneity, Price Discovery and Inequality 121

3.3 Meta Learning

In our model, an agent can consciously choose between social learning or individual
learning during each period. In such a setting, it is necessary to specify the basis on
which—i.e., how and when—such a choice is made. Agents may choose between
these strategies either randomly, or based on the relative expected payoffs by
employing these strategies. In our model, adaptive agents choose individual or social
learning modes based on the past performance of these modes. In other words, the
focus is on learning how to learn. This meta-learning framework is not restricted to
two modes alone, but we begin with the simplest setting. We formulate this choice
between different learning modes as a two-armed bandit problem.

3.3.1 Two-Armed Bandit Problem

In our market environment, an agent repeatedly chooses between two learning

modes, individual learning and social learning. Denote the action space (feasible set
of choice) by Γ , Γ = {ail , asl }, where ail and asl refer to the actions of individual
learning and social learning, respectively. Each action chosen at time t by agent
i yields a payoff π(ak,t )(k = il, sl). This payoff is uncertain, but the agent can
observe this payoff ex-post and this information is used to guide future choices.
This setting is analogous to the familiar two-armed bandit problem. In the literature,
reinforcement learning has been taken as a standard behavioural model for this type
of choice problem [4] and we follow the same approach.

3.3.2 Reinforcement Learning

Reinforcement learning has been widely investigated both in artificial intelligence

[23, 27] and economics [7, 19]. The intuition behind this learning scheme is
that better performing choices are reinforced over time and those that lead to
unfavourable or negative outcomes are not, or, alternatively, the better the experience
with a particular choice, the higher is the disposition or the propensity towards
choosing it in the future. This combination of association and selection, or,
equivalently, search and memory, or the so-called the Law of Effect, is an important
aspect of reinforcement learning [23, chapter 2].
In our model, each agent reinforces only two choices, i.e., individual learning and
social learning (Γ = {ail , asl }). In terms of reinforcement learning, the probability
of a mode being chosen depends on the (normalized) propensity accumulated over
time. Specifically, the mapping between the propensity and the choice probability is
represented by the following Gibbs-Boltzmann distribution:

eλi ·qi,k (t)

Probi,k (t + 1) = , k ∈ {ail , asl }, (15)
eλi ·qi,ail (t) + eλi ·qi,asl (t)
122 Shu-Heng Chen et al.

where Probi,k (t) is the choice probability for learning mode k (k = ail , asl ); we
index this choice probability by i (the agent) and t (time) considering that different
agents may have different experiences with respect to the same learning mode, and
that, even for the same agent, experience may vary over time. The notation qi,k (t)
(k = ail , asl ) denotes the propensity of the learning mode k for agent i at time t.
Again, it is indexed by t because the propensity is revised from time to time based
on the accumulated payoff. The propensity updating scheme applied here is the one-
parameter version of [19].

(1 − φ)qi,k (t) + Ui (t), if k is chosen.
qi,k (t + 1) = (16)
(1 − φ)qi,k (t), otherwise.

where Ui (t) ≡ Ui (Xi (t)), k ∈ {ail , asl }, and φ is the so-called recency parameter,
which can be interpreted as a memory-decaying factor.4 The notation λ is known
as the intensity of choice. With higher λs, the agent’s choice is less random and is
heavily biased toward the better-performing behavioural mode; in other words, the
degree of exploration that the agent engages in is reduced. In the limit as λ → ∞,
the agent’s choice is degenerated to the greedy algorithm which is only interested in
the “optimal choice” that is conditional on the most recent updated experience; in
this case, the agent no longer explores.

3.4 Reference Points

We further augment the standard reinforcement learning model with a reference-

point mechanism to decide when Eq. (15) will be triggered. Reference dependence
in decision making has been made popular by prospect theory [13], where gains and
losses are defined relative to a reference point. This draws attention to the notion
of position concerning the stimuli and the role it plays in cognitive coding that
describes the agent’s perception of the stimuli.
We augment our meta-learning model with reference points, where agents are
assumed to question the appropriateness of their ‘incumbent’ learning mode only
when their payoffs fall short of the reference point (along the lines of [6], p. 152–
153). Let Ui (t)(≡ Ui (Xi (t))) be the payoff of an agent i at time t. Let his reference
point at time t be Ri (t). The agent will consider a mode switch only when his
realized payoff Ui (t) is lower than his reference point Ri (t).

4 Following [4], a normalization scheme is also applied to normalize the propensities qi,k (t + 1) as
follows:
qi,k (t + 1)
qi,k (t + 1) ← . (17)
qi,ail (t + 1) + qi,asl (t + 1)
Heterogeneity, Price Discovery and Inequality 123

The reference points indexed by t, Ri (t), imply that they need not be static
or exogenously given; instead, they can endogenously evolve over time with the
experiences of the agents. Based on the current period payoffs, the reference point
can be revised up or down. This revision can be symmetric, for example, a simple
average of the current period payoffs and the reference point. A richer case as shown
in Eq. (18) indicates that this revision can be asymmetric; in Eq. (18), the downward
revisions are more pronounced than the upward revisions.

Ri (t) + α + (Ui (t) − Ri (t)), if Ui (t) ≥ Ri (t) ≥ 0,
Ri (t + 1) = (18)
Ri (t) − α − (Ri (t) − Ui (t)), if Ri (t) > Ui (t) ≥ 0.

In the above equation, α − and α + are revision parameters and α − , α + ∈ [0, 1].
The case with α − > α + would indicate that the agents are more sensitive to negative
payoff deviations from their reference points.5 Note that this is similar to the idea
of loss aversion in prospect theory [24], where the slopes of the values function on
either side of the reference point are different. For the rest of the simulations in this
paper, we have utilized the asymmetric revision case.

4 Simulation Design

4.1 A Summary of the Model

4.1.1 Scale Parameters

Table 1 presents a summary of the agent-based Scarf model which we describe in

Sects. 2 and 3. It also provides the values of the control parameters to be used in
the rest of this paper. The Walrasian dynamics of the Scarf economy run in this
model has the following fixed scale: a total of 270 agents (N = 270), three goods
(M = 3), and hence 90 agents for each type of agent N1 = N2 = N3 = 90.
The initial endowment for each agent is 10 units (wi = 10), which also serves as
the Leontief coefficient for the utility function. Our simulation routine proceeds
along the following sequence for each market day (period): production process
(endowment replenishment), demand generation, trading, consumption and payoff
realization, learning and expectations updating, propensity and reference point
updating, and learning mode updating. Each market day is composed of 10,000
random matches (S = 10,000), to ensure that the number of matches is large enough
to exhaust the possibilities of trade. At the beginning of each market day, inventories
perish (δ = 1) and the agents receive fresh endowments (10 units). Each single run
of the simulation lasts for 2500 days (T = 2500) and each simulation series is

5 The results do not vary qualitatively for perturbations of these parameters.

124 Shu-Heng Chen et al.

Table 1 Table of control parameters

Parameter Description Value/Range
N Number of agents 270
M Number of types 3
N1 , N2 , N3 Numbers of agents per type (4), (5) 90, 90, 90
wi (i = 1, 2, 3) Endowment (1) 10 units
Leontief coefficients (2) 10
δ Discount rate (Perishing rate) 1
S Number of matches (one market day) 10,000
T Number of market days 2500
Pie (0) (i = 1, 2, 3) Initial price expectations (3) ∼ Uniform[0.5, 1.5]
ϕ Parameter of price adjustment (11) 0.002
θ1 Parameter of price adjustment (13) ∼ Uniform[0,0.1]
θ2 Parameter of price adjustment (13) 1
K Number of arms (Sect. 3.3.1) 2
λ Intensity of choice (15) ∼ Uniform(0,8), ∼Uniform(0,15),
∼Normal(4,1), ∼Normal(8,1)
POPail (0) Initial population of ail 1/2
POPasl (0) Initial population of asl 1/2
φ Recency effect (16) 0
α+ Degree of upward revision (18) 0.1
α− Degree of downward revision (18) 0.2
The numbers inside the parentheses in the second column refer to the number of the equations in
which the respective parameter lies

repeated 50 times.6 These scale parameters will be used throughout all simulations
in Sect. 5.

4.1.2 Behavioural Parameters

The second part of the control parameters is related to the behavioural settings
of the model, which begins with initial price expectations and price expectation
adjustment, followed by the parameters related to the meta-learning models. As
noted earlier, we set good 3 as a numéraire good, whose price is fixed as one. Most
of these parameters are held constant throughout all simulations, as indicated in
Table 1, and specific assumptions concerning distributions are specified in places
where necessary.
We can systematically vary the values of the different parameters. The focus of
this paper is on the intensity of choice (λ) (Table 1). The initial price vector, Pei (0),

6 We have examined the system by simulating the same treatments for much longer periods and we
find that the results are robust to this.
Heterogeneity, Price Discovery and Inequality 125

for each agent is randomly generated where the prices lie within the range [0.5,1.5],
i.e., having the WE price as the centre. Except in the cases involving experimental
variations of initial conditions, agents’ learning modes, namely, individual learning
(innovation) and social learning (imitation) are initially uniformly distributed
(POPail (0) = POPasl (0) = 1/2).

4.2 Implementation

All simulations and analysis were performed using NetLogo 5.2.0 and Matlab
R2015a. To comply with the current community norm, we have made the computer
program available at the OPEN ABM.7 We classify the figure into five blocks. The
first block (the left-most blocks) is a list of control bars for uses to supply the values
of the control parameters to run the economy. The parameters are detailed in Table 1,
and include N , M, S, T , ϕ, θ1 , θ2 , K, λ, POPRE , and POPail (0). In addition to
these variables, other control bars are for the selection of the running model, and
include individual learning (only), social learning (only), the exogenously given
market fraction, and the meta-learning model. For the exogenously given market
fraction, POPail (0) needs to be given in addition.

5 Results

5.1 Heterogeneity and Scaling Up

Unlike the representative agent models in mainstream economic theory, agent-

based models are capable of incorporating heterogeneity in terms of the varying
characteristics of the agents. The importance of heterogeneity in agent-based models
is further underscored because of its potential to break down linear aggregation
due to interaction effects. In this context, we can ask whether the quantitative
phenomena observed in a heterogeneous economy are simply a linear combination
of the respective homogeneous economies. If the aggregation of the effects is
linear, the behaviour of the heterogeneous model can be simply deduced from the
homogeneous economies. If not, studying a heterogeneous version of the economy
becomes a worthwhile exercise to gain additional insights.
In our model, the only non-trivial source of heterogeneity comes from the
intensity of choice (λ) that the agents possess. For simplicity, we simulate a
heterogeneous economy with only two (equally populated) groups of agents with
λi values of 1 and 7, respectively. To compare these, first, we need to determine
the variable which will form the basis of comparison between the homogeneous

7 https://www.openabm.org/model/4897/
126 Shu-Heng Chen et al.

and heterogeneous economies. Second, we need a method for determining the

values of the above chosen variable across different homogeneous economies, which
correspond to the heterogeneous economy in question. For the first, we consider the
price deviations from the WE, measured in terms of the Mean Absolute Percentage
Error (MAPE) of good 1, calculated based on the last 500 rounds of the simulation.
The MAPE of the homogeneous economies, each with distinct values of λ is shown
in the left panel of Fig. 1.
For the second, there are at least two ways to compare a heterogeneous economy
(HE) (where 50% of the agents have λ = 1 and the other half have λ = 7) with
homogeneous economies. Let M H E represent the MAPE value in the heterogeneous
economy. Let MiH O be the MAPE in a homogeneous economy, where all the agents
have the same intensity of choice, i and Ω is the weighting constant. We can then
consider an average of the MAPE values between the homogeneous economies with
λ = 1 and 7, respectively. i.e.,
HO
M̄1,7 = ΩM1H O + (1 − Ω)M7H O (19)

For the case of a simple average that weights the MAPE values of the two
economies equally, we have:

0.8
0.25
0.7
MAPE (Price of Good 1)

0.2
0.6
MAPE (good 1)

0.5 0.15
0.4
0.1
0.3

0.2 0.05
0.1
0
Hetero Avg Rep
0 1 2 3 4 5 6 7 8 9 10
Economies
Intensity of Choice

Fig. 1 Homogeneity vs. heterogeneity and aggregation: the above figures compare the values
of the Mean Absolute Percentage Error (MAPE) of good 1 across different economies. The
MAPE is calculated based on the last 500 periods of each run and it is defined as MAPE(Pj ) =
2500 ∗ ∗
t=2001 | Pj (t) − Pj | (j = 1, 2), where Pj = 1. The left panel indicates the price deviation
1
500
for homogeneous economies according to λ. The right panel shows the comparisons between the
heterogeneous and relevant homogeneous economies. The three economies that are compared are:
(1) an economy consisting of agents with heterogeneous intensities of choice (Hetero), (2) an
economy constructed by averaging the MAPE values of two homogeneous economies with λ = 1
and 7(Avg), and (3) a homogeneous economy with agents holding an ‘average’ value of intensity
of choice (i.e., λ = 4, which is the average of 1 and 7.).(Rep). The boxplot shows the variability
of the MAPE across 50 different runs for each economy
Heterogeneity, Price Discovery and Inequality 127

M1H O + M7H O
HO
M̄1,7 = (20)
2
Another option would be to choose a single, homogeneous economy (Rep) that
is representative, given the λ values in the heterogeneous counterpart. Instead of
averaging the values of the two economies as shown above, we can choose MAPE
values corresponding to the homogeneous economy with λ = 4, which is the
midpoint of 1 and 7. We now have three different ‘economies’ to compare: (1)
An economy with heterogeneous intensities of choice—λi = 1, 7(Hetero), (2) an
economy representing the average MAPE values of two homogeneous economies
with λ = 1 and 7(Avg), and (3) a representative homogeneous economy with agents
holding an ‘average’ value of intensity of choice (i.e., λ = 4, the average of 1
and 7) (Rep). The boxplot in the right panel of Fig. 1 compares the MAPE values
across these three versions. The whiskerplot shows the variability of MAPE across
50 different runs for each economy.
We find that the values of MAPE corresponding to the heterogeneous version are
remarkably different from the other two versions in terms of the median values and
its dispersion across runs. In particular, MAPE values of the Avg economy are much
higher compared to the Hetero economy, indicating the absence of linear scaling
up. From the right panel of Fig. 1, we observe that the MAPE across homogeneous
economies is nonlinear with respect to λ. This partly explains the inadequacy of
the analysis that solely relies on the information from homogeneous economies to
understand the macro level properties. This nonlinearity, combined with potential
interaction effects, indicates the breakdown of a linear scale-up and merits an
independent investigation of the heterogeneous Scarf economy.

5.2 Price Convergence

In this section, we explore the ability to coordinate and steer the economy to the WE,
starting from out-of-equilibrium configurations. To do so, we simulate the agent-
based Scarf economy outlined in the previous section, where agents adaptively learn
from their experiences using a meta-learning model. Given that the heterogeneity
amongst agents is characterized in terms of their varying intensities of choice, we
explore the influence of this parameter in detail. We simulate an economy with
parameters indicated in Table 1. The agents differ in terms of the intensity of choice
parameter (the exploration parameter, λ) and the intensities are uniformly distributed
across agents8 as λi ∈ U (0, 8). The bounds of the said uniform distribution are
chosen based on the initial insights gained from simulations with homogeneous
agents (in terms of λ), where the behaviour of the economy is qualitatively similar

8 The granularity of λ values is chosen in increments of 0.1, and hence there are 81 distinct values

of λi corresponding to U(0,8).
128 Shu-Heng Chen et al.

2.5 2.5

2 2

1.5 1.5

Price
Price

1 1

0.5 0.5

0 0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Intensity of Choice Intensity of Choice

Fig. 2 Price convergence and heterogeneity: the above figure indicates the association between
the mean prices for good 1 among different agents and their associated intensities of choice.
The intensity of choice ranges between 0 and 8 and is distributed uniformly across agents in
the economy, i.e., λi ∈ U (0, 8). The prices of goods 1 and 2 reported here are the mean
market prices for each agent which are calculated based on the last 500 periods of each run, i.e.,
2500
Pij = 500
1
t=2001 Pij (t), (j = 1, 2). The mean prices of some agents in the economy diverge
away from the neighbourhood of the WE prices (Pij∗ = 1, for j = 1, 2) for values of intensity of
choice greater than 3

for higher values of the parameter. We simulate the economy for 2500 periods for
each run and conduct 50 such repetitions. The data on prices from these repetitions
are pooled together and analysed in order to obtain robust results.
Figure 2 shows the relationship between the prices of goods 1 and 2 among
different agents and their corresponding intensities of choice. The prices of goods 1
and 2 reported here are the mean market prices for each agent, which are calculated
2500
by considering the last 500 periods of each run, i.e., Pij = 500
1
t=2001 Pij (t), (j =
1, 2). The agent-level data on mean prices across 50 repetitions are pooled together
and plotted. We find that these mean prices resulting from the interaction and
adaptive learning reveal some distinctive patterns. In particular, Fig. 2 shows a range
of values of intensity of choice, roughly between 0 and 3, for which the agents
remain in the neighbourhood of the WE. Beyond this range, the agents with higher
values of λ do not necessarily have their mean prices confined to the neighbourhood
of the WE. Even though some agents stay in the neighbourhood of the WE for
λi ∈ (3, 8), we observe a greater dispersion of mean prices on both sides of the WE.
This is illustrated better in terms of the 3-D plot (Fig. 3) where the frequency of
agents corresponding to different prices is shown on the z-axis. Although there is a
dispersion on both of its sides, the WE still remains the predominant neighbourhood
around which the mean prices gravitate. We also observe that there is a sizeable
number of agents (more than 100) who hover close to zero as their mean price for
the values of intensity of choice approaches 8.
Notice that the distribution of the intensity of choice among agents is uniform
and that the distribution is split between two regions, one with less dispersion in
terms of mean prices with the WE as the centre and the other with higher dispersion
(λ ∈ (4, 8)). We test whether the WE remains as the predominant neighbourhood
Heterogeneity, Price Discovery and Inequality 129

800

700

600
800

600 500
Frequency

400 400

200 300

0
200
8
6 100
4 2.5
2
1.5 0
2 1
0.5
Lambdas 0 0 Prices

Fig. 3 Intensity of choice and price convergence for λi ∈ U (0, 8): the above figure shows
both the prices of good 1, the frequency of agents with those prices and the associated intensities
of choice of each agent, where λi ∈ U (0, 8). The prices of good 1 reported here are the mean
market prices for each agent, calculated based on the last 500 periods of each run, i.e., Pi1 =
1 2500
500 t=2001 Pi1 (t). We carry out 50 repetitions of the simulation and pool together the data from
all repetitions on the mean prices for all agents (270 × 50 = 13,500) for the above plot. The agents
are grouped into eight equally spaced bins based on their intensity of choice and the mean prices
are grouped into 50 equally spaced bins

with which heterogeneous agents coordinate by altering the proportion of agents

between high and low dispersion regions in Fig. 2. To do so, we alter the bounds of
the uniform distribution according to which agents are assigned different intensities
of choice in the economy. We increase the upper bound of this distribution to 15,
λi ∈ U (0, 15) and examine how the mean prices vary according to λ. Figure 4
illustrates this relationship. We find that the results are qualitatively similar and
that the WE continues to be the attraction point for the agents with an increasing
dispersion with increasing λ. As in the previous case, a group of agents seems to
cluster close to zero, which seems to be the dominant attraction point other than the
WE for the agents in the economy.
In addition to the normal distribution, we investigate the behaviour of the prices
in economies with alternative distributional assumptions concerning the intensity of
choice. Table 2 summarizes the prices of goods 1 and 2 at the aggregate level and
at the level of agent types for 2 variations of normal and uniform distributions.
For the case where λi ∈ U (0, 8), the aggregate prices do stay within the 6%
neighbourhood around the WE. They remain in a slightly larger neighbourhood
(8%) for λi ∈ U (0, 15), albeit with relatively higher variations among different
runs. In the case of a normal distribution where λi ∈ N (4, 1), the mean intensity
of choice among agents lies where the increasing dispersion of prices takes off. The
balance of stable and unstable tendencies explains the extremely high proximity of
130 Shu-Heng Chen et al.

300
400
250
300
Frequency

200
200
150
100

100
0
15
10 3 50
2
5 1 0
Lambdas 0 0
Prices

Fig. 4 Intensity of choice and price convergence for λi ∈ U (0, 15): the above figure shows
both the prices of good 1, the frequency of agents with those prices and the associated intensities
of choice of each agent, where λi ∈ U (0, 15). The prices of good 1 reported here are the mean
market prices for each agent, calculated based on the last 500 periods of each run, i.e., Pi1 =
1 2500
500 t=2001 Pi1 (t). We carry out 50 repetitions of the simulation and pool together the data from
all repetitions on the mean prices for all agents (270 × 50 = 13,500) for the above plot

Table 2 Prices of Goods 1 and 2 are shown at the aggregate level and according to the types of
agents
Aggregate Type 1 Type 2 Type 3
Prices/Distrib. P1 P2 P1 P2 P1 P2 P1 P2
U [0,8] 0.943 0.995 0.889 0.922 0.873 0.931 1.068 1.132
(0.004) (0.014) (0.152) (0.133) (0.173) (0.148) (0.008) (0.026)
U [0,15] 1.080 1.044 0.924 0.888 0.948 0.925 1.370 1.320
(0.014) (0.030) (0.324) (0.280) (0.280) (0.299) (0.086) (0.072)
N (4,1) 1.000 1.001 0.932 0.949 0.939 0.928 1.130 1.127
(0.004) (0.005) (0.169) (0.152) (0.166) (0.174) (0.016) (0.008)
N(8,1) 1.407 1.510 1.197 1.296 1.109 1.270 1.914 1.964
(0.089) (0.008) (0.404) (0.573) (0.571) (0.528) (0.434) (0.019)
The prices reported are the mean market prices calculated based on the last 500 periods of each run,
2500
which are averaged over all agents. For the aggregate case, Pij = N1 N 1
i=1 500 t=2001 Pij (t),
where j = 1, 2 and N = 270. For the type wise prices, the mean prices are averaged over 90
agents. We carry out 50 repetitions of the simulation and the standard deviations across 50 runs are
indicated in the parenthesis

aggregate prices to the WE in this case. However, when we place the mean of λi
in the normal distribution at 8, which is in the unstable region (i.e., λi > 4), the
aggregate prices can be as much as 51% more than the WE.
Remember that agents with excess supply have to adjust the price expectations
of production goods downward and adjust the price expectations of consumption
goods upward (Eq. (10)). The upward bias in the divergence of prices stems from the
Heterogeneity, Price Discovery and Inequality 131

fact that the agents endowed with the numerairé good who are dissatisfied can only
adjust the price expectations of consumption goods upward and, unlike the agents
endowed with other two goods, they cannot adjust their prices of their endowed
goods downward. Those other agents with unsatisfied demand naturally bid up their
prices for the goods in question. Therefore, all commodities receive more upward-
adjusted potentials than downward-adjusted potential (2/3 vs 1/3 of the market
participants), except for commodity 3, which serves as a numéraire. Hence, there
is a net pulling force for the prices of commodities 1 and 2, leading them to spiral
up.
Although we observe a high variation in mean prices among agents for larger
values of λi , we still observe that a huge proportion of agents continue to remain
in the neighbourhood of the WE for these high values of λi (see Figs. 3 and 4).
This motivates us to search for other discriminating factors that might shed light
on the dispersion of prices. In order to see whether the learning strategies chosen
by the agents, viz., innovation or imitation, hold any advantage in enhancing price
coordination, we classify the agents based on their strategy choices during the course
of the simulation. We define Normalized Imitation Frequency (NIF), which indicates
the ratio of the number of periods in which an agent has been an imitator over
the entire course (2500 periods) of the simulation. If the agent has almost always
been an innovator throughout the simulation, his NIF will be close to zero. On the
contrary, if the agent has been an imitator throughout, the NIF will be equal to 1 and
the intermediate cases lie between 0 and 1. The mean prices of good 1 held by the
agents are plotted against the NIF in Fig. 5. The left and right panels in the figure
correspond to the distributions λi ∈ U (0, 8) and λi ∈ U (0, 15), respectively. The
entries in red correspond to agents with λi > 4 and those in blue denote agents with
λi < 4.
We observe that the there are two distinct clusters for agents in red (λi > 4).
Agents with high intensities of choice who happen to be ‘almost always innovators’
(with an NIF close to 1) can be seen to remain in the neighbourhood of WE, while
the high dispersion in prices seems to be coming from innovating agents with a
high λi . In our meta-learning model characterized by the reinforcement learning
mechanism, agents reduce their tendency to explore for higher values of λi and get
locked into one of the two learning strategies. These ‘immobile agents’ populate
either corner of the NIF scale. Figure 5 indicates that the price strategy to which
agents get locked in while cutting down their exploration seems to have a decisive
impact on whether or not agents can coordinate themselves in the neighbourhood of
the WE.
To summarize, in this section, we find that the ability of the agents to steer
themselves toward the WE on aggregate depends on their tendency to explore that
is captured by the intensity of choice. The distributional assumptions concerning
the intensity of choice among agents influence whether or not the aggregate prices
are closer to the WE. More specifically, the relative proportion of agents with λi
corresponding to less and more volatile regions, roughly on either side of λi = 4
is crucial and the larger the proportion of the latter, the higher is the deviation from
the WE in the aggregate. Despite the high variability in prices among agents with
132 Shu-Heng Chen et al.

3
2.5

2.5
2
2
1.5
Price

Price
1.5

1
1

0.5 0.5

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Imitation Frequency Imitation Frequency

Fig. 5 Mean price of Good 1 and imitation frequency (normalized): the figure indicates the price
distribution among all agents and their associated imitation frequencies. The intensity of choice
(λ) is distributed uniformly across agents in the economy and the left and the right panel denote
economies with distributions λi ∈ U (0, 8) and λi ∈ U (0, 15). Agents with λi ≤ 4 are denoted
by blue and those in red denote agents with λi > 4. The prices of good 1 reported here are the
mean market prices for each agent, calculated based on the last 500 periods of each run, i.e.,
2500
Pi1 = 5001
t=2001 Pi1 (t). We carry out 50 repetitions of the simulation and pool together the data
from all repetitions on mean prices for all agents (270 × 50 = 13,500) for the above plot. The
imitation frequency (normalized) denotes the ratio of the number of periods in which an agent has
been an imitator over the entire course (2500 periods) of each repetition of the simulation

a high λ, a sizeable number still remain in the neighbourhood of the WE. Agents
who do not explore and remain predominantly as innovators for all periods exhibit
a higher dispersion in their mean prices.

5.3 Current and Accumulated Payoffs

In this section, we investigate whether there are any significant differences amongst
agents in terms of their current and the accumulated payoffs.9
Figure 6 shows the relationship between the accumulated payoffs (or ‘lifetime
wealth’) of agents and their corresponding intensity of choice (λi ). While there is
a noticeable variability in accumulated payoffs for agents with higher intensities
of choice (λi > 4), however, λ alone does not seem to sufficiently explain the
cause. For instance, by increasing λi beyond 4, some (but not all) agents seem to
be increasingly worse off in terms of their accumulated payoffs that keep falling.
However, some other agents with the same values of λi enjoy much higher payoffs,

9 In
our model, there is no material wealth or savings since all goods are consumed entirely within
each period. However, it is possible to interpret the accumulated payoffs of each agent over the
simulation period as a proxy for their ‘quality of life’ or ‘lifetime wealth’.
Heterogeneity, Price Discovery and Inequality 133

1200

1000
Accumulated Payoffs

800

600

400

200

0
0 1 2 3 4 5 6 7 8
Intensity of Choice

Fig. 6 Accumulated payoffs: the figure shows the accumulated payoffs for all agents (i), where
λi ∈ U (0, 8). The variability in accumulated payoffs is not entirely explained by variations in their
intensity of choice alone. Notice that there are remarkable deviations between agents in terms of
their accumulated payoffs for λi > 4

which are comparable to those with λi <. In other words, there is a wealth inequality
that emerges among agents over time. We need to identify the characteristics of
agents that are responsible for these payoff differences.
Since our agents are similar in terms of all relevant characteristics other than
λ, potential indirect channels of influence stemming from λ need to be examined.
Strategy choices by agents—to innovate or imitate—over the period constitute one
such channel since they are governed by the λ parameter in the meta-learning
(reinforcement) model. In the previous section, we learnt that imitating agents
managed to coordinate to the WE on average, despite having a high λ. Along the
same lines, we examine whether imitation and innovation also impact the payoffs
among agents. Figure 7 shows how the current payoffs of agents vary according
to their NIF for λi ∈ U (0, 8) and λi ∈ U (0, 15) in the left and right panels,
respectively. Agents with λi ≤ 4 are shown in blue and those with λi > 4 are
shown in red. This indicates that agents who are predominantly innovators (for a
high λ) seem to have high variability among themselves in terms of their current
payoffs and agents who are predominantly imitators seem to enjoy relatively higher
payoffs with considerably less variability.
In Fig. 7 we infer a clearer idea of the underlying phenomena. The left and right
panels denote the relationship between intensity of choice and normalized frequency
of imitation for λi ∈ U (0, 8) and λi ∈ U (0, 15), respectively. As the intensity
of choice increases, there is a polarization in terms of NIF: for lower values of
λ (roughly, 0<λ < 3), NIF values span a wide range and they are not clustered
close to the end of the interval (0,1). These intermediate values indicate that agents
134 Shu-Heng Chen et al.

1
1 0.4
0.4
0.35 0.8
Imitation Frequency

Imitation Frequency
0.8
0.3 0.3
0.25 0.6
0.6
0.2 0.2
0.4
0.4 0.15
0.1 0.2 0.1
0.2
0.05
0 0
0 0
0 1 2 3 4 5 6 7 8 0 5 10 15

Intensity of Choice Intensity of Choice

Fig. 7 Intensity of choice, imitation frequency and current payoffs: the figures in the left and
right panels denote the relationship between λ and the normalized mean value imitation frequency
(NIF) for all agents based on 50 repetitions, corresponding to λi ∈ U (0, 8) and λi ∈ U (0, 15),
respectively. NIF values of 0 and 1 indicate that the agent is an innovator and imitator for 100%
of the time, respectively. The heat map denotes the associated mean current payoffs, which are
calculated based on the last 500 periods

switch between their strategies (innovation and imitation) and are ‘mobile’, thus
balancing between exploration and exploitation. By contrast, for λ values of 4 and
beyond, we observe that there is a drastic fall in the number of agents and the
agents are either predominantly innovators or imitators. This phenomenon can be
explained by the power law of practice that is well recognized in psychology.10 The
heat map in Fig. 7 indicates that the distribution of current payoffs varies among
the agents. Mobile agents (with intermediate values of NIF) and agents who are
predominantly imitators enjoy higher current payoffs compared to agents who are
predominantly innovators (close to the X-axis). This pattern persists despite changes
in assumptions regarding the distributions indicated above.

5.3.1 Accumulated Payoffs

We now turn to the differences among agents in terms of their accumulated payoffs
that we observed in Fig. 6. We group agents into different groups based on their
accumulated payoffs and analyse specific characteristics associated with agents in
each of these groups that could explain the differences in their accumulated payoffs.
Figure 8 shows the agents clustered into four different groups (Very High, High,
Medium, Low) based on their accumulated payoffs for λi ∈ U (0, 8) and U (0, 15)
in the left and right panels, respectively. We use the K-means technique (Lloyd’s
algorithm) to group the agents, which allows us to group agents into K mutually

10 Inthe psychology literature, the power law of practice states that the subjects’ early learning
experiences have a dominating effect on their limiting behaviour. It is characterized by initially
steep but then flatter learning curves.
Heterogeneity, Price Discovery and Inequality 135

1200 1200

1000 1000
Accumulated Payoffs

Accumulated Payoffs
800 800

600 600

400 400
Group1 Group1
Group4 Group4
200 Group2 200
Group2
Group3 Group3
0 0
0 1 2 3 4 5 6 7 8 0 5 10 15
Intensity of Choice Intensity of Choice

Fig. 8 Accumulated payoffs—clustered: the figures on the left and right show the accumulated
payoffs for all agents (i), where the intensities are distributed uniformly as λi ∈ U (0, 8) and
λi ∈ U (0, 15), respectively. Based on their accumulated payoffs, the agents are clustered in four
distinct groups using the K-means clustering technique

Table 3 Accumulated payoffs of heterogeneous agents: table compares four different clusters of
agents, who are pooled from 50 different runs of the simulation
U(0,8) U(0,15)
N Π AIF λ̄ N Π AIF λ̄
Low 891 320.29 24.797 7.175 2006 123.71 9.485 11.266
Medium 902 558.42 53.396 6.038 725 383.83 40.662 6.906
High 4329 805.11 1215.298 4.127 4911 714.08 1243.968 7.322
Very High 7378 957.85 965.453 3.311 5858 968.42 1607.920 6.293
These are the number of agents in each group (N) and the group-wise averages of the accumulated
payoffs (Π ), intensity of choice (λ̄), and imitation frequency (AIF)

exclusive clusters based on the distance from the K different centroids. We use
K = 4 based on visual heuristics and it provides the best classification for the
intended purpose. Higher values of K do not show improvements in terms of their
silhouette values that measure the degree of cohesion with other points in the same
cluster compared to other clusters.
We showed earlier that eventual strategy choices associated with a high λ
could explain variations in current payoffs. We examine whether agents (with high
intensity of choice) who are predominantly imitators end up with significantly
higher accumulated payoffs compared to innovating agents. Table 3 shows the four
different groups consisting of agents pooled from 50 different runs of the simulation.
The table also indicates the number of agents in each group and the group-wise
averages of accumulated payoffs, intensity of choice, and imitation frequency.
Agents with relatively better average accumulated payoffs for the two high groups
differ significantly in terms of their (AIF) compared to those for the medium and
low groups. Average Imitation Frequency measures the average number of periods
during which agents have been imitators in the 2500 periods of the simulation.
136 Shu-Heng Chen et al.

From Fig. 1, we see that agents who have λ > 4 are those that constitute the
medium and low payoff clusters. From Table 3, by comparing low and high clusters
in U(0,8), we observe that agents with higher AIF (predominantly imitators) have
relatively higher accumulated payoffs. However, the significance of AIF comes into
play largely for agents with λ > 4. For lower values of λ the strategy choices are not
tilted predominantly towards innovation or imitation. In this range, both innovating
and imitating agents have a chance to obtain medium or high accumulated payoffs.
The strength of AIF as a discriminating factor becomes evident when comparing
different clusters in U(0,15), where the number of agents with a high λ are in
relative abundance. This relative abundance helps for a robust understanding of
payoff inequality among agents with a high λ. Table 3 denotes the monotonic and
increasing relationship between accumulated payoffs and AIF between clusters.
The average accumulated payoff for agents from the ‘Very High’ cluster is 968.4,
compared to 123.7 for those in the ‘Low’ cluster. Their corresponding AIF are 9.5
and 1607.9. Agents with abysmal accumulated payoffs for comparable λ values are
explained by their strategy choice—reluctance to imitate.

6 Discussion

We have examined the possibility of coordination among agents in a stylized

version of a decentralized market economy. We have analysed whether these agents
can reach the equilibrium (on average) in a Scarf-like economy starting from
disequilibrium states. In the presence of heterogeneity, Sect. 5.1 demonstrated that
it may not be adequate to look only at corresponding economies with homogeneous
agents and extrapolate patterns concerning price deviations. A straightforward
linear aggregation breaks down in our model due to potential interaction effects
and more importantly, due to the presence of a non-linear relationship between
intensity of choice—the sole source of heterogeneity in our model—and the MAPE.
The consequences of diversity among agents in shaping macroeconomic outcomes
can be readily investigated using agent-based models compared to equation-based
models.
Agents can and do succeed in coordinating themselves without a centralized
mechanism, purely through local interaction and learning. They learn to revise
private beliefs and successfully align their price expectations to those of the WE.
However, this is not guaranteed in all cases and the coordination to the WE critically
depends on a single behavioural parameter, viz., intensity of choice (λ). The level of
intensity indirectly captures the degree or the frequency with which agents engage
in exploration by trying different strategies. Lower levels of λ are associated with
a higher degree of exploration and vice versa. When the intensity of choice is zero,
it is equivalent to the case where agents choose between innovation and imitation
with equal probability. As the λ values rise, agents increasingly stick to the ‘greedy’
choice and explore less and less. As a consequence, we have fewer agents who are
Heterogeneity, Price Discovery and Inequality 137

switching between their strategies at higher λ values. They remain immobile and the
strategy choices are thus polarized with AIF being either 0 or 1.
Since agents differ only in terms of λ and coordination success is linked to λ,
the extent and nature of diversity among agents in an economy is crucial. It is
hard to speak about the possibility of success in coordination independently of the
distributional assumptions regarding λ. As pointed out in Sect. 5.2, if a relatively
large proportion of agents have high λ values, the coordination to equilibrium
is less likely to be successful, when compared to an economy with most agents
holding lower intensities of choice. In sum, we find that (a) the balance between
exploration and exploitation is crucial for each agent to reach the neighbourhood of
equilibrium prices, and (b) the level of heterogeneity (distributions of λ) among
agents determines the proportion of agents who explore little. This, in turn,
determines the extent of the average price deviation from the WE for the economy
as a whole (see Table 2).
The influence of a high λ (or ‘strong tastes’) of agents, although important, is
not the whole story. Similarly, sole reliance on meso-level explanations that look
at the fixed proportion of agents holding different strategies (market fraction) is
also inadequate in explaining the success or failure of coordination. From Fig. 5,
we find that individuals can reach the neighbourhood of the WE prices for a large
intermediate range (0 < N I F < 1) where the agents are mobile. We also find that
having a strong affinity to a particular strategy that performs well during the early
periods is not detrimental per se. Instead, the nature of the strategy to which an
agent is committed matters more. Notice from the figure that agents who are pure
imitators (N I F ≈ 1) also reach the neighbourhood of the WE, just as mobile agents
do. Thus, even with high intensity of choice, agents who explore (social learning)
are more effective in correcting their misperceptions unlike those who learn only
from their own experiences.
We observe that inequality emerges among agents in their accumulated payoffs.
There have been different explanations concerning the emergence of income and
wealth inequality among individuals and nations, such as social arrangements
that deny access to resources and opportunities to some groups vis-à-vis others,
differences in initial conditions like endowments or productivity [12], and financial
market globalization [15], to name a few. Other scholars see wealth and income
inequality as a by-product of the capitalistic system and due to differential rewards
to various factors of production [17].11 In addition to these explanations, we
point to another possible source of inequality over time: differences in learning
behaviour. Our analysis shows that payoff inequality can potentially stem from
diversity in expectations, which is traced back to a micro-level parameter that
governs the learning behaviour of agents. Even though agents have the same amount
of endowments, some become relatively worse off over time due to their choice of
learning strategy. Agents who do not explore enough and learn from others end up
receiving drastically lower payoffs and thus remain poor.

11 See also: [16].

138 Shu-Heng Chen et al.

7 Conclusion

We have examined the possibility of price discovery in a decentralized, hetero-

geneous agent-based Scarf economy and whether or not agents can coordinate
themselves to the WE starting from disequilibrium. We find that the coordination
success is intimately tied to learning mechanisms employed by agents and, more
precisely, how they find a balance between exploration and exploitation. A reduction
in or an absence of exploration has an adverse effect on accumulated payoffs and
coordination possibilities, endogenously giving rise to inequality. Social learning
has often been considered to be inferior and there has been a relatively bigger focus
on individual learning. It turns out that there is plenty that we can learn from others
after all. Neither of these learning modes is in itself sufficient to ensure efficient
outcomes and hence it may be necessary to balance them both. Although our model
is highly stylized, with the results lacking the desired generality, it brings us a step
closer towards comprehending the role of heterogeneity, and learning dynamically
shapes various aggregate outcomes.

Acknowledgements Shu-Heng Chen and Ragupathy Venkatachalam are grateful for the research
support provided in the form of the Ministry of Science and Technology (MOST) grants, MOST
103-2410-H-004-009-MY3 and MOST 104-2811-H-004-003, respectively.

References

1. Alós-Ferrer, C., & Schlag, K. H. (2009). Imitation and learning. In P. Anand, P. Pattanaik and
C. Puppe (Eds.), The handbook of rational and social choice. New York: Oxford University
Press.
2. Anderson, C., Plott, C., Shimomura, K., & Granat, S. (2004). Global instability in experimental
general equilibrium: The Scarf example. Journal of Economic Theory, 115(2), 209–249.
3. Arifovic, J., & Ledyard, J. (2004). Scaling up learning models in public good games. Journal
of Public Economic Theory, 6, 203–238.
4. Arthur, B. (1993) On designing economic agents that behave like human agents. Journal of
Evolutionary Economics 3(1), 1–22.
5. Bossan, B., Jann, O., & Hammerstein, P. (2015). The evolution of social learning and its
economic consequences. Journal of Economic Behavior & Organization, 112, 266–288.
6. Erev, I., & Rapoport, A. (1998). Coordination, “magic,” and reinforcement learning in a market
entry game. Games and Economic Behavior, 23, 146–175.
7. Erev, I., & Roth, A. (1998). Predicting how people play games: Reinforcement learning in
experimental games with unique, mixed strategy equilibria. American Economic Review, 88(4),
848–881.
8. Fisher, F. M. (1983). Disequilibrium foundation of equilibrium economics. Cambridge, UK:
Cambridge University Press.
9. Gintis, H. (2007). The dynamics of general equilibrium. Economic Journal, 117(523), 1280–
1309.
10. Hahn, F. H., & Negishi, T. (1962). A theorem on non-tâtonnement stability. Econometrica,
30(3), 463–469.
11. Hoppitt, W., & Laland, K. N. (2008). Social processes influencing learning in animals: A
review of the evidence. Advances in the Study of Behavior, 38, 105–165.
Heterogeneity, Price Discovery and Inequality 139

12. Hopkins, E., & Kornienko, T. (2010). Which inequality? The inequality of endowments versus
the inequality of rewards. American Economic Journal: Microeconomics, 2(3), 106–37.
13. Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk.
Econometrica, 47, 263–291.
14. Laland, K. (2004). Social learning strategies. Animal Learning & Behavior, 32(1), 4–14.
15. Matsuyama, K. (2004). Financial market globalization, symmetry-breaking, and endogenous
inequality of nations. Econometrica, 72(3), 853–884.
16. Piketty, T. (2000), Theories of persistent inequality and intergenerational mobility. Chapter 8,
Handbook of Income Distribution (vol. 1, pp. 429–476). Amsterdam: Elsevier.
17. Piketty, T. (2014), Capital in the Twenty-First Century. Cambridge: Harvard University Press.
18. Rendell, L., Boyd, R., Cownden, D., Enquist, M., Eriksson, K., Feldman, M. W., et al. (2010).
Why copy others? Insights from the social learning strategies tournament. Science, 328(5975),
208–213.
19. Roth, A., & Erev, I. (1995). Learning in extensive-form games: Experimental data and simple
dynamic models in the intermediate term. Games and Economic Behaviour, 8, 164–212.
20. Scarf, H. (1960). Some examples of global instability of the competitive economy. Interna-
tional Economic Review, 1(3), 157–172.
21. Schlag, K. (1998). Why imitate, and if so, how? A boundedly rational approach to multi-armed
bandits. Journal of Economic Theory, 78(1), 130–156.
22. Schlag, K. (1999). Which one should I imitate? Journal of Mathematical Economics, 31(4),
493–522.
23. Sutton, R., & Barto, A. (1998). Reinforcement learning: An introduction, Cambridge, MA:
MIT Press.
24. Tversky, A., & Kahneman, D. (1991). Loss aversion in riskless choice: A reference-dependent
model. Quarterly Journal of Economics, 106, 1039–1061.
25. Uzawa, H. (1960). Walras’ tâtonnement in the theory of exchange. The Review of Economic
Studies, 27(3), 182–194.
26. Vriend, N. (2000). An illustration of the essential difference between individual and social
learning, and its consequences for computational analyses. Journal of Economic Dynamics &
Control, 24(1), 1–19.
27. Wiering, M., & van Otterlo, M. (2012). Reinforcement learning: State of the art. Heidelberg:
Springer.
Rational Versus Adaptive Expectations
in an Agent-Based Model of a Barter
Economy

Shyam Gouri Suresh

Abstract To study the differences between rational and adaptive expectations,

I construct an agent-based model of a simple barter economy with stochastic
productivity levels. Each agent produces a certain variety of a good but can only
consume a different variety that he or she receives through barter with another
randomly paired agent. The model is constructed bottom-up (i.e., without a Wal-
rasian auctioneer or price-based coordinating mechanism) through the simulation
of purposeful interacting agents. The benchmark version of the model simulates
homogeneous agents with rational expectations. Next, the benchmark model is
modified by relaxing homogeneity and implementing two alternative versions of
adaptive expectations in place of rational expectations. These modifications lead to
greater path dependence and the occurrence of inefficient outcomes (in the form of
suboptimal over- and underproduction) that differ significantly from the benchmark
results. Further, the rational expectations approach is shown to be qualitatively and
quantitatively distinct from adaptive expectations in important ways.

Keywords Rational expectations · Adaptive expectations · Barter economy ·

Path dependence · Agent-based modeling

1 Introduction

The dynamic stochastic general equilibrium approach (“DSGE”) approach and its
representative agent (“RA”) version have been widely criticized on the basis of
empirical, theoretical, and methodological objections in a longstanding, large, and
rich body of literature. To name just three of numerous recent papers, [9, 21], and
[17], that summarize and explain many of the criticisms leveled against DSGE
models and advocate the use of agent-based modeling (“ABM”) as a preferred

S. Gouri Suresh ()

Davidson College, Davidson, NC, USA
e-mail: shgourisuresh@davidson.edu

© Springer Nature Switzerland AG 2018 141

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_7
142 S. Gouri Suresh

alternative. Various papers such as [14] and [18] have explicitly constructed models
to compare the two approaches in a variety of different ways. This paper also
compares the two approaches by constructing a basic illustrative benchmark DSGE
model embedded within an ABM framework that also allows for certain typical
DSGE assumptions to be selectively relaxed. Results confirm that changes in
assumptions regarding homogeneity and rational expectations lead to quantitatively
and qualitatively different outcomes even within a model designed to resemble a
simplified real business cycle (“RBC”) environment.
Of the many assumptions required in typical DSGE models, this paper focuses
broadly on three:
1. Rational expectations hypothesis (“REH”)—According to Frydman and Phelps
in [11, p. 8]:
REH presumes a special form of expectational coordination, often called the ‘rational
expectations equilibrium’ (“REE”): except for random errors that average to zero, an
economist’s model—‘the relevant economic theory’—adequately characterizes each
and every participant’s forecasting strategy.

However, notably, in the same chapter, [11, p. 2] the authors cite some of their
prior work to emphasize that
REH, even if viewed as a bold abstraction or approximation, is grossly inadequate for
representing how even minimally reasonable profit seeking participants forecast the
future in real-world markets.

The primary goal of this paper is to compare outcomes with and without
the assumption of rational expectations in an illustrative ABM framework by
constructing agents who could either have rational expectations or one of two
different types of adaptive expectations.
2. Homogeneous stochastic productivity levels—Stiglitz and Gallegati [21, p. 37]
argue that
First and foremost, by construction, [in the RA approach] the shock which gives rise
to the macroeconomic fluctuation is uniform across agents. The presumption is that
idiosyncratic shocks, affecting different individuals differently, would ‘cancel out.’ But
in the real world idiosyncratic shocks can well give rise to aggregative consequences. . .

More specifically, within the DSGE framework, it can typically be demonstrated

that if complete and frictionless insurance markets exist, only aggregate shocks
matter for macroeconomic consequences. However, in practice, the real world
includes a large number of uninsured idiosyncratic shocks. In order to examine
the consequences of this critical assumption, the model in this paper allows
shocks to either be homogeneous or heterogeneous across agents.
3. Walrasian equilibrium—[4] and the references included therein provide a
detailed and critical analysis of the concept of Walrasian equilibrium. In
particular, [22] states that four conditions (Individual Optimality, Correct
Expectations, Market Clearing, and the Strong Form of Walras’ Law) must
hold for Walrasian equilibrium and then proceeds to carefully construct a non-
Walrasian agent-based model by plucking out the Walrasian Auctioneer or
Rational Vs. Adaptive Expectations in a Barter Economy 143

the implicit price setting mechanism employed in the DSGE framework. The
model developed in this paper also does away with the Walrasian Auctioneer as
transactions between agents take place through barter. A decentralized trading
system has been implemented in [12] and various other influential papers in the
literature. This decentralized trading system is straightforward to implement in
an ABM framework and explicitly allows for situations where markets do not
clear or equilibrium is not attained in the sense that some bartering agents remain
unsatisfied with the volume exchanged.
The model in this paper is inspired in part by a rudimentary RBC model where
the primary source of fluctuations is a stochastic productivity shock. The reason for
choosing simplicity and parsimony is to illustrate how consequential various DSGE
assumptions can be even within elementary models. Despite its simplicity, the
model in this paper allows for the implementation of a typical rational expectations
equilibrium with homogeneous agents as well as the optional inclusion of hetero-
geneous random productivity levels, incomplete and decentralized markets with the
potential for disequilibria, adaptive expectations, price stickiness, and asymmetric
information. Results suggest that multiple equilibria plague the rational expectations
solution as is common in numerous papers in the extensive literature on coordination
failures—see, for instance, [5–7], and [2]. Under adaptive expectations, the model
has richer path dependence as well as convergence to a particular Pareto-inferior
rational expectations equilibrium. Moreover, some forms of adaptive expectations
are found to be particularly sensitive to the scale of the model.
The remainder of the chapter is organized as follows. Section 2 sets up the model
mathematically and describes some of its features. Section 3 provides results from
the analyses of various versions of the model and compares the outcomes of the
benchmark homogeneous version with the outcomes of various rational expectations
and adaptive expectations versions. The paper concludes with further discussion in
Sect. 4.

2 Model Development

In this model economy, although there is essentially only one good, it comes in two
varieties, red and blue. Half the agents in the economy produce the blue variety and
consume the red variety, whereas the other half of the agents in the economy produce
the red variety and consume the blue variety. The number of agents in the economy
equals N where N is restricted to be an even number. Each agent is endowed with
three units of time each period which the agent can allocate in discrete quantities to
leisure and labor. All agents receive disutility from labor and, as stated previously,
agents who produce the red variety derive utility from consuming the blue variety
and vice versa. The Cobb Douglas utility functions for red and blue producers are
given by (1) and (2):
144 S. Gouri Suresh

Ui = Bi α (3 − Li )1−α (1)
Uj = Rj α (3 − Lj )1−α (2)

The per period utility for Agent i, assuming that Agent i is a red variety producer,
is given by Ui , where Bi is the number of units of the blue variety consumed by
Agent i and Li is the number of units of time devoted by the Agent i towards labor.
The utility function for Agent j , a blue variety producer, is similar except for the
fact that Agent j derives utility from consuming the red variety. The parameter α
reflects the relative weight assigned to consumption over leisure and it is assumed
that agents accord a higher weight to consumption than leisure or 0.5 < α < 1.
Notice that no agent would ever devote all 3 of their units of time towards labor and
the specified preference structure reflects non-satiation.
Each agent conducts her own private production process that is unobservable to
other agents based on the following constant returns to scale Leontief production
functions, (3) and (4):

Ri = Min(Ai , Li ) (3)
Bj = Min(Aj , Lj ) (4)

The quantity of the red variety produced by the Agent i, (again, assuming that
Agent i is a red variety producer) is given by Ri , where Ai is the idiosyncratic
private stochastic productivity level experienced by Agent i, and Li once again is
the number of units of time devoted by the Agent i towards labor. The production
function for Agent j , a blue variety producer, is similar except for the fact that
Agent j is subject to her own idiosyncratic stochastic productivity level, Aj . The
stochastic productivity level can take on a value of either 1 (the low value) or 2 (the
high value) and evolves independently for each agent according to the same Markov
process described through the Markov transition matrix Π .
After the production process is completed privately, an agent who produces the
red variety is randomly paired with another agent who produces the blue variety. The
two agents barter as much of their produce as possible, at the fixed exchange rate
of one blue variety good for one red variety good. No further barter opportunities
are allowed for these two agents in the period and any excess goods they possess
that were not bartered are discarded. This barter process replaces the Walrasian
Auctioneer and allows for decentralized transactions. The barter process can be
represented by (5) and (6):

Bi = Min(Ri , Bj ) (5)
Rj = Min(Ri , Bj ) (6)

In other words, Bi , the quantity of the blue variety of good that Agent i, the red
variety producer obtains through barter, is equal to the lesser of Ri and Bj , the
quantities produced by Agent i and Agent j , respectively. Similarly, the quantity
Rational Vs. Adaptive Expectations in a Barter Economy 145

of the red variety of good that Agent j , the blue variety producer obtains through
barter, is also equal to the lesser of Ri and Bj , the quantities produced by Agent i
and Agent j , respectively.
In an analytically isomorphic formulation, all agents could be thought of as
producers of one of two colors of an intermediate good which are then perfect
complements in the production process for the final good with the production
function, F = 2 × Min(Ri , Bj ), and where the agents then subsequently split the
final good thus obtained equally between themselves. Alternatively, this model can
also be thought of as a discrete variant of the coconut model in [7] adapted for the
ABM framework.
Notice that there are two distinct sources of randomness in this model—the
first is the randomness associated with the stochastic productivity level and the
second is the randomness inherent in the pairwise matching for barter. The following
outcomes are of interest in this model: (1) per capita output or average output, (2)
per capita utility or average utility, (3) per capita overproduction or the average
quantity of goods discarded due to mismatches in barter (for example, if one agent
has produced 1 unit, while the other agent has produced 2 units then 1 unit produced
by the latter agent is considered overproduction), and (4) per capita underproduction
or the average difference between the quantity of goods actually produced and the
maximum quantity of goods that could have been consumed by a pair given the
stochastic productivity levels of its members (if either agent has a low productivity
level, then there is no underproduction because the most that each agent could have
consumed equals one; however, if both agents have high productivity levels, then
there is underproduction of one unit if one agent produces two goods and the other
produces one and there is underproduction of two units when both agents produce
only one good each).
Table 1 provides the value of the preference parameter α, the values that comprise
the Markov transition matrix Π , and the long-term probabilities implied by the
values in Π .1

Table 1 Parameter values Preference parameter Value

α 0.8
Markov transition matrix 
Future state
Low High
Low 0.6 0.4
Current state
High 0.1 0.9
Long-term probabilities due to 
Low 0.2
High 0.8

1 Resultsare somewhat quantitatively sensitive to the chosen parameters but the broader implica-
tions do not change. This will be discussed further in the next section.
146 S. Gouri Suresh

Table 2 lists all the theoretical outcomes that are possible for any pair of bartering
agents given the discrete nature of their stochastic productivity levels and labor
inputs.
The model constructed above is rather simple and unrealistic compared to typical
agent-based models. This, however, is a feature rather than a shortcoming since the
goal of this paper is not to replicate the real world but rather to demonstrate the
consequences of common DSGE assumptions within the ABM framework even in
a simple model that contains only the most basic RBC elements as specified by
Plosser in [20, p. 54]:
. . . many identical agents (households) that live forever. The utility of each agent is some
function of the consumption and leisure they expect to enjoy over their (infinite) lifetimes.
Each agent is also treated as having access to a constant returns to scale production tech-
nology for the single commodity in this economy. . . In addition, the production technology
is assumed to be subject to temporary productivity shifts or technological changes which
provide the underlying source of variation in the economic environment to which agents
must respond.2

Despite its obvious simplicity, the model contains some useful features:
1. Heterogeneous stochastic productivity levels—The model can be solved under
the assumption that each period all agents face the same stochastic productivity
level (i.e., idiosyncratic shocks are assumed to be eliminated through the
existence of perfect insurance) or under the assumption that each agents faces
an uninsurable idiosyncratic stochastic productivity level each period.
2. Private information—The model features agents who know their own stochastic
productivity level each period before they make their labor allocation decision but
are unaware of the stochastic productivity levels or the labor allocation decisions
of their potential bartering partners.
3. Decentralized markets—The model decentralizes markets and does away with
the construct of the Walrasian Auctioneer by featuring agents who engage in
barter.
4. Implicit price stickiness—The fixed one-for-one exchange rate of one variety for
another is equivalent to price stickiness. Alternatively, this fixed exchange rate
can also consider an outcome of the Leontief production function for the final
good that is then split equally between each pair of intermediate goods providers.
5. Markets that could fail to clear—Although in one sense markets clear trivially in
all barter economies, this model exhibits disequilibria in a deeper sense where
agents are unable to satisfy their desired demands and supplies and may regret
that they have produced too little or too much.

2 In addition, the RBC model described in [20] features capital in the production function as well as

the associated consumption-investment trade-off. Adding capital to the model in this paper would
not affect its central findings but would complicate the analysis considerably since agents would
need to base their own decisions not just on their expectations of others’ stochastic productivity
levels and labor allocation decisions but also on their expectations of others’ capital holdings. This
increase in dimensionality renders the problem computationally impractical if REH is assumed
along with heterogeneous shocks and a finite number (i.e., not a continuum) of distinct individuals.
Table 2 All possible outcomes for a pair of bartering agents
Situation Productivity Effort Output Consumption Utility Market
# Agent 1 Agent 2 Agent 1 Agent 2 Agent 1 Agent 2 Agent 1 Agent 2 Agent 1 Agent 2 Overproduction Underproduction clearing?
1 1 1 1 1 1 1 1 1 2(1−α) 2(1−α) 0 0 Yes
2 1 1 1 2 1 1 1 1 2(1−α) 1 0 0 Yes
3 1 1 2 1 1 1 1 1 1 2(1−α) 0 0 Yes
4 1 1 2 2 1 1 1 1 1 1 0 0 Yes
5 1 2 1 1 1 1 1 1 2(1−α) 2(1−α) 0 0 Yes
6 1 2 1 2 1 2 1 1 2(1−α) 1 1 0 No
7 1 2 2 1 1 1 1 1 1 2(1−α) 0 0 Yes
8 1 2 2 2 1 2 1 1 1 1 1 0 No
9 2 1 1 1 1 1 1 1 2(1−α) 2(1−α) 0 0 Yes
Rational Vs. Adaptive Expectations in a Barter Economy

10 2 1 1 2 1 1 1 1 2(1−α) 1 0 0 Yes
11 2 1 2 1 2 1 1 1 1 2(1−α) 1 0 No
12 2 1 2 2 2 1 1 1 1 1 1 0 No
13 2 2 1 1 1 1 1 1 2(1−α) 2(1−α) 0 2 Yes
14 2 2 1 2 1 2 1 1 2(1−α) 1 1 1 No
15 2 2 2 1 2 1 1 1 1 2(1−α) 1 1 No
16 2 2 2 2 2 2 2 2 2α 2α 0 0 Yes
Note: 2α > 2(1−α) > 1 because 0.5 < α < 1. In particular, since α = 0.8, 1.74 > 1.15 > 1
147
148 S. Gouri Suresh

The presence of these features allows this model to occupy a space in-between
standard DSGE models and various agent-based models. The next section analyzes
the results of this model under various alternative assumptions.

3 Results and Analyses

Many different versions of the model were analyzed in order to capture the role of
various assumptions. Each version with a discrete number of agents was simulated
over multiple periods multiple times using Monte Carlo techniques in order to
obtain statistically reliable interpretations. For the Monte Carlo simulations, 100
different random seeds (one for each simulation) were applied uniformly across
all heterogeneous agent models over 50 periods (the initial period was treated as
the zeroth period and dropped from the analysis). The version with a continuum
of agents could not be simulated and so theoretical results from that version are
included instead. The time series and ensemble means and standard deviations of
the various outcome measures for the various versions analyzed are presented in
Table 3.

3.1 The Benchmark Version: Homogeneous Agents

In this version of the model, all agents are assumed to be perpetually identical in
terms of their realized stochastic productivity levels. Every agent finds herself in
either situation 1 (20% of the time, on average) or situation 16 (80% of the time) of
Table 2. If we add the assumption of Pareto-optimality in the benchmark case, the
REH where every agent expends high effort when productivity levels are high and
low effort when productivity levels are low can be sustained.3 Consequently, there
is no overproduction, no underproduction, markets always clear, and the level of
utility is always the highest attainable in each and every period. Another important
feature of the homogeneous case is that outcomes are scale invariant and results
remain identical regardless of the number of agents (as long as the number of red
variety producers is equal to the number of blue variety producers, an assumption
that is held true throughout this paper).

3 In theory, non-Pareto optimal outcomes are possible with REH here. For instance, it would be
rational for all agents to expend low effort regardless of productivity levels with the expectation
that everyone else too expends low effort in every state.
Table 3 A comparison of time series and ensemble means and standard deviations for various model versions
Output Utility Over Under
per per production production
capita capita per capita per capita
Over 50 periods
Mean Std. Dev. Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.
Homogeneous agents Mean 1.81 0.38 1.63 0.22 0.00 0.00 0.00 0.00
Std. Dev. 0.09 0.08 0.05 0.05 0.00 0.00 0.00 0.00
Heterogeneous Symmetric 4 agents Strategy 1 Mean 1.00 0.00 1.15 0.00 0.00 0.00 0.63 0.33
agents rational (0) Std. Dev. 0.00 0.00 0.00 0.00 0.00 0.00 0.08 0.04
expectations
equilibrium Strategy 2 Mean 1.64 0.33 1.50 0.22 0.16 0.16 0.00 0.00
(65,535) Std. Dev. 0.07 0.03 0.02 0.01 0.03 0.01 0.00 0.00
Strategy 3 Mean 1.40 0.42 1.38 0.26 0.08 0.14 0.24 0.35
(32,489) Std. Dev. 0.10 0.04 0.06 0.02 0.03 0.02 0.05 0.03
Continuum Strategy 1 Mean 1.00 0.00 1.15 0.00 0.00 0.00 0.64 0.00
Rational Vs. Adaptive Expectations in a Barter Economy

(theoretical) Std. Dev. 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Strategy 2 Mean 1.64 0.00 1.50 0.00 0.16 0.00 0.00 0.00
Std. Dev. 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Adaptive 4 agents Mean 1.01 0.03 1.15 0.02 0.01 0.03 0.62 0.33

Across 100 Simulations

expectations Std. Dev. 0.02 0.06 0.01 0.03 0.01 0.03 0.08 0.03
based on 100 agents Mean 1.01 0.03 1.15 0.01 0.01 0.02 0.61 0.09
individual Std. Dev. 0.00 0.01 0.00 0.00 0.00 0.00 0.01 0.01
Adaptive 4 agents Mean 1.14 0.17 1.21 0.11 0.06 0.10 0.38 0.28
expectations Std. Dev. 0.20 0.17 0.12 0.10 0.04 0.04 0.17 0.11
based on 100 agents Mean 1.64 0.07 1.31 0.04 0.16 0.03 0.00 0.00
population Std. Dev. 0.02 0.01 0.01 0.01 0.01 0.00 0.00 0.00
149
150 S. Gouri Suresh

3.2 The REH Version with 4 Agents

The next model analyzed assumes REH and includes four agents.4 Though other
information sets compatible with REH are possible, in the version of REH assumed
here, the information set of each agent is assumed to consist of the actual past
productivities of all agents. With this assumption in place, all four agents are fully
aware of all four productivity levels in the previous period, the fixed transition
matrix Π determining the probabilities of current period stochastic productivity
level outcomes, and their own current period realized stochastic productivity level.
With REH, each agent can come up with a strategy that involves a certain action for
each possible state of affairs. For any agent, the most parsimonious description for
the current state of affairs from his own perspective (his state vector, in other words)
involves one unit of information for the agent’s own current period shock and four
units of information for the four individual shocks in the previous period. Each unit
of information involves one of two possible values (high or low). Therefore, an agent
can find himself in any one of 25 or 32 different states.5
A full description of an agent’s strategy requires specifying an action for each
possible state. Since an agent can choose an action of either high effort or low
effort for each state, the total number of possible strategies for an agent is 232
or 4,294,967,296. However, from the structure of the model we know that no
agent would find it optimal to put in high effort when their own current stochastic
productivity level is low and so in 16 of those 32 states, the optimal action is
necessarily that of low effort. This leads to each agent effectively having 216 or
65,536 potentially optimal strategies. If all four distinctly identifiable agents are
allowed to pursue different strategies, then the overall number of possible strategies
to be considered in the model is (65, 536)4 = 1.8447e + 19 if ordering of
strategies matters or 65,539×65,538×65,537×65,536
4! = 7.6868e + 17 if the ordering
of strategies is irrelevant. Since either of these is computationally intractable for
the purposes of this research paper, the analysis here is restricted to symmetric
REE (“SREE”) which can be thought of as REE with the additional requirement of
symmetry across agents. It bears repeating that despite the considerable computing
power of the microcomputer at my disposal, tractability in the model requires the
imposition of symmetry as an assumption. Methodologically, this calls into question

4A two-agent version was also analyzed but this version is qualitatively different since it does not
include the randomness associated with the pairing for the barter process. Results for this version
are available upon request.
5 An implicit assumption here is that all agents are distinctly identifiable. If such distinctions are

assumed away, then the most parsimonious description for the current state involves one unit of
information for the agent’s own current period shock (which can take on one of the two values),
one unit of information for the number of red producers who received high productivity levels in
the last period (which can take on three values—none, one, and both) and one unit of information
for the number of blue producers who received high productivity levels in the last period (which
can also take on three values). Thus if, we assume away distinctiveness, an agent can find himself
in any one of 2 × 3 × 3 = 18 different states.
Rational Vs. Adaptive Expectations in a Barter Economy 151

the reasonableness of assuming strict REH in the real world where symmetry is not
commonly observed and agents rely primarily on mental calculations rather than
supercomputers.
With the assumption of symmetry in place, the 65,536 strategies (numbered
0 through 65,535) were examined one-at-a-time using a simple programming
approach that tested whether an individual agent could expect a higher potential
utility if she alone were to deviate from the strategy under consideration. 64 of
the 65,536 potentially optimal strategies were found to be SREE, i.e., strategies
such that no individual agent could gain from deviating away from that strategy
if everyone else in the economy were following that same strategy.6,7 The full
set of 1.8447e+19 possible combinations of strategies was not explored, but the
examination of a random sample suggested the existence of numerous asymmetric
REE as well.8 As stated previously, the existence of multiple equilibria is a
common outcome in models with complementarities and coordination failures.
Although the various equilibria can be Pareto-ranked, if REH is followed strictly,
choosing between these equilibria is theoretically impossible, because each REE, by
definition, satisfies all REH requirements. Cooper [5] provides some suggestions on
how this problem could potentially be resolved through the use of history as a focal
point or through best response dynamics to out-of-equilibrium scenarios. Within the
SREE framework adopted here however, the latter is ruled out by definition while
the former is helpful to the extent that it suggests that any chosen SREE is likely to
continue unchanged across periods.
Table 4 contains the means and standard deviations of the various outcomes
of interest across the 64 different SREE strategies as well as the minimum and
maximum values of each outcome variable and the associated strategy numbers.
These results were obtained through Monte Carlo simulations where 100 different
random seeds (one for each simulation) were applied uniformly across all 64 SREE
strategies over 50 periods (the initial period was treated as the zeroth period and
dropped from the analysis). There were no readily discernible common properties of
these 64 different SREE strategies.9 On the other hand, results suggest that outcomes
differ greatly based on which of the multiple SREEs are assumed to be in place. In
particular, it can be seen that two contrasting strategies, Strategy #0 and Strategy
#65,535, are both SREEs and result in significantly different outcomes. Strategy #0
corresponds to all four agents always choosing low effort regardless of stochastic

6 Technically, since distinctiveness is unnecessary if SREE is assumed, the more parsimonious

method mentioned in footnote 5 can be adopted resulting in the need to explore only 29 or 512
strategies. As one of the exercises in this paper allowed for a sampling of asymmetric strategies,
all 65,535 possibilities were treated as distinct for comparability.
7 The number of SREEs obtained was robust to small changes in parameter values. However, the

number of SREEs obtained was affected by large changes in parameter values and even dropped
to as low as one (e.g., when the low productivity level was a near certainty, the only SREE was to
always provide low effort).
8 In the 2 agent version of the model, 16 REEs were found of which 8 were also SREEs.
9 Detailed results for all SREE strategies are available from the author upon request.
152 S. Gouri Suresh

Table 4 Summary of results across the 64 SREE strategies

Output per Utility per Overproduction Underproduction
capita capita per capita per capita
Strategy Strategy Strategy Strategy
Value code # Value code # Value code # Value code #
Mean 1.32 1.33 0.08 0.32
Std. Dev. 0.21 0.12 0.05 0.21
Min. 1.00 0 1.15 0 0.00 0 0.00 65,535
Max. 1.64 65,535 1.50 65,535 0.16 65,535 0.64 0

productivity levels, while Strategy #65,535 corresponds to all 4 agents choosing

low effort only when their own idiosyncratic stochastic productivity level is low.10
In order to allow at a glance comparisons across models, only three SREEs are
included in Table 3: Strategy #0, Strategy #65,535, and a randomly chosen SREE,
Strategy #32,489.
Using the terminology employed by Page [19] and developed further by Jackson
and Ken [13], the properties for the various versions of the model in this paper
can be studied in terms of state dependence, path dependence, and equilibrium
dependence.11 Based on this, the SREE with Strategy #0 displays independence
(i.e., outcomes and strategies are independent of the current period shocks and past
period shocks). In all the SREE cases other than Strategy #0, state dependence
exists in a trivial sense since the stochastic productivity level evolves according
to a Markov process and the decisions by the agents themselves depend only on
their own current state vectors. Consequently, both outcomes and decisions in these
SREE cases are only state dependent rather than path dependent. Also, unless an
existing SREE is perturbed, there is no change in the strategy pursued. Therefore,
the SREE version of the model can also be considered equilibrium independent
since the long-term evolution of the system does not change regardless of the
evolution of stochastic productivity levels.

3.3 The REH Version with a Continuum of Agents

In terms of scale, the REH model above becomes significantly less tractable
with additional agents. For instance, with six potentially asymmetric distinctly
identifiable agents, even after the simplification that no agent would put in high

10 Itcan be shown mathematically that regardless of how many agents are included in the model,
both of these contrasting strategies (i.e., 1. always choosing low and 2. choosing low only when
individual productivity level is low) will be equilibria for the chosen values of α and Π in this
model.
11 The concept of equilibrium here is the one adopted by Page [19] and is completely unrelated to

economic equilibrium or market clearing.

Rational Vs. Adaptive Expectations in a Barter Economy 153

effort in a situation where she has received a low productivity level, the number of
6
potential optimal strategies that need to be searched is (22 )6 which approximately
equals 3.9402 e+115 and is intractable. In general, the number of strategies that
N
would need to be examined with N distinct agents equals (22 )N if asymmetric
equilibriums are allowed and order matters (i.e., agents are distinctly identifiable)
N
(22 +N −1)!
or N if asymmetric equilibriums are allowed but the order of strategies is
N !(22 −1)!
irrelevant (i.e., the identity of individual agents does not matter). Even if attention is
restricted to SREE and agents are considered to be indistinct from each other, under
the most parsimonious formulation, the number of strategies that would need to be
N 2
examined equals 2( 2 +1) .
This combinatorial explosion of strategy space occurs because the information
set for each agent increases as the number of agents increases. With a larger
information set, each agent can fine tune her strategy based on many more possible
combinations of prior productivity levels across all the agents. Consequently, as the
number of agents increases, the number of possible SREE strategies also increases
explosively.
Interestingly, one possible way to escape this combinatorial explosion is to
assume a continuum of agents. With a continuum of agents, if we force SREE (i.e.,
require all agents to follow the exact same strategy), the information set in effect
collapses completely. The probability of encountering a high or low productivity
level agent with a set strategy no longer depends on the precise combinations of
prior stochastic productivity levels and remains constant across all periods. Each
agent now only needs to consider what to do when her own productivity is low and
what to do when her own productivity level is high since the state of the aggregate
economy is unchanged across periods due to the averaging over a continuum.
Thus, only four strategies are possible (expend high effort always, expend low
effort always, expend effort to match own productivity level, and expend effort
inversely to own productivity level) only two of which are reasonable (i.e., do
not involve expending high effort during a low productivity period). These two
equilibria remain for the values of α and Π chosen in this paper, and as can be
predicted on the basis of footnote 10 (page 152), these equilibria correspond to
always choosing low regardless of productivity levels and choosing low only when
individual productivity level is low. However, note that if we assume a continuum
of agents, entire functions of asymmetric equilibria are also possible.
Theoretical results with this assumption of a continuum of agents are available in
Table 3. It is easy to see on the basis of the discussion above why homogeneity and
continuums are often assumed in the REH literature for the purposes of tractability
even though neither specification accurately reflects real world economies where
a finite number of distinct agents interact. In fact, as Table 3 suggests, the
homogeneous case and the continuum case are very different from those of an
SREE such as Strategy # 32,489. In other words, assuming homogeneity or a
simplified continuum ignores many more equilibrium possibilities than merely
assuming SREE.
154 S. Gouri Suresh

Table 3 also highlights the difference between the homogeneous case and the
continuum case with Strategy #2. Intuitively, this difference arises because there
are no mismatched productivities in the homogeneous case, while they occur in
the latter. Also, note that the standard deviations are zero for the continuum case
because of averaging over a continuum of agents.12

3.4 The Adaptive Expectations (“AE”) Versions

In the AE model, agents are assumed to not know the underlying stochastic
model. In particular, it is assumed that agents do not know the value of Π and,
furthermore, do not know the values of the past and current stochastic productivity
levels of other agents. The only information that an agent in the AE model is
allowed to retain is information about the output produced in the previous period.
Two versions of the AE model are considered based loosely on the long-standing
literature that distinguishes between individual and social learning; for instance [23],
argues that there are fundamental distinctions between the two modeling approaches
with significant implications in terms of greatly differing outcomes based on
whether individual or social learning is assumed. Some summary comparative
results for individual and social adaptive learners are provided in Table 3.
1. Adaptive Expectations—Individual (“AE–I”): In this version, an agent naïvely
predicts that the individual she will be paired with in the current period
would be someone who has produced the same quantity as the individual she
encountered in the previous period. In other words, agents in this version form
their expectations solely on the basis of their most recent individual experience.
2. Adaptive Expectations—Population (“AE–P”): In this version, an agent who
produces a particular variety is aware of the economy-wide distribution of the
other variety produced in the previous period and predicts that distribution will
remain unchanged in the current period. In other words, agents in this version
form their expectations on the basis of the most recent population-wide outcome.
It must be noted that the agents in both these versions are goal-directed expected
utility maximizers who use information about their observable idiosyncratic shocks
and their naïve expectations about others actions to determine their optimal level of
effort. It is also assumed that all adaptive agents start with the high effort strategy
(in the first period, each agent expends high effort if her own stochastic productivity
level is high). A few interesting results emerge that highlight the contrast with the
SREE version. Although some of these differences can be found in Table 3, it is
easier to visualize these differences through graphs rather than summary statistics.

12 I would like to thank an anonymous referee for highlighting this point.

Rational Vs. Adaptive Expectations in a Barter Economy 155

2
Homogenous

SREE (Low)

1.6
SREE (32489)

SREE (High)
1.2
AE Individual

AE Population
0.8
1 9 17 25

Fig. 1 Ensemble averages of output per capita over 1000 simulations in a four-agent model

Figure 1 depicts ensemble averages of per capita output over time for AE–I, AE–
P, the high SREE, the low SREE, and the randomly chosen SREE. As can be seen,
the high SREE and the randomly chosen SREE bounce around throughout, while
AE–I and AE–P bounce around for a while before settling at 1 which equals the low
SREE. In other words, it appears that the low SREE is the eventual outcome of the
adaptive expectations process. Notice that none of the SREEs have any perceptible
trend, while both AE–I and AE–P have a perceptible trend that eventually settles
down at the low SREE.
In the AE–I case, the convergence to the low equilibrium occurs because each
time an agent trades with another agent who has produced a low level of output,
she decides to produce low output herself in the next period regardless of her
productivity level. This action of hers in turn causes the agent she interacts with
in the next period to produce low output in the period subsequent to the next. On
the other hand, if an agent encounters another agent who has produced a high level
of output, she will only produce a high level herself in the next period if she herself
also has a high outcome for her stochastic productivity level. The asymmetry in
this model is such that with adaptive expectations, the outcome is inexorably drawn
towards the low equilibrium due to the asymmetry embedded in effort choice: high
effort is optimal only when one’s own productivity level is high and one reasonably
expects to meet someone else who has produced a high level of output; low effort
is optimal both when one’s own productivity is low and or one expects to have a
trading partner who has produced low output.
The reason for the convergence of the AE–P case to the low equilibrium is
similar. As long as the population-wide distribution of outputs is above a certain
threshold, agents continue to adopt the high effort strategy (in which they expend
high effort whenever their own stochastic productivity level is high) because they
156 S. Gouri Suresh

reasonably expect to meet a high output agent the next period. However, as soon
as the population-wide distribution of outputs falls below a certain threshold, each
agent assumes that the probability of meeting a high output agent is too low to justify
expending high effort even when their own stochastic productivity outcome is high
the next period. As soon as this threshold is breached, all agents expend low effort in
the next period, leading to low effort remaining the adopted strategy for all periods
thereafter. When the population size is large, the likelihood of the population-wide
distribution of productivity levels being extreme enough to cause the distribution of
outputs to fall below the threshold mentioned previously is low. Consequently, as the
population size grows larger, the results of AE–P converge towards the continuum
case.
Various papers in the multiple equilibrium literature have a similar result where
one or more specific equilibria tend to be achieved based on best response dynamics
[6], genetic algorithms [1, 3, 15], various forms of learning [8, 10, 16], etc.
As can be seen from the discussion on AE–P above, the effect of scale depends
crucially on the modeling assumptions. While issues of dimensionality preclude the
analysis of all SREE in models with more than four agents, as discussed earlier,
the low SREE and the high SREE can be shown to exist regardless of the number of
agents. In order to examine the effects of scale more explicitly, consider Fig. 2 which
compares one sample run each (with different random evolutions of the stochastic
processes across the different models) for 4 agents, 12 agents, and 100 agents. From
the one run graphs in Fig. 2, it appears that scale matters a great deal for AE–P,
whereas scale does not matter for SREE or AE–I, except in the sense that variance
of outputs is predictably lower with more agents. This effect of scale is more clearly
visible if we look at the ensemble averages presented in Fig. 3.13 Scale matters to a
very limited extent for AE–I (being visible only when zoomed into the short run),
and a great deal for AE–P as we go from 4 agents to 100 agents. AE–P most closely
resembles AE–I in a model with 4 agents but by the time the economy is comprised
of 100 agents, AE–P starts to resemble SREE high. For the models with for 4 agents
and 8 agents, AE–P starts with a high output per capita that settles eventually to the
low SREE.
Based on the above results as well as the underlying theory, it can be seen that
unlike the SREE versions, the AE models are path dependent; it is not just the
productivity level in the last period that matters but also the specific sequence of past
stochastic productivity levels and past random pairings. However, continuing with
the terminology employed by Page [19], the AE models are equilibrium independent
in the sense that the long-run equilibrium converges with SREE low. This happens
fairly rapidly for AE–P with a small number of agents and AE–I in all situations.
For AE–P with a large number of agents, this could take a very long time indeed but

13 For all figures, the initial (zeroth) period is dropped from consideration.
Rational Vs. Adaptive Expectations in a Barter Economy 157

2.4

AE Individual
1.6 AE Population
SREE (Low)
SREE (High)

0.8
1 34 67 100

2.4

AE Individual

1.6 AE Population
SREE (Low)
SREE (High)
0.8
1 34 67 100

2.4

1.6 AE Individual
AE Population
SREE (Low)
SREE (High)
0.8
1 34 67 100

Fig. 2 Scale effects—single runs of models with 4 agents, 12 agents, and 100 agents

will indeed eventually occur.14 Assuming SREE as a simplification is equivalent to

ignoring short and intermediate run dynamics and those might matter in the real
world where the long-run equilibrium could indeed be a long way away, as in the
AE–P case with a large number of agents.

14 It
can be shown, for instance, that if all agents were to receive a low productivity level in any
period (a scenario with an extremely low but still nonzero probability in a model with many agents)
then everyone would switch to the low strategy in the next period.
158 S. Gouri Suresh

1.8
AE - Population (4)
AE - Population (8)
AE - Population (100)
1.5
AE - Individual (4)
AE - Individual (8)
AE - Individual (100)
1.2 SREE - High (4)
SREE - High (8)
SREE - High (100)
SREE - Low (4, 8, 100)
0.9
1 34 67 100

1.8 AE - Population (4)

AE - Population (8)
AE - Population (100)
1.5 AE - Individual (4)
AE - Individual (8)
AE - Individual (100)
1.2 SREE - High (4)
SREE - High (8)
SREE - High (100)
0.9 SREE - Low (4, 8, 100)
1 2 3 4 5

Fig. 3 Scale effects—ensemble means of models with 4 agents, 12 agents, and 100 agents

4 Further Discussion

In keeping with much of the literature surveyed in the introduction, the results
from this paper suggest that the assumptions of homogeneity, REH, and Walrasian
equilibria are far from innocuous. Even within the REH framework, assuming
continuums of agents and implementing SREE restrictions could lead to focusing
on a small subset of interesting results, even though those assumptions and
restrictions are extremely helpful in terms of providing tractability. Furthermore,
Rational Vs. Adaptive Expectations in a Barter Economy 159

results obtained through REH are not path dependent and could feature multiple
equilibria without a robust theoretical basis for choosing among them although
various alternatives have been proposed in the literature. AE–I and AE–P, on the
other hand, are path dependent and still eventually converge to certain SREE,
thereby allowing more precise and testable analyses of trajectories in the economy
through simulation-based event studies and Monte Carlo approaches. It should be
noted that the AE models employed here have a limitation. These AE models assume
that memory lasts only one period; changes in results caused by increasing memory
could be addressed in further work.15 Finally, SREE is scale free (although not in
terms of variance of output) whereas AE–I and AE–P are not (AE–P, in particular,
shows much greater sensitivity to scale) thereby allowing for richer models.
Overall, the exercises undertaken in this paper suggest that REE can indeed
be implemented in agent-based models as a benchmark. However, even a small
increase in the state space or the number of agents renders a fully specified REE
computationally intractable. ABM researchers interested in realism may find REE
benchmarks of limited applicability since modifying models to allow for tractable
REE could involve imposing severe and debilitating limitations on the model in
terms of the number of agents, their heterogeneity, the types of equilibria allowed,
and other forms of complexities.

Acknowledgements I would like to thank anonymous reviewers for their invaluable suggestions.
I would also like to acknowledge my appreciation for the helpful comments I received from
the participants of the 20th Annual Workshop on the Economic Science with Heterogeneous
Interacting Agents (WEHIA) and the 21st Computing in Economics and Finance Conference.

References

1. Arifovic, J. (1994). Genetic algorithm learning and the cobweb model. Journal of Economic
Dynamics and Control, 18(1), 3–28.
2. Bryant, J. (1983). A simple rational expectations Keynes-type model. Quarterly Journal of
Economics, 98(3), 525–528.
3. Chen, S.-H, John, D., & Yeh, C.-H. (2005). Equilibrium selection via adaptation: Using genetic
programming to model learning in a coordination game. In A. Nowak & S. Krzysztof (Eds.),
Advances in dynamic games. Annals of the international society of dynamic games (Vol. 7,
pp. 571–598). Boston: Birkhäuser.
4. Colander, D. (2006). Post Walrasian economics: Beyond the dynamic stochastic general
equilibrium model. New York: Cambridge University Press.
5. Cooper, R. (1994). Equilibrium selection in imperfectly competitive economies with multiple
equilibria. Economic Journal, 104(426), 1106–1122.

15 Intuitively,
with more memory, AE–I and AE–P may sustain the high effort equilibrium for
longer because individuals would take into consideration more periods with high outputs while
making decisions in subsequent periods. For instance, even in the period right after an economy-
wide recession, if agents possessed longer memories, they could revert to expending high effort
the next period if their own individual stochastic productivity level was high.
160 S. Gouri Suresh

6. Cooper, R., & Andrew, J. (1988). Coordinating coordination failures in Keynesian models.
Quarterly Journal of Economics, 103(3), 441–463.
7. Diamond, P. (1982). Aggregate demand management in search equilibrium. Journal of Political
Economy, 90(5), 881–894.
8. Evans, G., & Seppo, H. (2013). Learning as a rational foundation for macroeconomics and
finance. In R. Frydman & P. Edmund (Eds.), Rethinking expectations: The way forward for
macroeconomics (pp. 68–111). Princeton and Oxford: Princeton University Press.
9. Fagiolo, G., & Andrea, R. (2012). Macroeconomic policy in DSGE and agent-based models.
Rev l’OFCE, 124(5), 67–116.
10. Frydman, R. (1982). Towards an understanding of market processes: Individual expectations,
learning, and convergence to rational expectations equilibrium. The American Economic
Review, 72(4), 652–668.
11. Frydman, R., & Edmund, P. (2013). Which way forward for macroeconomics and policy
analysis? In R. Frydman & P. Edmund (Eds.), Rethinking expectations: the way forward for
macroeconomics (pp. 1–46). Princeton and Oxford: Princeton University Press.
12. Howitt, P., & Robert, C. (2000). The emergence of economic organization. Journal of
Economic Behavior and Organization, 41(1), 55–84
13. Jackson, J., & Ken, K. (2012). Modeling, measuring, and distinguishing path dependence,
outcome dependence, and outcome independence. Political Analysis, 20(2), 157–174.
14. Lengnick, M. (2013). Agent-based macroeconomics: A baseline model. Journal of Economic
Behavior and Organization, 86, 102–120.
15. Marimon, R., Ellen, M., & Thomas, S. (1990). Money as a medium of exchange in an economy
with artificially intelligent agents. Journal of Economic Dynamics and Control, 14(2), 329–
373.
16. Milgrom, P., & John, R. (1990). Rationalizability, learning, and equilibrium in games with
strategic complementarities. Econometrica, 58(6), 1255–1277.
17. Napoletano, M., Jean Luc, G., & Zakaria, B. (2012). Agent Based Models: A New Tool for
Economic and Policy Analysis. Briefing Paper 3 OFCE Sciences Po, Paris.
18. Oeffner, M. (2008). Agent-Based Keynesian Macroeconomics: An Evolutionary Model Embed-
ded in an Agent-Based Computer Simulation. PhD Thesis, Julius-Maximilians-Universität,
Würzburg.
19. Page, S. (2006). Path dependence. Quarterly Journal of Political Science, 1(1), 87–115.
20. Plosser, C. (1989). Understanding real business cycles. The Journal of Economic Perspectives,
3(3), 51–77.
21. Stiglitz, J., & Mauro, G. (2011). Heterogeneous interacting agent models for understanding
monetary economies. Eastern Economic Journal, 37(1), 6–12.
22. Tesfatsion, L. (2006). Agent-based computational modeling and macroeconomics. In D.
Colander (Ed.), Post Walrasian economics: Beyond the dynamic stochastic general equilibrium
model (pp. 175–202). New York: Cambridge University Press.
23. Vriend, N. (2000). An illustration of the essential difference between individual and social
learning, and its consequences for computational analyses. Journal of Economic Dynamics
and Control, 14(1), 1–19.
Does Persistent Learning or Limited
Information Matter in Forward Premium
Puzzle?

Ya-Chi Lin

Abstract Some literature explains the forward premium puzzle by the learning
process of agents’ behavior parameters, and this work argues that their conclusions
may not be convincing. This study extends their model to the limited information
case, resulting in several interesting findings: First, the puzzle happens when
the proportion of full information agents is small, even people make expectation
near rationally. Second, allowing the proportion of full information agents to be
endogenous and highly relied on the performance of forecasting, agents turn to
become full information immediately, and the puzzle disappears. These results
are similar in different learning gain parameters. Our finding shows that limited
information would be more important than learning, when explaining forward
premium puzzle. Third, the multi-period test of Fama equation is also examined
by the exchange rate simulated by learning in the limited information case. The
Fama coefficients are positive, and the puzzle will not remain. It is consistent with
the stylized facts in the multi-period version of Fama regression, which is found in
McCallum (J Monet Econ 33(1):105–132, 1994). Finally, we also find that if agents
rely on the recent data too much when forecasting, they tend to overreact on their
arbitrage behavior. The Fama coefficient deviates further from unity. People might
not benefit from having more information.

Keywords Limited information · Persistent learning · Market efficiency ·

Forward premium puzzle · Dual learning

Y.-C. Lin ()

Hubei University of Economics, Wuhan, China
e-mail: yachi.lin@hbue.edu.cn

© Springer Nature Switzerland AG 2018 161

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_8
162 Y.-C. Lin

1 Introduction

The forward premium puzzle is a long-standing paradox in international finance.

Foreign exchange market efficiency is in a status that exchange rates fully reflect
all available information. Unexploited excess profit opportunity does not exist when
market efficiency is hold. Therefore, the forward exchange rate should be the best
predictor of the future spot exchange rate. However, most empirical researches have
the opposite findings. The slope coefficient in a regression of the future spot rate
change on the foreign premium is significantly negative, which is expected to be
unity if market is efficient. If st is the natural log of the current spot exchange rate
(defined as the domestic price of foreign exchange), st+1 is the depreciation of the
natural log of the spot exchange rate from period t to t+1, i.e., st+1 = st+1 − st ,
and ft is the natural log of the one-period forward rate at period t. The forward
premium puzzle is examined by the following Fama equation:

 st+1 = β̂(ft − st ) + ût+1 (1)

The Fama coefficient β̂ is unity if the efficient market hypothesis holds. However,
in the majority of researches, β̂ is negative. It is what we called “forward premium
puzzle.”
Examining market efficiency by regressing future spot rate change on the forward
premium is based on the assumption that agents are rational and risk neutral. The
rejection of market efficiency has the following explanations. First, if agents are
risk averse, the forward exchange rate contains a risk premium. Hodrick [13] and
Engel [9] apply Lucas asset pricing model to price forward foreign exchange risk
premium, which shows future spot rate deviating from forward rate. Second, agents
may have incomplete knowledge about the underlying economic environment.
During the transitional period, they can only guess the forward exchange rates
by averaging several spot exchange rates that may possibly happen. This makes
systematic prediction errors even though they are behaving rational. It is what we
called “peso problem,” which is originally studied by Krasker [15]. Lewis [16]
assumes that agents update their beliefs about the regime shifts in fundamentals
by Bayesian learning. During the learning period, the forecast errors are systematic
and serially correlated. Motivated by the fact that even today only a tiny fraction
of foreign currency holdings are actively managed, [1] assume that agents may not
incorporate all information in their portfolio decisions. Most investors do not find
it in their interest to actively manage their foreign exchange positions since the
resulting welfare gain does not outweigh the information processing cost. It leads to
a negative correlation between exchange rate change and forward premium for five
to ten quarters. The puzzle disappears over longer horizons. Scholl and Uhlig [23]
consider agents are Bayesian investors, who trades contingent on a monetary policy
shock, and uses Bayesian VAR to assess the posterior uncertainty regarding the
resulting forward discount premium. Forward discount premium diverges for several
countries even without delayed overshooting. Forward discount puzzle seems to be
robust.
Does Persistent Learning or Limited Information Matter in Forward Premium Puzzle? 163

The third explanation is that agents behave far from rationality. Agents have
difficulty to distinguish between pseudo-signals and news, and then they take
actions on these pseudo-signals. The resulting trading dynamics produce transitory
deviations of the exchange rate from fundamental values [20]. Agents’ prospective
spot exchange rates are distorted by waves of excessive optimism and pessimism.
Black [3] calls these kinds of agents “noise traders.” Literatures explain forward
premium puzzle by irrational traders are as below. Gourinchas and Tornell [12]
suppose that investors are adaptive expectations and misperceive interest rate shocks
as transitory. In the following period, they find that interest rate turns out to be higher
than they first expected, which leads them to revise upward their beliefs about the
persistence of the original interest rate shock and triggers a further appreciation
of the dollar. Therefore, gradual appreciation is accompanied with increases in
interest rates, and also forward premium. Chakraborty and Evans [8] replace rational
expectations by perpetual learning, which assumes that agents in the economy face
some limitations on knowledge in the economy. Agents do not themselves know
the parameter values and must estimate them econometrically. They adjust their
forecast for the parameter values as new data become available over time. The model
generates a negative bias, and the bias becomes stronger when fundamentals are
strongly persistent or the learning gain parameter is larger.
Lewis [16] and Gourinchas and Tornell [12] assume that the unexpected shift in
fundamentals is initially unsure for agents, or is misperceived as temporary. Agents
learn about the changes in fundamentals but learning takes time. Based on such
model setting, the β̂ in Eq. (1) could be less than unity. On the other hand, [8] assume
that the fundamentals are observable by agents. The thing that unknown by agents
is the behavior parameters, which is adaptively learned by agents over time.
In this paper, we investigate whether forward premium puzzle remains under
the interaction of limited information flows and adaptive learning. There are three
reasons to examine forward premium puzzle under limited information flows. First,
although [8] provide the numerical evidences of forward premium puzzle under
adaptive learning, the puzzle is not discovered in the theoretical model. Refer to
Fig. 1, the Fama coefficient decreases in gain parameter, but it never turns to be
negative in the reasonable gain parameter range from 0 to 1, despite the value
of fundamental persistence. Second, the negative bias provided in the numerical
exercise decreases as the sample size increases, refer to Table 3 in [8]. The
numerical results are correlated to the sample size. Third, the analysis of [8] is
based on full information, and the simulated exchange rate always reflects full
market information. Therefore, they purely explain the impact of learning process on
forward premium puzzle, instead of the foreign exchange market efficiency. The full
information in [8] does not mean complete information since the adaptive learning
model implies model uncertainty, the imperceptibility of behavior parameters to
agents.
The limited information flows and learning model derive several interesting
findings. First, the simulated β̂ is negative when the proportion of full information
agents is small, even in small learning parameter and rational expectation. It implies
that the effect of limited information on asymptotic bias dominates that of persistent
164 Y.-C. Lin

Fig. 1 Theoretical plimβ̂ in

[8] for θ = 0.9 and
ρ = 0.9, 0.95, 0.99

learning. Second, the limited information could not only explain forward premium
puzzle, it is consistent with the stylized facts, which is found in [21]. The puzzle
will not remain in the multi-period test of Fama equation. Third, agents are free
to choose the full or limited information when forecasting. Then the proportion
of full information agents is endogenous, which is influenced by the information
or switching costs. The learning in forecasting rule is accompanied with the
learning in behavior parameters, which is called “dual learning.”İ When agents react
strongly to the relative performance of the forecasting rule by using full or limited
information, all agents switch to apply full information quickly. The proportion of
full information agents approaches to unity in a short period, and market efficiency
is hold. This result is similar for different gain parameters. Forward premium puzzle
happens when agents are uninformed. Finally, it is possible for agents to choose the
limited information forecasting rule rather than full information one. If agents rely
on the recent data too much, they tend to overreact on their arbitrage behavior. Then
agents might not benefit from too much information.
Our study is organized as follows. In Sect. 2 we present the impact of limited
information and learning on forward premium puzzle. The simulation is applied
by assuming the proportion of full information agents is exogenous. Section 3
introduces the dual learning into the model and investigates the forward premium
puzzle by simulation. Finally, Sect. 4 concludes.

2 The Theoretical Model

2.1 Monetary Model

To show the impact of limited information and adaptive learning on forward

exchange market efficiency, we use a simple monetary exchange rate model based
on purchasing power parity and uncovered interest parity.
Does Persistent Learning or Limited Information Matter in Forward Premium Puzzle? 165

st = pt − pt∗ (2)
it = it∗ + Et st+1 − st (3)

where pt is the log price level, and Et st+1 is the log of the forward rate at time t for
foreign currency at t+1. The variables with a star represent foreign variables. The
money market equilibrium in the home country is

mt − pt = c1 yt + c2 it (4)

where mt is the log money stock and yt is log real output. The money market
equilibrium is also held in the foreign country with identical parameters.
Combining the equations above, the reduced form of exchange rate can be solved.

st = θ Et st+1 + vt (5)

where θ = c2 /(1 − c2 ), and vt = (1 − θ )((mt − m∗t ) − c1 (yt − yt∗ )). Suppose that
vt follows a stationary first-order autoregressive process,

vt = ρvt−1 + εt (6)

where εt ∼ N (0, σε2 ), and 0 ≤ ρ < 1. By method of undetermined coefficients, the

rational expectation solution to this model is

st = bvt−1 + cεt (7)

where b = ρ(1 − ρθ )−1 and c = (1 − ρθ )−1 . Here risk neutrality is assumed, thus
the forward exchange rate equals to the market expectation of st+1 held at time t.

ft = Et st+1 (8)

Examining forward premium puzzle by using the exchange rate generated from
the rational expectation solution above, the puzzle does not exist.1 However, [8]
show that when replacing rational expectations by constant gain learning, a negative
bias exists in β̂. The result coincides with the concept of adaptive market hypothesis,
as proposed by Lo [17]. Even in an ideal world of frictionless markets and costless
trading, market efficiency still fails because of the violation of rationality in the
agents’ behavior, such as overconfidence, overreaction, and other behavioral biases.
Overconfident individuals overreact to their information about future inflation. The
forward rate will have larger response than the spot rate because that the forward rate
is more driven by speculation and the spot rate is the consequence of transaction

1 From the rational expectation solution, we could derive the following relationships,
cov(Δst+1 , ft − st ) = var(ft − st ) = (1 − ρ)σε2 /[(1 + ρ)(1 − ρθ)2 ]. Therefore, β̂ = 1, and the
puzzle does not exist.
166 Y.-C. Lin

demand for money [7]. Therefore the slope coefficient of the spot return on the
forward premium turns to below unity but still positive, as shown in Fig. 1.
Based on the result of [8], β̂ are negative only in the numerical simulation, but
it is always positive in the theoretical model. The forward premium puzzle is not
discovered theoretically, which represents that the puzzle may not be explained
thoroughly by simply adaptive learning. This study combines adaptive learning with
limited information flows, and we attempt to explain the puzzle by these two factors.

2.2 Limited Information Flow Model

The limited information flow model in our study is constructed by simplifying the
sticky information model in [18].2
Define It−j as the information set consists of all explanatory variables dated
t − j or earlier. We first assume that the latest information is not available for all
agents, but the explanatory variables dated t −1 or earlier are public information.
This assumption conforms to the economic reality that there may be information
costs or barriers when receiving the latest information. Once the latest information
becomes history and is announced, it is available to all agents. We also assume that
there exist two groups of agents, i = 1, 2. The first group make forecast with full
information, i.e., j = 0. The second group is blocked from the latest information,
and they could only make forecast by the past one, i.e., j = 1.3
e
s1,t+1 = E1 (st+1 |It ) = bt−1 vt (9)
e
s2,t+1 = E2 (st+1 |It ) = bt−2 E2 (vt |It−1 ) = bt−2 ρt−2 vt−1 (10)

where bt−j and ρt−j are estimates based on information up to time t-j.
At the beginning of time t, agents with full information have estimates bt−1 of
the coefficients b, based on information through t−1. Together with the observed
fundamentals vt , agents make the forecasting exchange rates for the next periods.
For agents not updating the information sets at time t, the forecasting exchange rate
is equivalent to the two-period forecasting at time t−1. The forecasting formula is
referred to [4] and [19]. The forecasting exchange rate for the next period is formed
by the behavior parameter bt−2 multiplying the expected fundamentals based on
It−1 .

2 Mankiw and Reis [18] assume that an exogenous proportion κ of agents will update the
information sets in each period. Then, at each t there are κ percent of agents with E(st+1 |It ),
κ(1 − κ) percent of agents with E(st+1 |It−1 ), κ(1 − κ)2 percent of agents with E(st+1 |It−2 ), and
so on. Thus the mean forecast is κ ∞j =0 (1 − κ) E(st+1 |It−j ).
j
3 Full
information in this setting is not complete information since the behavioral parameters are
unknown to agents.
Does Persistent Learning or Limited Information Matter in Forward Premium Puzzle? 167

The market forecast Et (st+1 ) is a weighted average of forecasts for these two
kinds of agents. For simplicity, we assume that the proportion of agents using full
information is an exogenous n in every period.

Et st+1 = ns1,t+1
e
+ (1 − n)s2,t+1
e
(11)

The exchange rate is obtained by substituting the market forecast in Eq. (5).

st = θ (nbt−1 vt + (1 − n)bt−2 ρt−2 vt−1 ) + vt (12)

Although agents know the structural form relationship between the fundamentals
and the exchange rate, the parameter values are imperceptible to them. The
parameter values b should be surmised by regressing st on vt−1 from the observed
data. Likewise, the parameter values ρ should be surmised by regressing vt on vt−1
from the observed data. We introduce constant gain learning into the model. The
learning process obeys
−1
bt = bt−1 + γ (st − bt−1 vt−1 )vt−1 Rt−1 = bt−1 + γ Q(t, bt−1 , vt−1 )
−1
ρt = ρt−1 + γ (vt − ρt−1 vt−1 )vt−1 Rt−1

Rt = Rt−1 + γ (vt2 − Rt−1 ) (13)

where γ > 0 is the learning gain parameter. Rt could be viewed as an estimate

of the second moment of the fundamentals. The gain parameter reflects the trade-
off between tracking and filtering, and it is set to be 1/t under decreasing gain
learning. The decreasing gain learning b will converge to b̄ with probability one.
Under constant gain learning, agents weight recent data more heavily than past data
and the gain parameter is therefore set to be a constant. The parameter bt is allowed
to vary over time, which in turn allows for potential structural change taking an
unknown form. As suggested by Orphanides and Williams [22], the model with the
gain parameter in a range of γ = 0.02 ∼ 0.05 will provide a better forecasting
performance for GDP growth and inflation. Based on the above model setting, we
discuss the impact of persistent learning and information flow on market efficiency.

2.3 Forward Exchange Market Efficiency

In this subsection we would like to show how learning process and limited
information model affect the relationship between forward exchange rates and future
spot rates. The tests of market efficiency are usually based on Eq. (1), corresponding
to Fama’s definition. Fama [11] summarizes the idea saying that a market is called
efficiency, when prices fully reflect all available information. Therefore under
market efficiency, the only reason for the price changes is the arrival of news or
168 Y.-C. Lin

unanticipated events. It also implies that forecast errors are unpredictable on the
basis of any information that is available at the time forecast is made. According to
the above definition, market efficiency could be represented as below if agents are
rational expectation and risk neutrality.

st+1 = Et st+1 + ut+1 (14)

ut+1 is the news in exchange rate markets, randomly distributed and expected to
be zero at time t. In other words, it is impossible to make superior profits by trading
on the basis of information that is available at time t if market is efficiency. From
Eqs. (5) to (11), the time t+1 adaptive learning forecast error is

st+1 − Et st+1 = θ Et+1 st+2 + vt+1 − [ns1,t+1

e
+ (1 − n)s2,t+1
e
]
= (1 + θ nbt )t+1 − (1 − n)bt−2 ρt−2 vt−1
+{ρ − (1 − ρθ )bt + (1 − n)[(1 − ρθ )bt
+θ bt−1 ρt ] + nγ Q(t, bt−1 , vt−1 )}vt (15)

According to Fama’s definition, efficient market hypothesis is supported if the

expected forecast error based on information available at time t is equal to zero. In
the case that all agents have full information (n = 1), Eq. (15) is rewritten as below.

st+1 − Et st+1 = [ρ − (1 − ρθ )bt + γ Q(t, bt−1 , vt−1 )]vt + (θ bt + 1)εt+1 (16)

If agents follow rational expectations, bt is equal to b̄ = ρ/(1 − ρθ ) in every

period, and the gain parameter γ → 0. The coefficient of vt is asymptotically
equal to zero. With rational expectation, agents’ forecast error cannot be predicted
by fundamental before current period, vt−1 , vt−2 , . . .. However, agents in real lives
usually make forecast by adaptive learning (γ > 0), which will induce systematic
forecast errors. The failure of market efficiency could be ascribed to adaptive
learning under full information. In the case of limited information flows (n < 1),
the forecast error has systematic components because the coefficients on vt and vt−1
are not zero. There are two channels that could affect the result of market efficiency,
the proportion of agents with limited information, 1 − n, and adaptive learning, γ .
The effect of adaptive learning on the expected forecast error depends on parameters
ρ, θ, n. Therefore, the effect of adaptive learning on the expected forecast error is
not monotonic under limited information flows, which is contradict with the findings
in [8] (as shown in Appendix 1). We also find that the expected forecast error is not
equal to zero even when agents are rational and information flow is limited.
Does Persistent Learning or Limited Information Matter in Forward Premium Puzzle? 169

2.4 Numerical Calibration

In this section, we simulate the exchange rate under limited information flow model
with adaptive learning, given b̄ as the initial parameter. To find out the impact of
limited information flows on the results of forward premium puzzle, the proportion
of agents using full information, n, is assumed to be exogenous and constant. The
exchange rates are simulated based on various n tested by Eq. (1), Fama regression
without intercept.
In the beginning of the simulation, we adopt the parameter setting in [8] and
compare the results with them. The results are shown in Table 1. In the case of
full information, n = 1, the simulated β̂ are positive in small gain parameters. The
simulated β̂ are decreasing to negative as the gain parameter increases. This result
coincides with the findings in [8]. When the spot and forward exchange rates derived
from near unit root process leads the simulated slope coefficient of the spot return on
the forward premium to be negative. The statistical explanation is that there exists
long memory behavior of the conditional variance in forward premium. Although β̂
is centered around unity, but it is widely dispersed and converge very slowly to its
true value of unity. A relatively large number of observations is required to provide
reliable estimates [2, 14].

Table 1 Simulated β̂ and tβ̂ under persistent learning and limited information by using the
parameter settings in [8]
γ n = 0.1 n = 0.4 n = 0.7 n=1
(a) θ = 0.9, ρ = 0.99
RE −5.45 (−23.17) −3.57 (−10.24) −0.95 (−1.76) 1.29 (1.56)
0.02 −6.22 (−27.70) −4.59 (−12.50) −1.69 (−2.73) 1.03 (1.15)
0.03 −6.65 (−30.72) −4.84 (−13.13) −2.12 (−3.23) 0.77 (0.81)
0.04 −6.75 (−31.95) −5.09 (−13.82) −2.35 (−3.68) 0.39 (0.45)
0.05 −6.71 (−31.67) −5.27 (−14.37) −2.79 (−4.03) 0.15 (0.18)
0.10 −6.21 (−26.87) −4.64 (−13.62) −2.44 (−4.56) −0.81 (−1.02)
(b) θ = 0.6, ρ = 0.99
γ n = 0.1 n = 0.4 n = 0.7 n=1
RE −1.35 (−19.26) −1.23 (−8.96) −0.81 (−2.67) 1.20 (1.48)
0.02 −1.40 (−20.61) −1.31 (−9.75) −1.08 (−3.55) 0.23 (0.24)
0.03 −1.39 (−20.67) −1.30 (−9.86) −1.08 (−3.70) −0.14 (−0.18)
0.04 −1.38 (−20.59) −1.31 (−9.96) −1.05 (−3.74) −0.34 (−0.52)
0.05 −1.36 (−20.21) −1.27 (−9.94) −1.05 (−3.92) −0.32 (−0.56)
0.10 −1.28 (−19.24) −1.15 (−9.67) −0.85 (−3.91) −0.39 (−1.01)
Note: Results from 1000 simulations with sample size equal to 360 after discarding the first 20,000
data points. Table gives medians of β̂ and tβ̂ for testing H0 : β = 0, without intercept in the Fama
regression. The medians of tβ̂ are shown in parentheses
170 Y.-C. Lin

Table 2 Simulated Fama coefficient in multi-period version under persistent learning and limited
information
γ n = 0.1 n = 0.4 n = 0.7 n=1
(a) θ = 0.9, ρ = 0.99
RE 1.14 (19.02) 1.53 (28.92) 1.36 (26.10) 1.00 (18.95)
0.02 1.14 (19.12) 1.54 (28.88) 1.38 (26.42) 1.02 (19.25)
0.03 1.14 (19.08) 1.54 (29.01) 1.39 (26.68) 1.03 (19.44)
0.04 1.14 (19.09) 1.54 (29.08) 1.39 (26.68) 1.03 (19.50)
0.05 1.14 (19.06) 1.54 (28.95) 1.40 (26.91) 1.05 (19.74)
0.10 1.12 (18.81) 1.52 (28.47) 1.42 (27.48) 1.08 (20.35)
(b) θ = 0.6, ρ = 0.99
RE 1.30 (13.44) 1.81 (24.39) 1.39 (23.21) 1.00 (18.98)
0.02 1.29 (13.26) 1.83 (24.22) 1.41 (23.15) 1.02 (19.08)
0.03 1.28 (13.17) 1.82 (23.93) 1.41 (23.06) 1.02 (19.10)
0.04 1.27 (13.18) 1.80 (23.61) 1.41 (22.97) 1.03 (19.22)
0.05 1.25 (13.06) 1.77 (23.15) 1.41 (23.01) 1.03 (19.23)
0.10 1.16 (12.61) 1.66 (21.51) 1.40 (22.51) 1.05 (19.40)
st+1 − st−1 = β̂(ft − st−1 ) + ût+1
Note: Results from 1000 simulations with sample size equal to 360 after discarding the first 20,000
data points. Table gives medians of β̂ and tβ̂ for testing H0 : β = 0, without intercept in the Fama
regression. The medians of tβ̂ are shown in parentheses

In the case of limited information, n < 1, the simulated β̂ are negative even
when gain parameters are small. Therefore, it may not be possible to explain forward
premium puzzle by gain parameter when information flow is limited. In the limited
information case of θ = 0.6 and ρ = 0.99, the simulated β̂ is increasing in the gain
parameter γ , contrary to the case of full information. The impact of gain parameter
on the simulated β̂ is not monotonic when considering limited information.
McCallum [21] argues that the forward premium puzzle does not exist in another
form of the test that in a multi-period version. A multi-period version for Eq. (1) is
as below:

st+1 − st+1−j = β̂(ft − st+1−j ) + ût+1 , j = 2, 3, . . . . (17)

McCallum [21] finds that β̂ is close to one if j ≥ 2, which coincides with the
theory. As shown in Table 2, our simulated estimates support McCallum’s findings.
The simulated β̂ are close to unity and significantly greater than zero. Although the
forward premium puzzle resulted from persistent learning or limited information
in a single-period version, the puzzle will disappear when testing in a multi-period
version.
This section integrates the concept of limited information flow with the model of
[8]. We find that when information flow is restricted, the impact of gain parameter on
Fama coefficient is small relative to the impact of the proportion of full information
Does Persistent Learning or Limited Information Matter in Forward Premium Puzzle? 171

agents on Fama coefficient. The innocence of behavior parameter may not be the
critical reason under limited information for forward premium puzzle, while the
innocence of the current information about fundamentals might be the main point.

3 Dual Learning

The proportion of agents with full information n is assumed to be exogenous in prior

section. Here n is allowed to vary over time. This section introduces dual learning
into the model: parameter learning and dynamic predictor selection. The market
forecast is a weighted average of the two forecasts, and the proportion of agents is
time varying.

Et st+1 = nt−1 s1,t+1

e
+ (1 − nt−1 )s2,t+1
e
(18)

Each period, agents could choose to use the first or second forecasting rule. This
choice is based on the models’ forecasting performance in each period. Following
[5], agents update their mean square error (MSE) estimates according to a weighted
least squares procedure with geometrically decreasing weights on past observations.

MSEi,t = MSEi,t−1 + γ [(st − si,t

e
) − MSEi,t−1 ] (19)

Refer to [6], predictor proportions are determined according to the following

discrete choice formula:
exp(−αMSE1,t )
nt = (20)
exp(−αMSE1,t ) + exp(−αMSE2,t )

where α measures the sensitivity of agents reacting to the relative performance of

the two forecasting rules. The value of α is correlated to the switching barriers
on choosing forecasting rules, such as information costs or switching costs. When
the switching parameter α equals to zero, agents do not react to any changes
of the relative performance and their weight are sticky at 0.5. As the switching
parameter approaches to infinity, agents are sensitive and switch the forecasting rule
instantaneously. The proportion of full information agents then become either 0 or 1.
The dual learning algorithm is given by a loop over Eqs. (5), (9), (10), (13), (18),
e
(19), and (20). In period t, agents in first group make exchange rate forecast s1,t+1
by predetermined parameter estimates bt−1 and the current observed fundamentals
vt . Since agents in the second group are restricted to the latest information, the
e
exchange rate forecast s2,t+1 is made by predetermined parameter estimates bt−2 ,
ρt−2 and the fundamentals vt−1 , which is observed in the last period. The market
expectation is formed as a weighted average of the forecast for group 1 and 2, with
predetermined weights nt−1 and (1 − nt−1 ). When agent observe st , they begin
to update their belief parameters bt , ρt , and MSEi,t . The relative performance of
172 Y.-C. Lin

Table 3 Simulated β̂ and tβ̂ by dual learning

α = 0.01 α = 0.05 α = 0.1 α = 500
(a) γ = 0.02
n̄ 0.67 0.94 0.98 1
β̂ j =1 −1.94 (−3.59) 0.07 (0.10) 0.19 (0.28) 0.35 (0.47)
β̂ j =2 1.41 (26.75) 1.07 (19.99) 1.04 (19.46) 1.02 (19.18)
(b) γ = 0.03
n̄ 0.68 0.95 0.98 1
β̂ j =1 −1.95 (−3.53) 0.02 (0.03) 0.10 (0.15) 0.18 (0.22)
β̂ j =2 1.40 (26.46) 1.07 (19.99) 1.05 (19.52) 1.03 (19.32)
(c) γ = 0.04
n̄ 0.69 0.95 0.98 1
β̂ j =1 −1.87 (−3.58) −0.24 (−0.32) −0.14 (−0.17) −0.09 (−0.15)
β̂ j =2 1.39 (26.27) 1.08 (20.01) 1.05 (19.62) 1.04 (19.43)
(d) γ = 0.05
n̄ 0.69 0.95 0.98 1
β̂ j =1 −1.79 (−3.52) −0.32 (−0.46) −0.28 (−0.37) −0.23 (−0.29)
β̂ j =2 1.40 (26.22) 1.09 (20.14) 1.06 (19.7) 1.05 (19.58)
(e) γ = 0.1
n̄ 0.69 0.93 0.98 0.98
β̂ j =1 −1.60 (−3.7) −0.74 (−1.30) −0.75 (−1.17) −0.69 (−1.07)
β̂ j =2 1.40 (26.16) 1.13 (20.87) 1.10 (20.38) 1.09 (20.23)
st+1 − st+1−j = β̂(ft − st+1−j ) + ût+1 , j = 1, 2
Note: Results from 1000 simulations with sample size equal to 360. The simulation is based on
the parameter assumption of θ = 0.9 and ρ = 0.99. Table gives medians of β̂ and tβ̂ for testing
H0 : β = 0, with intercept in the Fama regression. The medians of tβ̂ are shown in parentheses

forecasting rule is used to determine the new predictor proportions nt for the next
period. All initial values of the parameters are set by the rational expectation values.
The initial weight on model 1 is drawn from a standard uniform distribution U (0, 1).
The median Fama coefficients estimated by simulated exchange rates are shown
in Table 3. The result in Table 3 is simulated by the parameter assumptions of
θ = 0.9 and ρ = 0.99. For given γ , the average proportion for the first forecasting
rule over the period, n̄, increases in α, the sensitivity of the relative performance
to the predictors. If agents react more strongly to the relative performance of the
forecasting rules, they tend to forecast by using full information. Figure 2 plots the
cross-section median of the 1000 simulations for the proportion of the first forecast
rule nt . When α equals to 0.01, nt increases in relatively slow speed. While it
switches to unity immediately when α is larger than 500. The gain parameter γ also
influences the time path of nt . Agents rely more on recent data to update their belief
parameter when γ is larger. The time path of nt will converge to unity in shorter
time in larger gain parameter. Highly dependent on the recent data to forecast may
also lead to overreaction to the market news, the time path of nt is more volatile in
Does Persistent Learning or Limited Information Matter in Forward Premium Puzzle? 173

Fig. 2 Weights on full information under dual learning for θ = 0.9 and ρ = 0.99. (a) γ = 0.02.
(b) γ = 0.03. (c) γ = 0.04. (d) γ = 0.05. (e) γ = 0.1

larger gain parameters. The situation might happen when agents are overconfident
on the recent data, they tend to overreact on their arbitrage behavior. Then having
information might lead to market inefficiency. Agents might not benefit from too
much information and would rather drop some of information when forecasting.
In the case of single-period version of Fama regression (j = 1), when α is
equal to 0.01, the average proportion of full information agents is about 0.7, while it
increases to unity when α is equal to 500. It also shows that when α is equal to 0.01,
the simulated β̂ are significantly negative despite the value of gain parameter. Both
n̄ and β̂ increase with α. The puzzle tends to disappear as the switching parameter
is increasing. β̂ are positive in small gain parameter, and turn to be negative when
the gain parameter is larger than 0.03. However, β̂ are still not significantly negative
as the gain parameter increases to 0.1. Here we find that forward premium puzzle is
better explained by the switching parameter, instead of gain parameter.
The Fama regression in a multi-period version is also examined under dual
learning. The results are reported in the case of j = 1 in Table 3. The simulated
β̂ are close to unity and significantly greater than zero. The value of simulated β̂ is
uncorrelated to the average proportion for the first forecast rule over the period, n̄,
and also the sensitivity to the relative performance of the predictors, α. The puzzle
will disappear when testing in a multi-period version.
174 Y.-C. Lin

4 Conclusion

The forward premium puzzle was explained in several directions, such as irrational
agents, incomplete knowledge, or risk neutral agents. Chakraborty and Evans [8]
assume that agents are misperceive of behavior parameters and have to update
their knowledge of behavior parameters by persistent learning. They show that
the puzzle coexists with larger gain parameters. However, the Fama coefficient is
always positive in their theoretical model and the simulated Fama coefficient is also
highly correlated with sample size. This paper continues the discussion of [8]. The
persistent learning model of [8] was enlarged to a limited information case, that the
current fundamentals are not perceivable by some agents.
Based on the assumptions above, we find that: First, the simulated β̂ is negative
when the proportion of full information agents is small, even in small learning
parameter and rational expectation. It implies that the effect of limited information
on asymptotic bias dominates that of persistent learning. Second, the limited
information could not only explain forward premium puzzle, it is consistent with
the stylized facts, which is found in [21]. The puzzle will not remain in the multi-
period test of Fama equation. Third, when agents react strongly to the relative
performance of the forecasting rule by using full or limited information, all agents
switch to apply full information quickly. The proportion of full information agents
approaches to unity. Therefore, market efficiency hold when the sensitivity of the
relative performance of the forecasting rule is high. This result is similar for different
gain parameters. Finally, the same as the case that the proportion of full information
agents is exogenous, we find that the Fama coefficients are positive when examining
a multi-period Fama regression. These concepts will help for further understanding
of the reasons for forward premium puzzle.

Appendix 1

Following [8], we could derive the distribution of bt for small γ > 0, which is
approximately normal with mean b̄ and variance γ C, where C = (1 − ρ 2 )(1 − (1 −
n)ρθ )2 /2(1−ρθ )3 . Using this property, we could derive the least-square estimates β̂
under the null hypothesis of H0 : β = 1 in Eq. (1). We have the following findings:
First, under H0 : β = 1 and sufficiently small γ > 0, the asymptotic bias plim
B(γ , θ, ρ, n) = plimβ̂ − 1 is approximately equal to

B(γ , θ, ρ, n)
γ (1 − ρθ )(1 + ρ)(1 − θ )((1 − n)ρ 2 + n) − (1 − n)A
=−
(γ (1 − θ )2 (1 + ρ) + 2(1 − ρ)(1 − ρθ )((1 − n2 )ρ 2 + n2 ) + (1 − n)F
(21)
Does Persistent Learning or Limited Information Matter in Forward Premium Puzzle? 175

where A ≡ (1 + ρ)(γ n(1 − θ )(1 − ρ 2 ) + 2ρ(1 − ρθ )(nρ(1 − θ ) − (1 − ρθ ))),

and F ≡ 2(1 − ρθ )2 (1 + ρ)((1 − n)(1 − ρθ ) + 2n(1 − ρ)). When all agents in the
society update their information sets freely, i.e., n = 1, B(γ , θ, ρ, 1) will be equal
to that provided by Chakraborty and Evans [8].
Plimβ̂ is below unity when the parameters satisfy 0 ≤ θ < 1, 0 ≤ ρ < 1 and
0 ≤ n < 1. As shown in Fig. 3, when the fundamentals are highly persistent, β̂ turns

Fig. 3 Theoretical plim β̂ for (a) θ = 0.6, n = 1 and (b) θ = 0.6, n = 0.6
176 Y.-C. Lin

Fig. 4 Theoretical plim β̂ for (a) ρ = 0.98, n = 1 and (b) ρ = 0.98, n = 0.6
Does Persistent Learning or Limited Information Matter in Forward Premium Puzzle? 177

to negative under limited information (n = 0.6) and is always in the range between
0 and 1 under full information.
Second, as ρ −→ 1, the asymptotic bias B(γ , θ, ρ, n) = plimβ̂ −1 approximate
to below −1, and the OLS estimates β̂ turns to be negative, i.e., forward premium
puzzle exists.

γ + 2(1 − n)2
lim B(γ , θ, ρ, n) = − ≤1 (22)
γ + 2(1 − n)2 (1 − θ )

In the case of full information (n = 1), the asymptotic bias is equal to −1, which
means the lower bound of theoretical β̂ is 0.
Third, contrary to [8], our study finds that the learning process might either
magnify or minify the asymptotic bias when information flows are limited. The
impact of gain parameter on plimβ̂ is not monotonic, depending on the persistence
of fundamental ρ and the dependence of that current exchange rate on the
expectation of future exchange rate θ . β̂ may either decrease or increase in learning
parameter under limit information, as shown in the lower panel of Figs. 3 and 4.

Appendix 2

Under constant gain learning, there is a non-zero correction for the estimated
coefficients through the forecasting error, even in the limit as t approaches to infinity.
The constant gain algorithm converges to a distribution rather than the constant
value of rational expectation. Evans [10] show that under the two conditions below,
a constant gain algorithm converges to a normal distribution centered the rational
expectation equilibrium. First, the gain parameter should be in the range between 0
and 1. Second, the parameter values should coincide with the E-stability conditions.
The learning process obeys
−1
bt = bt−1 + γ (st − bt−1 vt−1 )vt−1 Rt−1 = bt−1 + γ Q(t, bt−1 , vt−1 )
−1
ρt = ρt−1 + γ (vt − ρt−1 vt−1 )vt−1 Rt−1

Rt = Rt−1 + γ (vt2 − Rt−1 ) (23)

The learning system is E-stable if it is a locally asymptotically stable equilibrium

of the ordinary differential equation.

db
= (θρ − 1)b
dτ
dρ
=0 (24)
dτ
178 Y.-C. Lin

Then an rational expectation equilibrium is E-stable if the Jacobian of db/dτ

evaluated at the rational expectation equilibrium is a stable matrix, implying that it
has eigenvalues with strictly negative real parts.

db/dτ
J (b̄) = | = θρ − 1 (25)
db b=b̄
Since the parameters θ and ρ are in the range of 0 and 1, the unique stationary
rational expectation equilibrium is E-stable.

References

1. Bacchetta, P., & Wincoop, E. V. (2009). On the Unstable Relationship between Exchange
Rates and Macroeconomic Fundamentals. NBER Working Papers 15008, National Bureau of
Economic Research, Inc.
2. Baillie, R. T., & Bollerslev, T. (2000). The forward premium anomaly is not as bad as you
think. Journal of International Money and Finance, 19, 471–488.
3. Black, F. (1986). Noise. Journal of Finance, 41, 529–543.
4. Branch, W. A. (2007). Sticky information and model uncertainty in survey data on inflation
expectations. Journal of Economic Dynamics and Control, 31, 245–276.
5. Branch, W. A., & Evans, G. W. (2007). Model uncertainty and endogenous volatility. Review
of Economic Dynamics, 10(2), 207–237.
6. Brock, W. A., & Hommes, C. H. (1997). A rational route to randomness. Econometrica, 65(5),
1059–1096.
7. Burnside, C., Han, B., Hirshleifer, D., & Wang, T. Y. (2011). Investor overconfidence and the
forward premium puzzle. Review of Economic Studies, 78, 523–558.
8. Chakraborty, A., & Evans, G. W. (2008). Can perpetual learning explain the forward-premium
puzzle? Journal of Monetary Economics, 55(3), 477–490.
9. Engel, C. (1992). On the foreign exchange risk premium in a general equilibrium model.
Journal of International Economics, 32(3–4), 305–319.
10. Evans, G. W., & Honkapohja, S. (2001). Learning and expectations in macroeconomics.
Princeton, NJ: Princeton University Press.
11. Fama, E. F. (1970). Efficient capital markets: A review of theory and empirical work. Journal
of Finance, 25, 383–417.
12. Gourinchas, P. O., & Tornell, A. (2004). Exchange rate puzzles and distorted beliefs. Journal
of International Economics, 64(2), 303–333.
13. Hodrick, R. J. (1987). The empirical evidence on the efficiency of forward and futures foreign
exchange markets. Chur: Harwood Academic Publishers.
14. Kellard, N., & Sarantis, N. (2008). Can exchange rate volatility explain persistence in the
forward premium? Journal of Empirical Finance, 15(4), 714–728.
15. Krasker, W. S. (1980). The ‘peso problem’ in testing the efficiency of forward exchange
markets. Journal of Monetary Economics, 6(2), 269–276.
16. Lewis, K. (1989). Changing beliefs and systematic rational forecast errors with evidence from
foreign exchange. American Economic Review, 79, 621–36.
17. Lo, A. (2004). The adaptive markets hypothesis: Market efficiency from an evolutionary
perspective. Journal of Portfolio Management, 30, 15–29.
18. Mankiw, N. G., & Reis, R. (2002). Sticky information versus sticky prices: A proposal to
replace the New Keynesian Phillips curve. The Quarterly Journal of Economics, 117(4), 1295–
1328.
Does Persistent Learning or Limited Information Matter in Forward Premium Puzzle? 179

19. Mark, N. C. (2009). Changing monetary policy rules, learning, and real exchange rate
dynamics. Journal of Money, Credit and Banking, 41(6), 1047–1070.
20. Mark, N. C., & Wu, Y. (1998). Rethinking deviations from uncovered interest parity: The role
of covariance risk and noise. Economic Journal, 108(451), 1686–1706.
21. McCallum, B. T. (1994). A reconsideration of the uncovered interest parity relationship.
Journal of Monetary Economics, 33(1), 105–132.
22. Orphanides, A., & Williams, J. C. (2005). The decline of activist stabilization policy: Natural
rate misperceptions, learning, and expectations. Journal of Economic Dynamics and Control,
29(11), 1927–1950.
23. Scholl, A., & Uhlig, H. (2008). New evidence on the puzzles: Results from agnostic
identification on monetary policy and exchange rates. Journal of International Economics,
76(1), 1–13.
Price Volatility on Investor’s Social
Network

Yangrui Zhang and Honggang Li

Abstract Based on an Ising model proposed by Harras and Sornette, we established

an artificial stock market model to describe the interactions among diverse agents.
We regard these participants as network nodes and link them with their correlation.
Then, we analyze the financial market based on the social network of market
participants. We take the random network, scale-free network, and small-world
network into consideration, and then build the stock market evolution model
according to the characteristics of the investors’ trading behavior under the different
network systems. This allows us to macroscopically study the effects of herd
behavior on the rate of stock return and price volatility under different network
structures. The numerical simulation results show that herd behavior will lead to
excessive market volatility. Specifically, the greater the degree of investor’s trust
in neighbors and their exchange, the greater the volatility of stock price will be.
With different network synchronization capabilities, price fluctuations based on the
small-world network are larger than those based on the regular network. Similarly,
price fluctuations based on the random network are larger than those based on
the small-world network. On the other hand, price fluctuations based on both the
random network and the small-world network firstly increase and then decrease with
the increase of the average node degree. All of these results illustrate the network
topology that has an important impact on the stock market’s price behavior.

Keywords Artificial stock market · Social network · Price volatility · Herd

behavior · Network structures

Y. Zhang · H. Li ()
School of Systems Science, Beijing Normal University, Beijing, People’s Republic of China
e-mail: hli@bnu.edu.cn

© Springer Nature Switzerland AG 2018 181

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_9
182 Y. Zhang and H. Li

1 Introduction

For decades, a multitude of complex and macroscopic behavioral characteristics

have constantly sprung up in financial market. But it is difficult to accurately
estimate the nonlinear relationships between variables based on the traditional
probability theory and econometrics. Establishing artificial financial markets, using
objective bounded rationality assumptions to simulate the interaction between
investor behaviors and studying the evolution and dynamics process, enable us
to have a deeper understanding of the mechanism driving macroscopic behavior
characteristics.
Establishing the artificial stock market model to study the interactions of investor
behaviors and macroscopic behavior characteristics has become an important field
of computational finance. Economists have proposed different mechanisms to
describe the diverse agents, which are used to simulate macroscopic behavior
and dynamic evolution in the market, and the interactions among them. Some
scholars such as Arifovic [1], Lettau [2], and Kirman [3] analyzed the influence
of information on investors’ decisions through the observation of real participants’
trading behavior. In the model of Johnson [4], when the value of the best strategy
was below a certain threshold value, traders remained inactive, so the number of
traders in the system is time-varying. Based on the Ising model, Iori [5] established
the feedback of price and threshold value, and concluded that price and volatility
are related. Johansen and Sornette [6] pointed out that all traders can be regarded as
interactive sources of opinion, they found that the feedback effects and the memory
function play an important role in the market. Basically, these researchers proposed
a variety of artificial financial market models with heterogeneous interacting agents
in order to account for the observed scaling laws, most of these models simply
assume either a fully connected network or a regular structure. More reasonable
network structures are needed to be taken into account with the improvement of
the research. In the framework of the imitative conformity mechanism, Kim [7]
researched the correlation of stock price with the scale-free network structure.
Liang [8] established an artificial stock market model on the basis of the small-
world network and explored some factors that lead to the complexity of financial
markets. Alfarano and Milakovic [9] analyzed the correlation of herd behavior and
the systemic size based on the network topology.
In the financial market, the trading behavior of investors, who have bounded
rationality, leads to the complex nonlinear mechanism of the whole financial system.
Investors are more prone to being influenced by their peers, the media, and other
channels that combine to build a self-reflexive climate of optimism. Particularly,
social network communication from investors may greatly affect investment opin-
ion. At the same time, this communication may lead to significant imitation, herding,
and collective behaviors [10]. Therefore, it is necessary to establish a reasonable
social network to study interactions between investors and herd behavior from
the microscopic aspect. Based on the complex network concept, we treat these
Price Volatility on Investor’s Social Network 183

participants as network nodes and link them according to their correlation. Then,
we analyze the financial market with the social network.
The preceding research has adopted the agent-based model to study the complex-
ity of the financial market. However, most of agent-based models do not consider
the social network between traders as a key element. Our model focuses on the
interaction among traders by a comparative analysis of different complex network
structures. We take the random network, the scale-free network [11], and the small-
world network [12] into consideration, and build the dynamic model according to
the characteristics of the traders’ investing behavior. Technically, inspired by Harras
and Sornette model [6, 13], we establish the artificial stock market model to study
the effect of herd behavior on the rate of return and price volatility under different
network structures, from a kind of macroscopic aspect.

2 The Model

2.1 Investor’s Social Network

The relationship between investors in the stock market can be described through
the network structure, in which the connection may lead to the spread of trading
behavior.
It is necessary to establish reasonable interpersonal network topology to research
the interactions between investors and herd behavior. Due to the real stock market
network scale and the complex interactions between nodes, its topology structure
is still unexplored. Therefore, we mainly take the random network, the scale-free
network, and the small-world network into consideration, and build the stock market
evolution model so that we can macroscopically study the effects of herd behavior
on the rate of stock return and price volatility under different network structures.
In addition, in our opinion, the social network structure itself evolves very slowly.
Therefore, in this paper we assume that the network topology remains unchanged
during the process of simulation.

2.2 Market

We focus on the interactive influence among traders and set an artificial stock market
model to describe the interaction based on a prototypical interaction-based herding
model proposed by Harras and Sornette [6].
We consider a fixed environment is composed of N agents, who are trading a
single asset, as a stock market, and at each time phase, agents have the possibility to
either trade (buy or sell) or to remain passive. The trading decision si (t) of agent i is
opinion-based on the future price development and the opinion of agent i at time t,
184 Y. Zhang and H. Li

wi (t). It consists of three different sources: idiosyncratic opinion, global news, and
acquaintance network.


J
wi (t) = c1i kij (t − 1)Ei [sj (t)] + c2i u(t − 1)n(t) + c3i εi (t)
j =1

Where εi (t) represents the private information of agent i, n(t) is the public
information. J is the number of neighbors that agent i polls for their opinion and
Ei [sj (t)] is the action of the neighbor j at time t − 1; (c1i , c2i , c3i ) is the form of
the weights the agent attributes to each of the three pieces of information.
It is assumed that each agent is characterized by a fixed threshold wi to control
the triggering si (t) of an investment action. An agent i uses a fixed fraction g of
his cash to buy a stock if his conviction wi (t) is sufficiently positive so as to reach
the threshold: wi (t) ≥ wi . On the contrary, he sells the same fixed fraction g of the
value of his stocks if wi (t) ≤ −wi .

wi (t) ≥ wi : si (t) = +1(buying)

cashi (t)
vi (t) = g ·
p(t − 1)
wi (t) ≤ −wi : si (t) = −1(selling)
vi (t) = g · stocksi (t)

Once all of the agents have determined their orders, the new price of the asset is
determined by the following equations:

1 
N
r(t) = si (t) · vi (t)
λ·N
i=1

log[p(t)] = log[p(t − 1)] + r(t)

Here r(t) is the return and vi (t) is the volume at time t; λ represents the relative
impact of the excess demand upon the price, i.e., the market depth.
When the return and the new price are determined, the cash and number of stocks
held by each agent i are updated according to

cashi (t) = cashi (t − 1) − si (t)vi (t)p(t)

stocksi (t) = stocksi (t − 1) + si (t)vi (t)
Price Volatility on Investor’s Social Network 185

2.3 Adaption

At each step traders change the trust factor according to the related traders’ decisions
and public information from the previous period. We assume that agents adapt their
beliefs concerning the credibility of the news n(t) and their trust in the advice
Ei [sj (t)] of their social contacts, according to time-dependent weights u(t) and
kij (t), which take into account their recent performance. The implementation is
achieved by a standard auto-regressive update:

r(t)
u(t) = αu(t − 1) + (1 − α)n(t − 1)
σr
r(t)
kij (t) = αkij (t − 1) + (1 − α)Ei [sj (t − 1)]
σr

where α represents traders’ memory in the process, σr is the volatility of returns.

3 Simulations and Results

3.1 Parameter Setting

In our simulations, we fix the network parameters to N = 2500, the market

parameters to λ = 0.25, g = 2%, wi = 2, the initial amount of cash and stocks
held by each agent to cashi (0) = 1 and stocksi (0) = 1, and the memory discount
factor to α = 0.90.

3.2 Market Price

Firstly, we assume that all traders lack communication with others. The numerical
experimental results in Fig. 1 show that, when traders make independent decisions,
fluctuations in market price sequence (red) are equal to fluctuations in fundamental
value (blue), which is determined by the market information. We consider that this
market is efficient in this case.
We then allow the communication between traders with the random network
structure. In Fig. 2, we can observe the excessive volatility and price bubbles; the
volatility of return has significant cluster characteristics.
Figures 3 and 4 show that the distribution of return has the characteristics of a
higher peak and a fat tail compared with the normal distribution The right panel
also shows the slow attenuation of the autocorrelation function; the absolute value
of the returns is related for a long time, which also suggests that the price fluctuation
has the volatility clustering property. All of these results concur with the empirical
research.
186 Y. Zhang and H. Li

0.8
market price
0.7
fundamental price
0.6 difference value

0.5

0.4

0.3
p

0.2

0.1

-0.1

-0.2
0 500 1000 1500 2000 2500 3000
t

Fig. 1 Series of market price and fundamental value

0.4

0.2
p

-0.2

-0.4
0 200 400 600 800 1000 1200 1400 1600 1800 2000
t
0.1

0.05
R

-0.05
0 200 400 600 800 1000 1200 1400 1600 1800 2000
t

Fig. 2 Market price and returns

Price Volatility on Investor’s Social Network 187

50

45

40

35

30
PDF(R)

25

20

15

10

0
-0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08
R
Mean Maximum Minimum Std.Dev Skewness Kurtosis Jarque-Bera Probability
Data 1.15e-004 0.146000 -0.11000 0.027559 0.189764 6.497000 548.6494 0

Fig. 3 Distribution of return

Sample Autocorrelation Function

1

0.8
Sample Autocorrelation

0.6

0.4

0.2

-0.2
0 5 10 15 20 25 30 35 40 45 50
Lag

Fig. 4 Correlation of volatility (measured as the absolute value of returns)

188 Y. Zhang and H. Li

0.13 0.13
Grid network
ER network BA network
0.12 0.12
WS network WS network

0.11 0.11
Volatility

Volatility
0.1 0.1

0.09 0.09

0.08 0.08

0.07 0.07
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
C1 C1

Fig. 5 Volatility changes with c1 under different network structures

3.3 Market Volatility

To research the relationship between investors’ trust in neighbors and market

volatility under different network structures with c1 values of 1 to 3, we set the
average node of three network types to 8. Taking averaged value of the return
volatility from 200 times simulations, as shown in Fig. 5, we can conclude that in
the same network structure, the greater degree of investors trust in neighbors and
their exchange, the greater the volatility of stock price will be. With the different
network synchronization capabilities, the difference between volatility under the
small-world (SW) network and the scale-free (BA) network is small. Furthermore,
the price fluctuations based on the small-world network are larger than those based
on the regular grid network. Finally, equally, the price fluctuations based on the
random (ER) network are larger than those based on the small-world network.
For a larger system size, it can be seen as a collection of small networks, the
excessive volatility still exists since the influence of the macro- information. So the
model is robust with respect to an enlargement of system size (Fig. 6).

3.4 Network Characteristics

We refer to the construction algorithm of the WS small-world network to change the

values of rewiring probability p to transfer the network structure from the regular
network (p = 0) to the random network (p = 1). Figure 7 shows the impact of p on
the market volatility. During the transition from the regular network to the random
network, the volatility increases with the increase of the rewiring probability, under
the same transmission sensitivity. We also find that the market volatility has a fast
declining trend at first, and then leveled off when the clustering coefficient increases
(Fig. 8).
Price Volatility on Investor’s Social Network 189

0.12
C1=1.6
C1=2.0
0.1 C1=2.4

0.08
Volatility

0.06

0.04

0.02

0
2000 3000 4000 5000 6000 7000 8000 9000 10000
N

Fig. 6 Volatility changes with N (the total number of agents) for C1 = 1.6, 2.0, 2.4 (ER network)

0.13 0.096
0.125 C1=1.6 C1=1.6
C1=2.0 0.094 C1=2.0
0.12
C1=2.4 0.092 C1=2.4
0.115
0.11 0.09
Volatility

Volatility

0.105 0.088
0.1
0.086
0.095
0.084
0.09
0.085 0.082

0.08 0.08
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
p p

Fig. 7 Volatility changes with p (rewiring probability) for C1 = 1.6, 2.0, 2.4

Figure 9 displays the volatility changes with the average node degree under the
different network structures. Where the volatility firstly increases and then slowly
rests at a lower level with the increase of the average node degree. The cause of
this phenomenon may be attributed to the long-range connection and the higher
synchronizability of the small-world network. On the other hand, when the network
node degree increases to a certain degree, the connection of the activity nodes will
be wider. Finally, the behavior of traders in the market will tend to make an overall
decision, resulting in a loss of market volatility.
190 Y. Zhang and H. Li

0.096

C1=1.6
0.094 C1=2.0
C1=2.4

0.092

0.09
Volatility

0.088

0.086

0.084

0.082
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 8 Volatility changes with C (Clustering coefficient) for C1 = 1.6, 2.0, 2.4

0.03

Grid network
ER network
0.025 WS network

0.02
Volatility

0.015

0.01

0.005
0 2 4 6 8 10 12 14 16 18 20
Average node degree

Fig. 9 Volatility changes with average node degree changes under different network structures
Price Volatility on Investor’s Social Network 191

0.13
0.12
0.125
0.14 0.13 0.115
0.12
0.13 0.12
0.115 0.11
0.12

Volatility
Volatility

0.11 0.11
0.11 0.105
0.105 0.1
0.1 0.1
0.09 0.1 0.09
0.095
0.08 0.095 0.08
1 1
3 0.09 3 0.09
0.5 2.5 0.5 2.5
2 0.085 2 0.085
1.5 1.5
lamda 0 1 C1 lamda 0 1 C1

0.13
0.125
0.14
0.12
0.12 0.115
Volatility

0.11
0.1
0.105
0.08 0.1
0.095
0.06
1 0.09
3 0.085
0.5 2.5
2 0.08
1.5
lamda 0 1 C1

Fig. 10 Volatility with C1 and λ changes under different network structures (upper left panel: ER
network. upper right panel: SW network; lower panel: BA network)

3.5 Guru

In the communication network, we add a given special agent named “guru.” It

attracts the largest number of in-coming links. In our settings, each agent i chooses
to link a probability λ to the fixed agent guru, and with probability 1 − λ, to keep
the original connection.
Figure 10 shows the volatility changes with C1 and λ under different network
structures. It can be seen that the volatility increases with the increase of λ under the
same transmission sensitivity. This means that the connection of the guru changes
the network structure and increases the market volatility.

4 Conclusion

Based on the artificial stock market model, we research the effect of the network
topology on market volatility. When investors establish contact with others, the herd
behavior can lead to excessive volatility of the market. The greater the degree of
investor trust in neighbors and their exchange, the greater the volatility of stock
192 Y. Zhang and H. Li

price will be. On the other hand, the volatility firstly increases and then slowly rests
at a lower level with the increase of the average node degree, under three kinds of
network structure. Furthermore, the results show that the topology of the traders’
network has an important effect on the herding behavior and the market price.
It goes without saying that, in this paper, we only analyze one type of market
model. However, the market volatility based on different types of market models
and different mechanisms of information transmission still requires further research.

Acknowledgements We wish to thank Xuezhong He for useful comments and discussions, and
anonymous referees for helpful comments. None of the above is responsible for any of the errors
in this paper. This work was supported by the Fundamental Research Funds for the Central
Universities under Grant No. 2012LZD01 and the National Natural Science Foundation of China
under Grant No. 71671017.

References

1. Arifovic, J. (1996). The behavior of the exchange rate in the genetic algorithm and experimental
economies. Journal of Political Economy, 104, 510–541.
2. Lettau, M. (1997). Explaining the facts with adaptive agents: The case of mutual fund flows.
Journal of Economic Dynamics and Control, 21(7), 1117–1147.
3. Kirman, A. (1991). Epidemics of opinion and speculative bubbles in financial markets. In
Money and financial markets (pp. 354–368). Cambridge: Blackwell.
4. Johnson, N. F., Hart, M., Hui, P. M., & Zheng, D. (2000). Trader dynamics in a model market.
International Journal of Theoretical and Applied Finance, 3(03), 443–450.
5. Iori, G. (2002). A microsimulation of traders activity in the stock market: The role of
heterogeneity, agents’ interactions and trade frictions. Journal of Economic Behavior &
Organization, 49(2), 269–285.
6. Harras, G., & Sornette, D. (2011). How to grow a bubble: A model of myopic adapting agents.
Journal of Economic Behavior & Organization, 80(1), 137–152.
7. Kim, H. J., Kim, I. M., Lee, Y., & Kahng, B. (2002). Scale-free network in stock markets.
Journal-Korean Physical Society, 40, 1105–1108.
8. Liang, Z. Z., & Han, Q. L. (2009). Coherent artificial stock market model based on small world
networks. Complex Systems and Complexity Science, 2, 70–76.
9. Alfarano, S., & Milaković, M. (2009). Network structure and N-dependence in agent-based
herding models. Journal of Economic Dynamics and Control, 33(1), 78–92.
10. Tedeschi, G., Iori, G., & Gallegati, M. (2009). The role of communication and imitation in
limit order markets. The European Physical Journal B, 71(4), 489–497.
11. Barabási, A. L., & Albert, R. (1999). Emergence of scaling in random networks. Science,
286(5439), 509–512.
12. Watts, D. J., & Strogatz, S. H. (1998). Collective dynamics of small-world networks. Nature,
393(6684), 440–442.
13. Sornette, D., & Zhou, W. X. (2006). Importance of positive feedbacks and overconfidence
in a self-fulfilling Ising model of financial markets. Physica A: Statistical Mechanics and Its
Applications, 370(2), 704–726.
The Transition from Brownian Motion to
Boom-and-Bust Dynamics in Financial
and Economic Systems

Harbir Lamba

Abstract Quasi-equilibrium models for aggregate or emergent variables over long

periods of time are widely used throughout finance and economics. The validity of
such models depends crucially upon assuming that the system participants act both
independently and without memory. However important real-world effects such as
herding, imitation, perverse incentives, and many of the key findings of behavioral
economics violate one or both of these key assumptions.
We present a very simple, yet realistic, agent-based modeling framework that
is capable of simultaneously incorporating many of these effects. In this paper we
use such a model in the context of a financial market to demonstrate that herding
can cause a transition to multi-year boom-and-bust dynamics at levels far below a
plausible estimate of the herding strength in actual financial markets. In other words,
the stability of the standard (Brownian motion) equilibrium solution badly fails a
“stress test” in the presence of a realistic weakening of the underlying modeling
assumptions.
The model contains a small number of fundamental parameters that can be easily
estimated and require no fine-tuning. It also gives rise to a novel stochastic particle
system with switching and re-injection that is of independent mathematical interest
and may also be applicable to other areas of social dynamics.

Keywords Financial instability · Far-from-equilibrium · Endogenous dynamics ·

Behavioral economics · Herding · Boom-and-bust

H. Lamba ()
Department of Mathematical Sciences, George Mason University, Fairfax, VA, USA
e-mail: hlamba@gmu.edu

© Springer Nature Switzerland AG 2018 193

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_10
194 H. Lamba

1 Introduction

In the physical sciences using a stochastic differential equation (SDE) to model the
effect of exogenous noise upon an underlying ODE system is often straightforward.
The noise consists of many uncorrelated effects whose cumulative impact is well-
approximated by a Brownian process Bs , s ≥ 0 and the ODE df = a(f, t) dt is
replaced by an SDE df = a(f, t) dt + b(f, t) dBt .
However, in financial and socio-economic systems the inclusion of exogenous
noise (i.e., new information entering the system) is more problematic—even if the
noise itself can be legitimately modeled as a Brownian process. This is because
such systems are themselves the aggregation of many individuals or trading entities
(referred to as agents) who typically
(a) interpret and act differently to new information,
(b) may act differently depending upon the recent system history (i.e., non-
Markovian behavior), and
(c) may not act independently of each other.
The standard approach in neoclassical economics and modern finance is simply
to “average away” these awkward effects by assuming the existence of a single
representative agent as in macroeconomics [7], or by assuming that the averaged
reaction to new information is correct/rational, as in microeconomics and finance
[4, 13]. In both cases, the possibility of significant endogenous dynamics is removed
from the models resulting in unique, Markovian (memoryless), (quasi)-equilibrium
solutions. This procedure is illustrated in Fig. 1 where the complicated “human
filter” that lies between the new information and the aggregate variables (such as
price) does not alter its Brownian nature. This then justifies the use of SDEs upon
aggregate variables directly.

Fig. 1 In standard financial (Random) External Influences

and economic models
exogenous effects act directly
upon the model variables.
This is achieved by assuming
that the system participants
behave independently and are Idealized Humans
(at least when averaged) It’s "As if" they are: Averaged
perfectly rational and Away
Perfectly rational/no psychology
memoryless No memory
Uncoupled

Effect on Prices
(or other aggregate quantities)
From Brownian Motion to Boom-and-Bust Dynamics 195

(Random) External Influences

Actual Humans Internal

There will be systemic non−averagable effects: Dynamics
Systemic biases
Memory effects
(including
Behaviour of individuals is not independent network
effects)

Effect on Prices
(or other aggregate quantities)

Fig. 2 A more realistic “human filter” is capable of introducing endogenous, far-from-equilibrium,

dynamics into the system and economic history suggests that these are often of the boom-and-
bust type. Unfortunately, the difference between such dynamics and the instantly-equilibrating
models represented in Fig. 1 may only be apparent over long timescales or during extreme events
corresponding to endogenous cascading processes

However, the reality is far more complicated, as shown in Fig. 2. Important

human characteristics such as psychology, memory, systemic cognitive or emotional
biases, adaptive heuristics, group influences, and perverse incentives will be present,
as well as various possible positive feedbacks caused by endogenous dynamics or
interactions with the aggregate variables.
In an attempt to incorporate some of these effects, many heterogeneous agent
models (HAMs) have been developed [6] that simulate agents directly rather than
operate on the aggregate variables. These have demonstrated that it is relatively
easy to generate aggregate output data, such as the price of a traded asset, that
approximate reality better than the standard averaging-type models. In particular
the seemingly universal “stylized facts” [2, 11] of financial markets such as
heteroskedasticity (volatility clustering) and leptokurtosis (fat-tailed price-return
distributions resulting from booms-and-busts) have been frequently reproduced.
However, the effects of such research upon mainstream modeling have been minimal
perhaps, in part, because some HAMs require fine-tuning of important parameters,
others are too complicated to analyze, and the plethora of different HAMs means
that many are mutually incompatible.
The purpose of this paper is rather different from other studies involving HAMs.
We are not proposing a stand-alone asset pricing model to replace extant ones
(although it could indeed be used that way). Rather we are proposing an entire
framework in which the stability of equilibrium models can be tested in the presence
of a wide class of non-standard perturbations that weaken various aspects of the
efficiency and rationality assumptions—in particular those related to (a), (b), and
(c) above.
196 H. Lamba

Fig. 3 The M signed y

particles are subject to a Bulk stochastic motion
horizontal stochastic forcing
All particles diffuse
and they also diffuse
Minority particles
independently. Minority
particles also drift downwards
also drift down D
at a rate proportional to the
imbalance. When a particle
hits the boundary it is
re-injected with the opposite Reinject
sign and a kick is added to the and switch
bulk forcing that can trigger a
cascade (see text)
x

In this paper we focus upon the effects of herding as it is an easily understood

phenomenon that has multiple causes (rational, psychological, or due to perverse
incentives) and is an obvious source of lack of independence between agents’
actions. We start by introducing a simplified version of the modeling framework
introduced in [9, 10] that can also be described as a particle system in two
dimensions (Fig. 3). A web-based interactive simulation of the model can be
found at http://math.gmu.edu/~harbir/market.html. It provides useful intuition as
to how endogenous multi-year boom-and-bust dynamics naturally arise from the
competition between equilibrating and disequilibrating forces that are usually only
considered important over much shorter timescales.

2 A Stochastic Particle System with Re-injection

and Switching

We define the open set D ⊂ 2 by D = {(x, y) : −y < x < y, y > 0}. There
are M signed particles (with states +1 or −1) that move within D subject to four
different motions. Firstly, there is a bulk Brownian forcing Bt in the x-direction
that acts upon every particle. Secondly, each particle has its own independent two-
dimensional diffusion process. Thirdly, for agents in the minority state only, there is
a downward (negative y-direction) drift that is proportional to the imbalance.
Finally, when a particle hits the boundary ∂D it is re-injected into D with
the opposite sign according to some predefined probability measure. When this
happens, the position of the other particles is kicked in the x-direction by a (small)
amount ± 2κM , κ > 0, where the kick is positive if the switching particle goes from
the −1 state to +1 and negative if the switch is in the opposite direction. Note that
the particles do not interact locally or collide with one another.
From Brownian Motion to Boom-and-Bust Dynamics 197

2.1 Financial Market Interpretation

We take as our starting point the standard geometric Brownian motion (gBm) model
of an asset price pt at time t with p0 = 1. It is more convenient to use the log-
price rt = ln pt which for constant drift a and volatility b is given by the solution
rt = at + bBt to the SDE

drt = a dt + b dBt . (1)

Note that the solution rt depends only upon the value of the exogenous Brownian
process Bt at time t and not upon {Bs }ts=0 . This seemingly trivial observation
implies that rt is Markovian and consistent with various notions of market efficiency.
Thus gBm can be considered a paradigm for economic and financial models in
which the aggregate variables are assumed to be in a quasi-equilibrium reacting
instantaneously and reversibly to new information.
The model involves two types of agent and a separation of timescales. “Fast”
agents react near instantaneously to the arrival of new information Bt . Their effect
upon the asset price is close to the standard models and they will not be modeled
directly. However, we posit the existence of M “slow” agents who are primarily
motivated by price changes rather than new information and act over much longer
timescales (weeks or months). At time t the ith slow agent is either in state si (t) =
+1 (owning the asset) or si (t) = −1 (not owning the asset) and the sentiment σ (t) ∈
M
[−1, 1] is defined as σ (t) = M1
i=1 si (t). The ith slow agent is deemed to have an
evolving strategy that at time t consists of an open interval (Li (t), Ui (t)) containing
the current log-price rt (see Fig. 4). The ith agent switches state whenever the price
crosses either threshold, i.e., rt = Li (t) or Ui (t), and a new strategy interval is
generated straddling the current price. Note that slow agents wishing to trade do not
need to be matched with a trading partner—it is assumed that the fast agents provide
sufficient liquidity.
We assume in addition that each threshold for every slow agent has its own
independent diffusion with rate αi (corresponding to slow agents’ independently
evolving strategies) and those in the minority (i.e., whose state differs from |σ |) also
have their lower and upper thresholds drift inwards each at a rate Ci |σ |, Ci ≥ 0.

state = −1

Lj (t) Li (t) Uj (t) U i(t)

log−price
r(t)
state = +1

Fig. 4 A representation of the model showing two agents in opposite states at time t. Agent i is
in the +1 state and is represented by the two circles at Li (t) and Ui (t) while agent j is in the −1
state and is represented by the two crosses
198 H. Lamba

These herding constants Ci are crucial as they provide the only (global) coupling
between agents. The inward drift of the minority agents’ strategies makes them more
likely to switch to join the majority—they are being pressured/squeezed out of their
minority view. Herding and other mimetic effects appear to be a common feature
of financial and economic systems. Some causes are irrationally human while
others may be rational responses by, for example, fund managers not wishing to
deviate too far from the majority opinion and thereby risk severely under-performing
their average peer performance. The reader is directed to [9] for a more detailed
discussion of these and other modeling issues.
Finally, changes in the sentiment σ feed back into the asset price so that gBm (1)
is replaced with

drt = a dt + b dBt + κΔσ (2)

where κ > 0 and the ratio κ/b is a measure of the relative impact upon rt of
exogenous information versus endogenous dynamics. Without loss of generality we
let a = 0 and b = 1 by setting the risk-free interest rate to zero and rescaling time.
One does not need to assume that all the slow agents are of equal size, have
equal strategy-diffusion, and equal herding propensities. But if one does set αi = α
and Ci = C ∀i, then the particle system above is obtained by representing the ith
agent, not as an interval on , but as a point in D ⊂ 2 with position (xi , yi ) =
( Ui +L
2 −rt ,
i Ui −Li
2 ). To make the correspondence explicit: the bulk stochastic motion
is due to the exogenous information stream, the individual diffusions are caused by
strategy-shifting of the slow agents; the downward drift of minority agents is due to
herding; the re-injection and switching are the agents changing investment position;
and the kicks that occur at switches are due to the change in sentiment affecting the
asset price via the linear supply/demand price assumption.

2.2 Limiting Values of the Parameters

There are different parameter limits that are of interest.

1. M → ∞ In the continuum limit the particles are replaced by a pair of evolving
density functions ρ + (x, y, t) and ρ − (x, y, t) representing the density of each
agent state on D—such a mesoscopic Fokker–Planck description of a related,
but simpler, market model can be found in [5]. The presence of nonstan-
dard boundary conditions, global coupling, and bulk stochastic motion present
formidable analytic challenges for even the most basic questions of existence
and uniqueness of solutions. However, numerical simulations strongly suggest
that, minor discretization effects aside, the behavior of the system is independent
of M for M > 1000.
From Brownian Motion to Boom-and-Bust Dynamics 199

2. Bt → 0 As the external information stream is reduced the system settles

into a state where σ is close to either ±1. Therefore this potentially useful
simplification is not available to us.
3. α → 0 or ∞ In the limit α → 0 the particles do not diffuse, i.e., the agents do not
alter their thresholds between trades/switches. This case was examined in [3] and
the lack of diffusion does not significantly change the boom–bust behavior shown
below. On the other hand, for α  max(1, C) the diffusion dominates both the
exogenous forcing and the herding/drifting and equilibrium-type dynamics is re-
established. This case is unlikely in practice since slow agents will alter their
strategies more slowly than changes in the price of the asset.
4. C → 0 This limit motivates the next section. When C = 0 the particles are
uncoupled and if the system is started with approximately equal distributions
of ±1 states, then σ remains close to 0. Thus (2) reduces to (1) and the
particle system becomes a standard equilibrium model—agents have differing
expectations about the future which causes them to trade but on average the price
remains “correct.” In Sect. 3 we shall observe that endogenous dynamics arise
as C is increased from 0 and the equilibrium gBm solution loses stability in the
presence of even small amounts of herding.
5. κ → 0 For κ > 0 even one agent switching can cause an avalanche of similar
switches, especially when the system is highly one-sided with |σ | close to 1.
When κ = 0 the particles no longer provide kicks (or affect the price) when they
switch although they are still coupled via C > 0. The sentiment σ can still drift
between −1 and +1 over long timescales but switching avalanches and large,
sudden, price changes do not occur.

3 Parameter Estimation, Numerical Simulations, and the

Instability of Geometric Brownian Pricing

In all the simulations below we use M = 10,000 and discretize using a timestep
h = 0.000004 which corresponds to approximately 1/10 of a trading day if one
assumes a daily standard deviation in prices of ≈0.6% due to new information.
The price changes of ten consecutive timesteps are then summed to give daily price
return data making the difference between synchronous vs asynchronous updating
relatively unimportant.
We choose α = 0.2 so that slow agents’ strategies diffuse less strongly than the
price does. A conservative choice of κ = 0.2 means that the difference in price
between neutral (σ = 0) and polarized markets σ = ±1 is, from (2), exp(0.2) ≈
22%.
After switching, an agent’s thresholds are chosen randomly from a uniform
distribution to be within 5% and 25% higher and lower than the current price. This
allows us to estimate C by supposing that in a moderately polarized market with
|σ | = 0.5 a typical minority agent (outnumbered 3–1) would switch due to herding
200 H. Lamba

Fig. 5 The top left figure shows the prices p(t) for both our model and the gBm pricing model
(1) with the same exogenous information stream Bs . The herding model is clearly more volatile
but the other pictures demonstrate the difference more clearly. In the bottom left, the same data
is plotted in terms of daily price changes. Over the 40-year period not a single instance of a daily
price change greater than 2% occurred in the gBm model. All the large price fluctuations are due
to endogenous dynamics in the herding model. This is shown even more clearly in the top right
picture where sentiment vs time is plotted for the herding model—the very sudden large switches
in sentiment are due to cascading changes amongst the agents’ states. It should be noted that the
sentiment can remain polarized close to ±1 for unpredictable and sometimes surprisingly long
periods of time. Finally, the bottom right picture shows the cumulative log–log plot of daily price
changes that exceed a given percentage for each model. The fat-tailed distribution for the herding
model shows that the likelihood of very large price moves is increased by orders of magnitude over
the gBm model

pressure after approximately 80 trading days (or 3 months, a typical reporting period
for investment performance) [14]. The calculation 80C|σ | = | ln(0.85)|/0.00004
gives C ≈ 100. Finally, we note that no fine-tuning of the parameters is required for
the observations below.
Figure 5 shows the results of a typical simulation, started close to equilibrium
with agents’ states equally mixed and run for 40 years. The difference in price
history between the above parameters and the equilibrium gBm solution is shown in
the top left. The sudden market reversals and over-reactions can be seen more clearly
in the top right plot where the market sentiment undergoes sudden shifts due to
From Brownian Motion to Boom-and-Bust Dynamics 201

0.9
Average Maximum Sentiment |σmax|

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 5 10 15 20 25 30 35 40
Coupling (Herding) C

Fig. 6 A measure of disequilibrium |σ |max averaged over 20 runs as the herding parameter C
changes

switching cascades. These result in price returns (bottom left) that could quite easily
bankrupt anyone using excessive financial leverage and gBm as an asset pricing
model! Finally in the bottom right the number of days on which the magnitude of
the price change exceeds a given percentage is plotted on log–log axes. It should be
emphasized that this is a simplified version of the market model in [9] and an extra
parameter that improves the statistical agreement with real price data (by inducing
volatility clustering) has been ignored.
To conclude we examine the stability of the equilibrium gBm solution using
the herding level C as a bifurcation parameter. In order to quantify the level of
disequilibrium in the system we record the maximum value of |σ | ignoring the first
10 years of the simulation (to remove any possible transient effects caused by the
initial conditions) and average over 20 runs each for values of 0 ≤ C ≤ 40. All the
other parameters and the initial conditions are kept unchanged.
The results in Fig. 6 show that even for values of C as low as 20 the deviations
from the equilibrium solution are close to being as large as the system will allow
with |σ | usually getting close to ±1 at some point during the simulations. To
reiterate, this occurs at a herding strength C which is a factor of 5 lower than
the value of C = 100 estimated above for real markets! It should also be noted
that there are other significant phenomena that have not been included, such as
new investors and money entering the asset market after a bubble has started, and
localized interactions between certain subsets of agents. These can be included in
the model by allowing κ to vary (increasing at times of high market sentiment, for
example) and, as expected, they cause the equilibrium solution to destabilize even
more rapidly.
202 H. Lamba

4 Conclusions

Financial and economic systems are subject to many different kinds of inter-
dependence between agents and potential positive feedbacks. However, even those
mainstream models that attempt to quantify such effects [1, 14] assume that the
result will merely be a shift of the equilibria to nearby values without qualitatively
changing the nature of the system. We have demonstrated that at least one such
form of coupling (incremental herding pressure) results in the loss of stability
of the equilibrium. Furthermore the new dynamics occurs at realistic parameters
and is clearly recognizable as “boom-and-bust.” It is characterized by multi-year
periods of low-level endogenous activity (long enough, certainly, to convince
equilibrium-believers that the system is indeed in an equilibrium with slowly
varying parameters) followed by large, sudden, reversals involving cascades of
switching agents triggered by price changes.
A similar model was studied in [8] where momentum-traders replaced the slow
agents introduced above. The results replicated the simulations above in the sense
that the equilibrium solution was replaced with multi-year boom-and-bust dynamics
but with the added benefit that analytic solutions can be derived, even when agents
are considered as nodes on an arbitrary network rather than being coupled globally.
The model presented here is compatible with existing (non-mathematized)
critiques of equilibrium theory by Minsky and Soros [12, 15]. Furthermore, work
on related models to appear elsewhere shows that positive feedbacks can result
in similar non-equilibrium dynamics in more general micro- and macro-economic
situations.

Acknowledgements The author thanks Michael Grinfeld, Dmitri Rachinskii, and Rod Cross
for numerous enlightening conversations and Julian Todd for writing the browser-accessible
simulation of the particle system.

References

1. Akerlof, G., & Yellen, J. (1985). Can small deviations from rationality make significant
differences to economic equilibria? The American Economic Review, 75(4), 708–720.
2. Cont, R. (2001). Empirical properties of asset returns: Stylized facts and statistical issues.
Quantitative Finance, 1, 223–236.
3. Cross, R., Grinfeld, M., Lamba, H., & Seaman, T. (2005). A threshold model of investor
psychology. Physica A, 354, 463–478.
4. Fama, E. F. (1965). The behavior of stock market prices. Journal of Business, 38, 34–105.
5. Grinfeld, M., Lamba, H., & Cross, R. (2013). A mesoscopic market model with hysteretic
agents. Discrete and Continuous Dynamical Systems B, 18, 403–415.
6. Hommes, C. H. (2006). Handbook of computational economics (Vol. 2, pp. 1109–1186).
Amsterdam: Elsevier.
7. Kirman, A. (1992). Who or what does the representative agent represent? Journal of Economic
Perspectives, 6, 117–136.
From Brownian Motion to Boom-and-Bust Dynamics 203

8. Krejčí, P., Melnik, S., Lamba, H., & Rachinskii, D. (2014). Analytical solution for a class of
network dynamics with mechanical and financial applications. Physical Review E, 90, 032822.
9. Lamba, H. (2010). A queueing theory description of fat-tailed price returns in imperfect
financial markets. The European Physical Journal B, 77, 297–304.
10. Lamba, H., & Seaman, T. (2008). Rational expectations, psychology and learning via moving
thresholds. Physica A: Statistical Mechanics and its Applications, 387, 3904–3909.
11. Mantegna, R., & Stanley, H. (2000). An introduction to econophysics. Cambridge: Cambridge
University Press.
12. Minsky, H. P. (1992). The Financial Instability Hypothesis. The Jerome Levy Institute Working
Paper, 74.
13. Muth, J. A. (1961). Rational expectations and the theory of price movements. Econometrica,
6. https://doi.org/10.2307/1909635
14. Scharfstein, D., & Stein, J. (1990). Herd behavior and investment. The American Economic
Review, 80(3), 465–479.
15. Soros, G. (1987). The alchemy of finance. New York: Wiley.
Product Innovation and Macroeconomic
Dynamics

Christophre Georges

Abstract We develop an agent-based macroeconomic model in which product

innovation is the fundamental driver of growth and business cycle fluctuations. The
model builds on a hedonic approach to the product space and product innovation
developed in Georges (A hedonic approach to product innovation for agent-based
macroeconomic modeling, 2011).

Keywords Innovation · Growth · Business cycles · Agent-based modeling ·

Agent-based macroeconomics

1 Introduction

Recent evidence points to the importance of product quality and product innovation
in explaining firm level dynamics. In this paper we develop an agent-based
macroeconomic model in which both growth and business cycle dynamics are
grounded in product innovation. We take a hedonic approach to the product space
developed in [23] that is both simple and flexible enough to be suitable for modeling
product innovation in the context of a large-scale, many-agent macroeconomic
model.
In the model, product innovation alters the qualities of existing goods and
introduces new goods into the product mix. This novelty leads to further adaptation
by consumers and firms. In turn, both the innovation and adaptation contribute
to complex market dynamics. Quality adjusted aggregate output exhibits both
secular endogenous growth and irregular higher frequency cycles. There is ongoing
churning of firms and product market shares, and the emerging distribution of these
shares depends on opportunities for niching in the market space.

C. Georges ()
Hamilton College, Department of Economics, Clinton, NY, USA
e-mail: cgeorges@hamilton.edu

© Springer Nature Switzerland AG 2018 205

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_11
206 C. Georges

2 Background

Recent research suggests that product innovation is a pervasive force in modern

advanced economies. For example, Hottman et al. [28] provide evidence that
product innovation is a central driver of firm performance. They offer an accounting
decomposition that suggests that 50–70% of the variance in firm size at the
aggregate level can be attributed to differences in product quality, whereas less than
25% can be attributed to differences in costs of production. Further, in their analysis,
individual firm growth is driven predominantly by improvements in product quality.
Similarly, Foster et al. [21] found that firm level demand is a more powerful driver
of firm survival than is firm level productivity. Broda and Weinstein [7] and Bernard
et al. [6] further document the substantial pace of churning in product markets with
high rates of both product creation and destruction and changes of product scope at
the firm level.
We explore the implications of product innovation for growth and business cycle
fluctuations in an agent-based macroeconomic model. While there is a literature
(e.g., [24]) that attributes economic growth to growth in product variety, variety is
only one expression of product innovation. Both the product turnover and skewed
distributions of firm sizes and product market shares that we observe indicate that it
is conventional for some products to drive out other products and develop outsized
market shares due to superiority in perceived quality. Our agent-based approach
is well suited to model the types of heterogeneity and churning dynamics that we
observe empirically.
Our approach revisits [31, 32]. Preferences are defined over a set of product
characteristics, and products offer various bundles of these characteristics. As Lan-
caster notes, the characteristics approach allows for new products to be introduced
seamlessly, as new products simply offer new possibilities for the consumption of
an unchanging set of characteristics.
Of course there is a large literature on product innovation, and there are a number
of existing models that bear some relation to the one developed here. See, for
example, [2, 4, 11–16, 25, 30, 33, 34, 36–38]. For comparisons with the current
approach, see [23].
The current paper is in the recent tradition of agent-based macroeconomics. This
literature builds macroeconomic models from microfoundations, but in contrast to
standard macroeconomic practice, treats the underlying agents as highly heteroge-
neous, boundedly rational, and adaptive, and does not assume a priori that markets
clear. See, e.g., [17–20, 26].

3 The Macroeconomic Environment

Our goal is to understand the role of product innovation in driving growth and
fluctuations in a very simple macroeconomic environment. Here are the fundamental
features of the model.
Product Innovation and Macroeconomic Dynamics 207

• There are n firms, each of which produces one type of good at any time.
• There are m characteristics of goods that consumers care about. Each good
embodies distinct quantities of these characteristics at any given time. Product
innovation affects these quantities.
• The probability that a firm experiences a product innovation at any time depends
on its recent investments in R&D, which in turn is the outcome of a discrete
choice rule.
• Each firm produces with overhead and variable labor. It forecasts the final
demand for its product by extrapolating from recent experience, sets its price
as a constant mark-up over marginal cost, and plans to produce enough of its
good to meet its expected final demand given this price.
• There is a single representative consumer who spends all of her labor income
each period on consumption goods and searches for better combinations of
products to buy within her budget.
• If a firm becomes insolvent, it exits the market and is replaced by a new entrant.

4 Consumer Preferences

The representative consumer’s preferences are defined on the characteristics space.

Specifically, the momentary utility from consuming the vector z ∈ Rm of hedonic
characteristics is u(z).1
In addition to this utility function, the consumer has a home production function
g(q) that maps bundles q ∈ Rn of products into perceived bundles z ∈ Rm of the
characteristics.2
More specifically, we assume that the representative consumer associates with
each good i a set of base characteristic magnitudes z-basei ∈ Rm per unit of
the good, as well as a set of complementarities with other goods. If the consumer
associates good k as complementary to good i, then she associates with the goods
pair (i,k) an additional set of characteristic magnitudes z-compi,k ∈ Rm , per
composite unit qi,k = θ (qi , qk ) of the two goods (defined below).3
Intuitively, a box of spinach may offer a consumer certain quantities of sub-
jectively valued characteristics like nutrition, flavor, and crunchiness. However,
the flavor characteristic might also be enhanced by consuming the spinach in
combination with a salad dressing, so that the total quantity of flavor achieved by

1 While we are working with a representative consumer in the present paper for convenience, it
is a simple step in the agent-based modeling framework to relax that assumption and allow for
idiosyncratic variation of consumer preferences.
2 This is essentially the approach taken by Lancaster, and shares some similarities with others such

as [5, 35]. The primary deviation of our approach from that of Lancaster is the construction of our
home production function g(q).
3 This vector is associated with good i, and it is convenient to assume that the complementarities

are independent across goods (i.e., that the vectors z-compi,k and z-compk,i are independent).
208 C. Georges

eating these in combination is greater than the sum of the flavor quantities from
consuming each separately.
Similarly, in isolation, an Apple iPad may provide a consumer some modest
degree of entertainment, but this entertainment value is dramatically enhanced
by consuming it along with a personal computer, an internet access subscription,
electricity, apps, and so on.
We assume that both base characteristic magnitudes and complementary charac-
teristic magnitudes are additive at the level of the individual good. Thus, for good i
and hedonic characteristic j , the consumer perceives

zi,j = z-basei,j · qi + z-compi,j,k · qi,k . (1)
k

These characteristic magnitudes are then aggregated over products by a CES

aggregator:
 1/ρ1

n
ρ
zj = zi,j1 (2)
i=1

with ρ1 < 1. Equations (1) and (2) define the mapping g(q) introduced above. The
CES form of (2) introduces some taste for variety across products.4
We assume that the utility function u for the representative consumer over
hedonic characteristics is also CES, so that
⎡ ⎤1/ρ2

m
u=⎣ (zj + z̄j )ρ2 ⎦ (3)
j =1

where z̄j is a shifter for characteristic j (see [29]), and ρ2 < 1. Thus, utility takes
a nested CES form. Consumers value variety in both hedonic characteristics and in
products.
Finally, we specify the aggregator for complements θ (qi , qk ) as floor(min(qi ·
λ , qk · λ )) · λ. I.e., complementarities are defined per common (fractional) unit λ
1 1

consumed.5

4 Note that if ρ1 = 1, the number of viable products in the economy would be strongly limited by
the number of hedonic elements, as in Lancaster, who employs a linear activity analysis to link
goods and characteristics.
5 Note that this introduces a (fractional) integer constraint on the consumer’s optimization and

search problem. λ > 0 but need not be less than one.

Product Innovation and Macroeconomic Dynamics 209

5 Product Innovation

A product innovation takes the form of the creation of a new or improved product
that, from the point of view of the consumer, combines a new set of characteristics,
or enhances an existing set of characteristics, when consumed individually or jointly
with other products. The new product will be successful if it is perceived as offering
utility (in combination with other goods) at lower cost than current alternatives. The
product may fail due to high cost, poor search by consumers, or poor timing in terms
of the availability or desirability of other (complementary and non-complementary)
goods.
In the present paper, at any time the base and complementary set of hedonic
characteristic magnitudes (z-basei and z-compi,k ) associated by the consumer with
good i are coded as m dimensional vectors of integers. These characteristics vectors
are randomly initialized at the beginning of the simulation.
A product innovation is then a set of random (integer) increments (positive or
negative) to one or more elements of z-basei or z-compi,k . Product innovation for
continuing firms is strictly by mutation. Product innovations can be positive or
negative. I.e., firms can mistakenly make changes to their products that consumers
do not like. However, there is a floor of zero on characteristic values. Further,
innovations operate through preferential attachment; for a firm that experiences a
product innovation, there is a greater likelihood of mutation of non-zero hedonic
elements.6
The probability that a firm experiences product innovation in any given period t
is increasing in its recent R&D activity.

6 R&D

The R&D investment choice is binary—in a given period a firm either does or does
not engage in a fixed amount of R&D. If a firm engages in R&D in a given period,
it incurs additional overhead labor costs R in that period.
In making its R&D investment decision at any time, the firm compares the recent
profit and R&D experiences of other firms and acts according to a discrete choice
rule. Specifically, firms observe the average recent profits πH and πL of other firms
with relatively high and low recent R&D activity.7 Firms in the lower profit group
switch their R&D behavior with a probability related to the profitability differential
between the two groups. Specifically, they switch behavior with probability 2Φ − 1,
where

6 This
weak form of preferential attachment supports specialization in the hedonic quality space.
7 Eachfirm’s recent profits and recent R&D activity are (respectively) measured as exponentially
weighted moving averages of its past profits and R&D activity.
210 C. Georges

e γ π1
Φ=
e γ π1 + e γ π2

γ > 0 measures the intensity of choice and π1 and π2 are measures of the average
recent profits of the high and low profit R&D groups.8 There is additionally some
purely random variation in R&D choice.

7 Production and Employment

Each firm i produces its good with labor subject to a fixed labor productivity Ai and
hires enough production labor to meet its production goals period by period. Each
firm also incurs a fixed overhead labor cost H in each period and additional overhead
R&D labor cost R in any period in which it is engaged in R&D. In this paper, our
focus is on product innovation rather than process innovation. Consequently, we
suppress process innovation and hold Ai , H , and R constant over time.9

8 Consumer Search

The consumer spends her entire income each period and selects the shares of
her income to spend on each good. Each period, she experiments with random
variations on her current set of shares. Specifically, she considers randomly shifting
consumption shares between some number of goods, over some number of trials,
and selects among those trials the set of shares that yields the highest utility with
her current income. While the consumer engages in limited undirected search and is
not able to globally maximize utility each period, she can always stick to her current
share mix, and so never selects new share mixes that reduce utility below the utility
afforded by the current one. I.e., the experimentation is a thought exercise not an act
of physical trial and error.

9 Entry and Exit

When a firm does not have enough working capital to finance production, it shuts
down and is replaced. The new firm adopts (imitates) the product characteristics of a
randomly selected existing firm, retains the exiting firm’s current share of consumer
demand, and is seeded with startup capital.

if πH > πL , then π1 = πH and π2 = πL , and firms with relatively low recent R&D activity
8 I.e.,

switch R&D on with a probability that is greater the larger is the difference between π1 and π2 .
9 Process innovation that affects these parameters across firms and over time can easily be

introduced.
Product Innovation and Macroeconomic Dynamics 211

10 Timing

The timing of events within each period t is as follows:

• R&D: firms chose their current R&D investment levels (0 or 1).
• Innovation: firms experience product innovation with probabilities related to their
recent R&D investments.
• Production: firms forecast sales, hire production labor, and produce to meet
forecasted demand.
• Incomes: wages and salaries are paid to all labor employed (production, standard
overhead, R&D overhead).
• Consumer search: the representative consumer searches and updates the con-
sumption basket.
• Sales: the consumer spends all of the labor income (above) on the consumption
basket.
• Entry and Exit: firms with insufficient working capital are replaced.

11 Features of the Model

If we eliminate all heterogeneity in the model, there is a steady state equilibrium

which is stable under the simple dynamics described above. Assuming that all
firms engage in R&D investment, steady state aggregate output is a multiple of the
combined overhead labor cost H + R and is independent of product quality. See
the Appendix for details. Quality adjusted output and consumer utility, on the other
hand, will grow over time in equilibrium as product quality improves.
With firm heterogeneity, the evolution of the hedonic characteristics of firms’
products will influence product market shares as well as aggregate activity and living
standards.
The stochastic evolution of product quality with zero reflective lower bounds
on individual characteristic magnitudes will tend to generate skewness in the
distribution of individual product qualities and thus in the distribution of market
shares.
Working against the skewness of the distribution of product shares is the variety-
loving aspect of consumer preferences and the number of product characteristics
and complementarities. The greater the number of characteristics, the greater the
opportunities for individual firms to find niches in the characteristics space. A small
number of characteristics relative to the number of products creates a “winner
take all” environment in which firms compete head to head in a small number of
dimensions, whereas a large number of characteristics may create opportunity for
the development of a (so-called) “long tail” of niche products.10

10 A recent literature argues that the growth of internet retail has allowed niche products that better
suit existing consumer preferences to become profitable, eroding the market shares of more broadly
popular “superstar” products [8, 9].
212 C. Georges

Turning to aggregate activity, with heterogeneity, there are several channels

through which product innovation may be a source of output and utility fluctua-
tions.11
First, product innovation produces ongoing changes in the pattern of demand
across goods. This may affect aggregate output for several reasons. For one, it takes
time for production to adjust to the changing demand pattern resulting in both direct
and indirect effects on output. Additionally, as demand patterns change, demand
may shift between goods with different production technologies.12 Further, if
network effects are large or product innovation results in highly skewed distributions
of firm sizes and product shares, then idiosyncratic firm level fluctuations may
not wash out in the aggregate, even as the number of firms and products grows
large.13
A second mechanism is that R&D investment has a direct impact on aggregate
demand by increasing labor income and thus consumer spending on all goods with
positive demand shares. Since R&D investment decisions are conditioned by social
learning, there can be cyclical herding effects in this model.
Finally, quality adjusted output and consumer utility will fluctuate and grow due
to the evolution of both physical output and product qualities.

12 Some Simulation Results

Runs from a representative agent baseline version of the model behave as expected.
If all firms have the same parameters and the same productivities and hedonic
qualities and all engage in R&D, then output converges to the analytical equilibrium
discussed in the Appendix.
For the heterogeneous firms model, ongoing endogenous innovation tends to
generate stochastic growth in consumer utility and ongoing fluctuations in total
output. For the simulation experiments discussed here, as the number of firms
increases, the output fluctuations become increasingly dominated by variations in
R%D investment spending through the herding effect noted above.

11 We exclude other sources of business cycle fluctuations and growth from the model to focus on
the role of product innovation.
12 Below, we suppress the latter effect, standardizing productivity across firms in order to focus

more directly on product innovation.

13 Intuitively, if some firms or sectors remain large and/or central to the economy, even under

highly disaggregated measurement, then idiosyncratic shocks will have aggregate effects. For
formalizations, see, for example, [1, 3, 10, 18, 22, 27].
Product Innovation and Macroeconomic Dynamics 213

1080

600
1060
Output

Utility
1040

400
1020
1000

200
18000 18500 19000 19500 20000 18000 18500 19000 19500 20000
time time

Fig. 1 Representative run. Output and utility, for rounds 18,000–20,000. Output is measured
weekly as average output for the last quarter. The number of hedonic characteristics is m = 50
and the intensity of choice for firms’ R&D decisions is γ = 0.2

Figure 1 is produced from a representative run with 1000 firms. In this case
the number of product characteristics is 50, ρ1 = ρ2 = 0.8, and mutation is
multiplicative.14 There is no variation in labor productivity Ai over time or across
firms i.
The time period is considered to be a week, so we are showing approximately
40 years of simulated data well after transitory growth dynamics have died out.
Aggregate output exhibits irregular fluctuations and cycles; peak to trough swings
in GDP are on the order of 1–3%. Utility also fluctuates with both output and the
evolution of the quality of products produced and consumed, but also grows due to
long-term net improvements in product quality. These net improvements are driven
both directly and indirectly by innovation. Existing firms’ product qualities may rise
or fall in response to innovations, but also as consumers shift toward higher quality
product mixes, less successful firms are driven from the market, and new firms enter.
The evolution of output in Fig. 1 follows the evolution of R&D investment
spending fairly closely (with correlation close to 0.9). There is substantial additional
variation of firm level output due to the churning of demand shares, with much of
that variation washing out in the aggregate.
When the number of firms engaging in R&D increases, spending on R&D
increases, driving up consumption, output, and utility.15 A second effect is that
more firms innovate over time, driving the rate of growth of utility upward. Figure 2
illustrates these level and growth effects on utility (equivalently, quality adjusted

14 The number of firm in this simulation is small (1000), but can easily be scaled up to several
million. Similarly the number of hedonic characteristics (50) can be increased easily (though in
both cases, of course, at some computational cost).
15 In the representative agent case, the multiplier for output is 1 , where η is the markup.
η−1
214 C. Georges

2500

35
2000

30
1500

Utility
R&D
1000

25
500

20
0

4000 5000 6000 7000 8000 9000 4000 5000 6000 7000 8000 9000
time time

Fig. 2 Run similar to that in Fig. 1, but with high intensity of choice γ = 10 for R&D decisions.
The number of firms n is 2500

output) in a run with the intensity of choice for firms’ R&D decisions increased
dramatically (from γ = 0.2 above, to γ = 10).
Here, as the relative profits of firms engaging in R&D evolves, there is strong
herding of the population of firms toward and away from engaging in R&D. These
shifts lead to shifts in both the level and growth rate of overall consumer utility.
Note that for the individual firm, R&D investment involves a tradeoff between its
current profit and its future potential profits. R&D spending reduces the firm’s profit
today (by R), but potentially raises its profit in the future via increased demand for
its product. Firms with relatively high demand shares can spread their overhead costs
over more output and make positive profit, while firms with chronically low demand
shares tend to run losses and ultimately fail due to their inability to cover even their
basic overhead costs (H ). Nevertheless, current R&D is a direct drag on current
profit, with the potential benefits via product innovation and increased demand
accruing in the future. This tension between the short and medium term costs and
benefits of R&D investment supports the complex patterns of R&D investment
behavior in the model.
Now consider the distribution of firm sizes as measured by final demand shares.
We start the simulations with characteristic magnitudes (z-basei and z-compi,k )
initialized to random sequences of zeros and ones, so that there is idiosyncratic
variation in product quality across firms. However, the representative consumer
initially sets equal shares across firms. As each simulation proceeds, the consumer
searches for better bundles of goods.
If there is neither product nor process innovation, then the optimum bundle for
the consumer is static, and demand is redistributed across firms over time, leading
to a skewed distribution of demand shares. The degree of skewness is influenced
by the CES utility elasticities. Lower values of ρ1 and ρ2 indicate a greater taste
for variety in characteristics and products, limiting the impact of product quality
Product Innovation and Macroeconomic Dynamics 215

150

300
100

200
Frequency

Frequency
100
50
0

0
0 .005 .01 .015 .02 0 50 100
share char10

Fig. 3 Distributions of product shares and values of hedonic characteristic 10 by firm at time
20,000 in the run in Fig. 1

on market shares. The shape of the distribution is also affected by the ratio m n of
product characteristics to goods, as well as the distribution of complementarities
and the entry and exit process. For example, if m n is small, then there is less room
for firms to have independent niches in the product space, and so the market tends
to become more concentrated.
Including ongoing endogenous product innovation leads to ongoing changes in
the relative qualities of goods and so ongoing churning in demand shares. The share
distribution follows a similar pattern to that above in early rounds. Starting from
a degenerate distribution with all the mass on 1/n, it first spreads out and then
becomes skewed as shares are progressively reallocated among products.
As innovation proceeds, and product churning emerges, the degree of skewness
continues to evolve. The ultimate limiting distribution depends on the stochastic
process for product innovation in addition to the factors above (ρ1 , ρ2 , m n , the
distribution of complementarities, and the entry and exit process).
Figure 3 shows the distribution of product shares and the values of one of the 50
hedonic characteristics (characteristic 10) over the firms at time 20,000 in the run in
Fig. 1. Here we exclude zero shares and zero characteristic values (held by 374 and
262 of the 1000 firms respectively) and display the distributions of non-zero values.

13 Conclusion

We have developed an agent-based macroeconomic model in which product innova-

tion is the fundamental driver of growth and business cycle fluctuations. The model
builds on a hedonic approach to the product space and product innovation proposed
in [23]. Ongoing R&D activities, product innovation, and consumer search yield
216 C. Georges

ongoing firm level and aggregate dynamics. Holding productivity constant, output
fluctuates but does not grow in the long run. However, utility, or equivalently quality
adjusted output, does exhibit long run endogenous growth as a result of product
innovation. The distribution of product market shares tends to become skewed,
with the degree of skewness depending on the opportunities for niching in the
characteristics space. As the number of firms grows large, business cycle dynamics
tend to become dominated by an innovation driven investment cycle.

Acknowledgements I am grateful to participants at CEF 2015 and a referee for useful comments,
suggestions, and discussions. All errors are mine.

Appendix: Representative Agent Benchmark

Consider the case in which all firms are identical and each starts with a 1/n share of
the total market. Suppose further that all firms engage in R&D in every period and
experience identical innovations over time.
In this case, the equal individual market shares will persist, since there is no
reason for consumers to switch between firms with identical product qualities and
identical prices. Further, there is a unique equilibrium for the real production and
sales of consumer goods Y at which demand and supply are in balance. At this
equilibrium, aggregate real activity depends on the markup η, the per firm overhead
labor costs H and R, the wage rate W (for production workers), labor productivity A
(for production workers), and the number of firms n. Specifically, at this equilibrium
Y = ( η−11 A
)· W ·(H +R)·n. See below for details. Further, this equilibrium is a steady
state of the agent dynamics in the model and is locally stable under those dynamics.
If, for example, firms all start with production less than steady state production, then
since the markup η > 1, demand will be greater than production for each firm, and
production will converge over time to the steady state equilibrium.
We can see this as follows. Since all firms are identical, they produce identical
quantities q of their goods. Then total labor income is:
!
W
E =n· ·q +H +R (4)
A

Each firm also charges an identical price p for its good, which is a markup η on
marginal cost

p = η · MC
W
MC = (5)
A
Product Innovation and Macroeconomic Dynamics 217

and so produces and sells

E
q= (6)
n·p

units of its good.

These relationships yield the following steady state equilibrium value for (per
firm) production.

(H + R) · A
q∗ = W
(7)
η−1

We can see that ∂q ∗ /∂(H + R) > 0, ∂q ∗ /∂A > 0, ∂q ∗ /∂W < 0, and ∂q ∗ /∂η <
0. These are all demand driven. An increase in the cost of overhead labor (H or R)
raises the incomes of overhead workers, raising AD and thus equilibrium output.
An increase in labor productivity A will cause firms to lower their prices, raising
aggregate demand and equilibrium output. Similarly an increase in the wage rate
W of production workers or in the mark-up η will cause firms to raise their prices,
lowering AD and equilibrium output.
Further, if in this representative agent case firms follow simple one period
adaptive expectations for demand, then firm level output dynamics are given by:

1 1 A
qt = · qt−1 + · (H + R) · (8)
η η W

Thus, given η > 1, the steady state equilibrium q ∗ is asymptotically stable.

Total market output in the steady state equilibrium is just

Y ∗ = n · q∗
n · (H + R) · A
= W
(9)
η−1

Clearly, this will be constant as long as there is no change in the parameters

H , R, A, W , η and n. If we were to allow them to change, growth in the number
of firms n or the productivity of production labor A would cause equilibrium total
production Y ∗ to grow over time, while balanced growth in production wages W and
labor productivity A would have no impact on equilibrium production. Note that, in
the full heterogeneous agent model, while all of the parameters above are fixed, the
fraction of firms adopting R&D investment varies endogenously, contributing to the
disequilibrium dynamics of the model.
Importantly for the present paper, note that the representative agent steady
state equilibrium production Y ∗ above is entirely independent of product quality.
Improvements in product quality will, however, increase consumer utility, or
equivalently, the quality adjusted value of total production at this equilibrium.
218 C. Georges

Specifically, given the nested CES formulation of utility, if the magnitudes of all
product characteristics grow at rate g due to innovation, then the growth rate of
consumer utility will converge in the long run to g. If the magnitudes of different
characteristics grow at different rates, the growth rate of utility will converge to
the rate of growth of the fastest growing characteristic. All else equal, the long run
growth path of utility will be lower in the latter case than in the former case.
It is also worth noting that, at the representative agent steady state equilibrium
above, firms make zero profit. Revenues net of the cost of production labor are just
great enough to cover all overhead labor costs. Once we move to the heterogeneous
firm case, in the comparable steady state equilibrium, profits are distributed around
zero across firms. Firms will vary as to R&D investment status, with firms who
are not engaging in R&D investment saving overhead cost R per period. Firms
will also vary with respect to demand shares (driven by product qualities which
are themselves related to past investments in R&D), and firms with relatively high
demand shares are able to spread overhead cost over greater sales. Thus, for two
firms with the same overhead cost (i.e., the same current R&D investment status),
the firm with greater demand for its product will have higher profit, while for two
firms with the same product demand shares, the one with the lower overhead cost
(lower current R&D investment) will have higher profit. Firms that face chronic
losses will eventually fail and be replaced.

References

1. Acemoglu, D., Carvalho, V. M., Ozdaglar, A., & Tahbaz-Selehi, A. (2012). The network origins
of aggregate fluctutations. Econometrica, 80(5), 1977–1016.
2. Acemoglu, D., Akcigit, U., Bloom, N., & Kerr, W. R. (2013). Innovation, reallocation and
growth. In NBER WP 18933.
3. Acemoglu, D., Akcigit, U., & Kerr, W. R. (2015). Networks and the macroeconomy: An
empirical exploration. In NBER macroeconomic annual (Vol. 30).
4. Akcigit, U., & Kerr, W. R. (2018). Growth through heterogeneous innovations. Journal of
Political Economy, 126(4), 1374–1443.
Akcigit, U., & Kerr, W. R. (2018). Growth through heterogeneous innovations. Journal of
Political Economy (Vol. 136) No.4
5. Becker, G. S. (1965). A theory of the allocation of time. Economic Journal, 75(299), 493–517.
6. Bernard, A. B., Redding, S. J., & Schott, P. K. (2010). Multi-product firms and product
switching. American Economic Review, 100(1), 70–97.
7. Broda, C., & Weinstein, D. E. (2010). Product creation and destruction: Evidence and price
implications. American Economic Review, 100(3), 691–723.
8. Brynjolfsson, E., Hu, Y., & Smith, M. D. (2010). The longer tail: The changing shape of
Amazon’s sales distribution curve. Manuscript, MIT Sloan School.
9. Brynjolfsson, E., Hu, Y., & Simester, D. (2011). Goodbye Pareto principle, hello long tail:
The effect of search costs on the concentration of product sales. Management Science, 57(8),
1373–1386.
10. Carvalho, V. M., & Gabaix, X. (2013). The great diversification and its undoing. American
Economic Review, 103(5), 1697–1727.
Product Innovation and Macroeconomic Dynamics 219

11. Chen, S. H., & Chie, B. T. (2005). A functional modularity approach to agent-based modeling
of the evolution of technology. In A. Namatame et al. (Eds.), The complex networks of
economic interactions: Essays in agent-based economics and econophysics (Vol. 567, pp. 165–
178). Heidelberg: Springer.
12. Chen, S. H., & Chie, B. T. (2007). Modularity, product innovation, and consumer satisfaction:
An agent-based approach. In International Conference on Intelligent Data Engineering and
Automated Learning.
13. Chiarli, T., Lorentz, A., Savona, M., & Valente, M. (2010). The effect of consumption and
production structure on growth and distribution: A micro to macro model. Metroeconomica,
61(1), 180–218.
14. Chiarli, T., Lorentz, A., Savona, M., & Valente, M. (2016). The effect of demand-driven struc-
tural transformations on growth and technical change. Journal of Evolutionary Economics,
26(1), 219–246.
15. Chie, B. T., & Chen, S. H. (2014). Non-price competition in a modular economy: An agent-
based computational model. Manuscript, U. Trento.
16. Chie, B. T., & Chen, S. H. (2014). Competition in a new industrial economy: Toward an agent-
based economic model of modularity. Administrative Sciences, 2014(4), 192–218.
17. Dawid, H., Gemkow, S., Harting, P., van der Hoog, S., & Neugart, M. (2018). Agent-based
macroeconomic modeling and policy analysis: The Eurace@Unibi model. In S. H. Chen, M.
Kaboudan, & Ye-Rong Du (Eds.), Handbook of computational economics and finance. Oxford:
Oxford University Press.
18. Delli Gatti, D., Gaffeo, E., Gallegati, M., & Giulioni, G. (2008). Emergent macroeconomics:
An agent-based approach to business fluctuations. Berlin:Springer.
19. Dosi, G., Fagiolo, G., & Roventini, A. (2005). An evolutionary model of endogenous business
cycles. Computational Economics, 27(1), 3–34.
20. Fagiolo, G., & Roventini, A. (2017). Macroeconomic policy in DSGE and agent-based models
redux: New developments and challenges ahead. Journal of Artificial Societies and Social
Simulation, 20(1).
21. Foster, L., Haltiwanger, J. C., & Syverson, C. (2008). Reallocation, firm turnover, and
efficiency: Selection on productivity or profitability? American Economic Review, 98(1), 394–
425.
22. Gabaix, X. (2011). The granular origins of aggregate fluctuations. Econometrica, 79, 733–772.
23. Georges, C. (2011). A hedonic approach to product innovation for agent-based macroeconomic
modeling. Manuscript, Hamilton College.
24. Grossman, G. M., & Helpman, E. (1991). Innovation and growth in the global economy.
Cambridge, MA: The MIT Press.
25. Hidalgo, C. A., Klinger, B., Barabasi, A. L., & Hausmann, R. (2007). The product space
conditions the development of nations. Science, 317, 482–487.
26. Hommes, C., & Iori, G. (2015). Introduction to special issue crises and complexity. Journal of
Economic Dynamics and Control, 50, 1–4.
27. Horvath, M. (2000). Sectoral shocks and aggregate fluctuations. Journal of Monetary Eco-
nomics, 45, 69–106.
28. Hottman, C., Redding, S. J., & Weinstein, D. E. (2014). What is ‘firm heterogeneity’ in trade
models? The role of quality, scope, markups, and cost. Manuscript, Columbia U.
29. Jackson, L. F. (1984). Hierarchic demand and the engel curve for variety. Review of Economics
and Statistics, 66, 8–15.
30. Klette, J., & Kortum, S. (2004). Innovating firms and aggregate innovation. Journal of Political
Economy, 112(5), 986–1018.
31. Lancaster, K. J. (1966). A new approach to consumer theory. Journal of Political Economy,
74(2), 132–157.
32. Lancaster, K. J. (1966). Change and innovation in the technology of consumption. American
Economic Review, 56(1/2), 14–23.
220 C. Georges

33. Marengo, L., & Valente, M. (2010). Industrial dynamics in complex product spaces: An
evolutionary model. Structural Change and Economic Dynamics, 21(1), 5–16.
34. Pintea, M., & Thompson, P. (2007). Technological complexity and economic growth. Review
of Economic Dynamics, 10, 276–293.
35. Strotz, R. H. (1957). The empirical implications of a utility tree. Econometrica, 25(2), 269–
280.
36. Sutton, J. (2007) Market share dynamics and the “Persistance of leadership debate. American
Ecmomic Review, 97(1), 222–241.
37. Valente, M. (2012). Evolutionary demand: A model for boundedly rational consumers. Journal
of Evolutionary Economics, 22, 1029–1080.
38. Windrum, P., & Birchenhall, C. (1998). Is product life cycle theory a special case? Dominant
designs and the emergence of market niches through coevolutionary-learning. Structural
Change and Economic Dynamics, 9(1), 109–134.
 Part II
New Methodologies and Technologies
Measuring Market Integration: US Stock
and REIT Markets

Douglas W. Blackburn and N. K. Chidambaran

Abstract Tests of financial market integration traditionally test for equality of risk
premia for a common set of assumed risk factors. Failure to reject the equality of
market risk premia can arise because the markets do not share a common factor, i.e.
the underlying factor model assumptions are incorrect. In this paper we propose
a new methodology that solves this joint hypothesis problem. The first step in
our approach tests for the presence of common factors using canonical correlation
analysis. The second step of our approach subsequently tests for the equality of
risk premia using Generalized Method of Moments (GMM). We illustrate our
methodology by examining market integration of US Real Estate Investment Trust
(REIT) and Stock markets over the period from 1985 to 2013. We find strong
evidence that REIT and stock market integration varies through time. In the earlier
part of our data period, the markets do not share a common factor, consistent with
markets not being integrated. We also show that during this period, the GMM tests
fail to reject the equality of risk premia, highlighting the joint hypothesis problem.
The markets in the latter half of our sample show evidence of both a common
factor and the equality of risk premia suggesting that REIT and Stock markets are
integrated in more recent times.

Keywords Market integration · Canonical correlation · Financial markets ·

REIT · Stock · GMM · Risk premia · Common factor · Factor model ·
Principal components

D. W. Blackburn · N. K. Chidambaran ()

Fordham University, New York, NY, USA
e-mail: blackburn@fordham.edu; chidambaran@fordham.edu

© Springer Nature Switzerland AG 2018 223

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_12
224 D. W. Blackburn and N. K. Chidambaran

1 Introduction

Financial market integration has significant impact on the cost of capital, the pricing
of risk, and cross-market risk sharing. Understanding whether markets are or are
not integrated is of great interest and importance to academics, practitioners, and
policymakers. Though definitions vary, two markets are said to be integrated if
two assets from two different markets with identical exposure to the same risk
factors yield the same expected returns (see, e.g., [4, 10, 24]). That is, financial
market integration requires that the same risk factors explain prices of assets in
both markets and that the risk-premia associated with the factors are equal across
the two markets. Empirical tests for determining whether markets have equal risk
premia for shared risk factors are econometrically challenging. First, the tests suffer
from a joint hypothesis problem as they require the right factor model specification
with common factors correctly identified. Second, financial market integration is a
changing and volatile process introducing noise in the data. In this paper, we present
a new approach to testing for market integration that untangles these issues and we
illustrate the methodology by examining integration of US Real Estate Investment
Trust (REIT)1 and stock markets.
Our approach for testing financial market integration is based on the Canonical
Correlation (CC) technique and is in two steps. In the first step, we determine
the factor model describing returns in each market. We begin with determining a
set of principal components that explain variation of returns in each dataset and
search for the presence of a common factor within the factor space of each market
using canonical correlations, as in [2].2 In the absence of a common factor, tests
of market integration are meaningless and this first step therefore represents a
necessary condition for integration tests. If a common factor is present, we then
proceed to the second step and develop economic proxies for the common factor
and run tests for equality of risk premia with respect to the economic factors. For
each step of our procedure, we develop tests for statistical significance and carefully
examine the implications for market integration.

1 REITs are companies that invest in real estate. REITs are modeled after closed-end mutual funds
with REIT shareholders able to trade their shares as do mutual fund shareholders. In addition to
being able to sell and divest their shares, REIT shareholders receive dividend payouts representing
income generated by the real estate properties owned by REITs.
2 The Canonical Correlation approach determines two factors, one from each data set, that have the

highest correlation compared to any other two factors across the two data sets. In determining
the set of factors with the highest correlation, the methodology recognizes that the principal
components are not a unique representation of the factor model and that a rotation of the principal
components is also a valid representation of the factor model. A rotation of the principal component
is a set of orthogonal linear combination of the principal components and are called canonical
variates. The search for a set of factors that has the highest correlation also includes all canonical
variates, i.e. all possible rotations of the principal components. A canonical variate in the first data
set that is highly correlated with a canonical variate in the second data set represents a common
factor, if the correlation is statistically significant.
Measuring Market Integration: US Stock and REIT Markets 225

Prior integration studies tend to fall into one of two camps. Either equality of risk
premia is tested relative to an assumed set of factors or implications of integration
are tested while ignoring the requirement that risk-premia be equal for shared
risk factors. In regard to the first, researchers assume a particular factor model of
returns and examine whether estimated risk premia of assumed common factors
are equal across markets using econometric techniques such as GMM or the [8]
methodology. Liu et al. [15] assume the Capital Asset Pricing Model (CAPM)
and Ling and Naranjo [13] use a multi-factor model in their study of equity and
REIT market integration. This is a direct test of market integration as it explicitly
determines the consistency of pricing across two markets relative to an assumed
pricing model. Bekeart and Harvey [4] assume a nested world and local CAPM
model in their investigation of world market integration. While these tests work
very well if the model and the assumptions are valid, results are subject to a joint
hypothesis problem. A rejection of integration may be a result of the markets not
being integrated or a result of using the incorrect model. It is also possible for tests
to not reject integration even when markets do not share a common factor in the
first place. By separately testing for the presence of a common factor, our approach
resolves the dual hypothesis issue.
The second strand of research studies market integration by examining criteria
other than equality of risk premia. Such tests are indirect measures of market
integration as they do not explicitly compare the pricing of assets across markets.
Two common measures, for example, are R-square (see [20]) and comovement
(see [3]). Studies based on these statistics seek to imply trends in integration
by measuring changes in the explanatory power of the assumed common factors
or changes in the economic implications of integration such as the degree of
diversification benefits, respectively. However, trends in R-Square and comovement
only suggest an increase in the pervasiveness of a set of common factors but it
is not clear what level of R-Square or comovement is required to imply market
integration or the existence of a common factor. Without knowing whether markets
are or are not integrated, interpreting these measures is difficult. As our tests directly
determine whether markets are integrated in each year, our approach allows us to
better interpret trends in measures such as comovement and R-Square.
In addition to the joint hypothesis problem, another significant challenge in
testing for financial market integration arises from the fact that integration is not
static but is instead a changing and volatile process. Markets can move from being
segmented to integrated, or from being integrated to segmented, because of changes
in risk premia or changes in risk factors. This volatility introduces noise into the
analysis especially when using data spanning periods when markets are integrated
and periods when they are not. We test for equality of risk premia only for the time-
period when a common factor exists, thereby improving the signal-to-noise ratio and
improving the power of our statistical tests. The increased power makes it possible
for us to test over a shorter time period, which allows for economically meaningful
analysis that accounts for changes in the integration process and time varying risk
premia. A related important aspect of the changes in integration over time is that the
pattern may be non-monotonic. While the literature has mostly assumed, and looked
226 D. W. Blackburn and N. K. Chidambaran

for, monotonic trends, we do not pre-suppose a monotonic transition of markets

from being segmented to being integrated.
We illustrate our methodology and contrast it to the standard approach through
a careful study of US REIT and stock market integration over the period from
1985 to 2012. We begin with the standard approach by assuming a single factor
model with the Center for Research in Security Prices (CRSP) value-weighted
portfolio (MKT) as the common factor. GMM is used to estimate the risk premia of
both markets every quarter using daily data. Using this standard approach, we are
unable to reject equal risk premia over 1985–1992 and 2000–2012. It is only during
the middle period 1993–1999 that equality of risk premia is consistently rejected.
Further analysis however reveals problems with these results. Specifically, we show
that MKT is not even a REIT factor throughout the 1985–1992 time period. For
example, in 1991, the average R-square from regressing REIT returns on MKT is
only 1.7%. Additionally, the R-square from regressing MKT on the first eight REIT
principal components is a dismal 3.9% suggesting MKT is not in the REIT factor
space. This illustrates the surprising result that we may find support for integration
even when the assumed factor is not present in the factor models of one, or both,
markets.
To contrast the previous result, we first test for the presence of a common factor
prior to testing for integration.Using canonical correlation analysis on the factors for
REITs and Stocks confirms that there are no common factors for nearly all quarters
prior to 1998. From 1998 to 2005, the canonical correlations vary above and below
0.87, the critical level shown by simulations to be necessary for the presence of a
common factor. It is only after 2005 that the canonical correlations are consistently
higher than 0.87. Tests for unequal premia are only consistently meaningful after
2005 when it is known that the two markets share a common factor. Testing for
integration prior to 2005 is therefore meaningless suggesting that the inability to
reject equal premia is a spurious result. Using GMM, we confirm the equality of
risk premia each quarter from 2005 to 2012 and over the entire sub-sample from
2005 to 2012.
Our choice of the US REIT and stock markets is interesting for a number of
reasons. First, there is mixed evidence in the literature as to whether the markets
are integrated. On the one side, early studies by Schnare and Struyk [21], Miles
et al. [18], and Liu et al. [15] conclude that the two markets are not integrated
while on the other side, Liu and Mei [16], Gyourko and Keim [11], and Ling and
Naranjo [13] find evidence that the two markets are at least partially integrated.3
Ling and Naranjo [13] are unable to reject integration after 1990 fewer times than
before 1990 suggesting an increase in integration. Our results add to this debate.
Second, though the real estate market is very large, the number of REITs is small
making integration tests less robust. Ling and Naranjo [13], for example, use a
sample of five real estate portfolios and eleven equity portfolios and are unable to

3 Liu et al. [15] find that equity REITs but not commercial real estate securities in general are

integrated with the stock market.

Measuring Market Integration: US Stock and REIT Markets 227

reject a difference in risk premia even though the difference is economically large—
1.48% for real estate and 5.91% for equities. Third, REIT data is noisy. Titman and
Warga [25] discuss the lack of power in statistical tests brought about by the high
volatility of REITs. By first identifying regimes over which integration is feasible,
we decrease the adverse effects of noise brought about by these data issues. Last,
US Stock and REIT markets present a nice setting to study a new methodology
for testing financial market integration without the confounding effects arising from
differences in legal regimes, exchange rates, time zone differences, or accounting
standards that are present in the study of international stock and bond markets.
The paper proceeds as follows. Section 2 describes the US REIT and Stock return
data used in the study. Section 3 presents the traditional approach used to study
integration as a first pass. Section 4 presents our new two-step methodology and
Section 5 concludes.

2 Data and Relevant Literature

In this section we describe the data and the structure of our tests, informed by the
relevant literature in the area.
Financial market integration requires that returns on assets in the two markets be
explained by the same set of risk-factors, i.e. a set of common factors, and that assets
exposed to these risk factors have the same expected returns. Testing for market
integration therefore involves estimating the risk-premia for a set of common factors
and is econometrically challenging. Tests inherently involve a dual hypothesis—that
the econometrician has the right factor model and that the estimated risk-premia for
the assumed risk factors are equal across the two data sets. A rejection of integration
may be a result of the markets not being integrated or a result of using the incorrect
model. It is also possible for tests to not reject integration even when markets do not
share a common factor in the first place.
Two approaches have been used in prior studies to test for market integration. In
the first approach, researchers assume a particular factor model of returns motivated
by financial theory applicable to one single market such as the US or to non-
segmented markets. They then test whether risk-premia, estimated using GMM or
other statistical methods, of the assumed common factors are equal across markets.
Liu et al. [15], for example, assume the Capital Asset Pricing Model (CAPM) as
the factor model and Ling and Naranjo [13] use a multi-factor model in their study
of equity and REIT market integration. Bekeart and Harvey [4] assume a nested
world and local CAPM as the factor model in their investigation of world market
integration. The problem with this approach is that the tests for equality of risk
premia for an assumed common factor has little statistical power when the factor is
not a common factor. Failure to reject unequal risk-premia may simply be because
the assumed common factor is not common to the two markets. Similarly, tests that
fail to reject equality of risk-premia may also simply reflect an incorrect common
factor assumption.
228 D. W. Blackburn and N. K. Chidambaran

In the second approach, researchers examine trends in criteria other than equality
of risk premia. The underlying argument in this approach is that integrated markets
differ from segmented markets with respect to specific statistical measures. Two
measures that have been extensively used are R-Square, a measure representing
the explanatory power of the assumed common factors, and comovement of asset
returns, which impacts the degree of diversification benefits and is a measure of the
economic implications of integration. If the R-Square or comovement is sufficiently
high, one can conclude that the markets are integrated. The issue, however, is
that it is not clear what level of R-Square or comovement is required to imply
market integration or the existence of a common factor. Further, noise in the data
arising from autocorrelation or heteroskedasticity can make measurements difficult.
Researchers, therefore, examine changes in R-Square and comovement over time
and interpret the time trends in integration as implying a movement towards or away
from integration. Pukthuanthong and Roll [20] focus on the time trend of the degree
to which asset returns are explained by R-square and Bekaert et al. [3] examine
the trend in the comovement of returns. These tests can inherently only suggest an
increase in the pervasiveness of a set of common factors but cannot really address
whether expected returns of assets facing similar risks are equal across the two
markets. It is plausible for financial markets to be completely segmented yet each
market exhibiting an increasing R-square relative to an assumed common factor.
Consider, for example, an economy that is comprised of two perfectly segmented
markets with returns of each market exhibiting an upward R-square trend relative
to its own market-specific value weighted portfolio. A researcher who assumes the
common factor is the value-weighted portfolio of all securities from both markets
will find that both markets have an increasing R-square relative to this common
factor. This is because the returns of each market are increasingly explained by the
part of the common factor from the same market while being orthogonal to the part
of the common factor from the other market. Without knowing whether markets
are or are not integrated, interpreting these measures is difficult. Theoretical asset
pricing literature has also suggested that the risk premia and factor sensitivities vary
over time (see [12]), and the changing risk premia may result in markets moving
in and out of integration. Previous empirical studies have found that markets can
become more and less integrated through time (see, e.g., [4, 6]).4

4 Bekeart and Harvey [4] study international market integration over 1977 to 1992 using a
conditional regime-switching model that nests the extreme cases of market segmentation and
integration and find that most countries are neither completely integrated nor segmented but are
instead somewhere in-between fluctuating between periods of being more integrated and periods
of being more segmented. India and Zimbabwe are particularly interesting cases. Both countries
exhibit extreme and sudden regime-shifts—in 1985, India instantaneously moves from being
integrated to segmented, and likewise, Zimbabwe suddenly changes from being integrated to
segmented in 1986 and then just as suddenly switches back to being integrated in 1991. Carrieri
et al. [6] document similar results of a volatile integration process characterized by periods of
increasing and decreasing integration.
Measuring Market Integration: US Stock and REIT Markets 229

Our approach differs from these standard approaches by first testing for the
presence of a common factor and then checking for equality of risk-premia only
when a common factor is present. By directly determining whether markets are
integrated in each year, our approach allows us to better interpret trends in measures
such as comovement and R-Square.
We use daily return data in our tests. We measure the integration level of the
US REIT and Stock Markets each quarter from 1985 to 2013. This choice of annual
time steps stems from a number of considerations. On the one hand, since we want to
study the time variation in integration, we need to develop a sequence of integration
measures with each measurement estimated over a short time-step. This allows us
to observe variation in the factors and factor structures over time. On the other hand,
we use principal components to compute the factor structure and GMM to measure
risk-premium, techniques which require a longer data series for robust estimation
leading to stronger results. Our choice of using daily data over each quarter strikes
a reasonable balance yielding a time series of 116 quarters with a time series of 60
days used to compute each measurement.
Daily stock return (share codes 10 and 11) and REIT return (share codes 18
and 48) data for the sample period 1985–2013 are from CRSP dataset available
through Wharton Research Data Services (WRDS). To be included in the data set for
a particular quarter, stocks and REITs must have daily returns for all days during that
quarter. Depending on the year, the number of stocks (REITs) used in our analysis
ranges from a low of 3429 (55) to a high of 7060 (210). Additionally, we use the
Fama-French-Carhart factors also available through WRDS. These factors are the
excess return on the CRSP value-weighted portfolio (MKT) and the three well-
known factor mimicking portfolio SMB, HML, and UMD.
Our statistical tests are structured as follows. We begin with a simulation to
identify the level of canonical correlation expected when a common factor is
present. The simulation uses the actual stock and REIT returns and embeds a
common factor into the returns. Performing the simulation in this way ensures that
the returns, and more importantly the errors, maintain their interesting statistical
properties. We can say that the two markets share a common factor when the
canonical correlations rise above the simulated benchmark level. Once a regime
characterized by the existence of common factors is identified, factor mimicking
portfolios based on canonical correlations of the principal components are compared
with economically motivated factors that are believed to explain returns. We can thus
identify economically meaningful common factors and run tests for the equality of
risk premia to identify market integration.

3 Market Integration: A First Pass

We begin our investigation by examining integration using a pre-specified factor

model for assets returns. As in earlier studies, we assume a model of expected
returns and test whether the assumed risk factors are priced equally across the
230 D. W. Blackburn and N. K. Chidambaran

markets. Liu et al. [15], one of the first studies of stock and REIT market integration,
assume the CAPM and test for integration relative to six different proxies for the
market factor while [13] assume a multi-factor model that includes such factors
as MKT, a term structure premium, a default premium, and consumption growth.
Following these studies, we select the single factor CAPM and use MKT as
the proxy for the market factor for our first series of tests. MKT is a value-
weighted portfolio that includes both stocks and REITs providing theoretic as well
as economic support for its use.
We begin by dividing all stocks and all REITs into 15 stock portfolios and
15 REIT portfolios according to each securities beta coefficients relative to MKT.
Portfolio betas and factor risk premia are simultaneously estimated using a GMM
framework and standard errors are computed using [19] with three lags. Coefficients
estimated using GMM are equal to those found using Fama-MacBeth. By using
GMM, however, we avoid the error-in-variable problem and standard errors are
appropriately estimated as in [22] with the added adjustment for serially correlated
errors. This estimation is performed each year from 1985 to 2013 using daily return
data. Estimating risk premia at the annual frequency allows for some time variation
in betas and in risk premia. A Wald test is used to test the hypothesis of equal stock
and REIT risk premia.
Table 1 shows the results of these tests. The second and third columns in the
table list the risk premia estimated from REITs and their corresponding t-statistics.
The fourth and fifth columns are the premia and t-statistics estimated for the stock
market. The premia are listed as daily percent returns. The sixth column is the Wald
statistic. For 1992, for example, the daily risk premia for REITs is 0.028% and for
stocks is 0.092% which roughly correspond to 0.62% and 2.02% monthly premia
for REITs and stock, respectively. Though the premia are economically significant,
only the stock premium is statistically significant, most likely due to noisy data.
Despite the large difference in premia, we are unable to reject the hypothesis that
the premia are equal suggesting the possibility that the stock and REIT markets are
integrated. In contrast, in 1995, the REIT and stock markets have daily risk premia
of 0.030% and 0.111%, respectively, which correspond to monthly premia of 0.66%
and 2.44%. Again, only the stock premia is significant; however, we do reject the
hypothesis of equal risk premia. We conclude, therefore, that for 1995 either markets
are not integrated or the assumed model is incorrect.
Overall, we reject integration in only six of the 29 years in the sample leading
to the belief that the markets are, by in large, integrated. This conclusion is
questionable upon further investigation. We perform two tests to check whether the
initial modeling assumptions are reasonable. The first test is based on the R-square
measure proposed by Pukthuanthong and Roll [20] who argue that a reasonable
measure of market integration is the proportion of returns that can be explained by
a common set of factors. If the level is low, then returns are primarily influenced
by local sources of risk. We construct this measure by taking the average R-square
from regressing each REIT on MKT. The average R-square is listed in column 7
of Table 1 labeled REIT R2. With the exception of 1987, the year of a big market
crash, R-squares are consistently below 10% from 1985 to 2002. R-squares then
Measuring Market Integration: US Stock and REIT Markets 231

Table 1 Mean canonical correlations

Test for equal risk premia
REIT REIT Stock Stock
Premium t-stat Premium t-stat Wald REIT R2 PC R2
1985 0.041 0.56 0.090 1.89 1.06 0.009 0.068
1986 0.125 1.49 0.049 0.75 1.34 0.021 0.110
1987 −0.076 −0.61 0.012 0.09 0.41 0.111 0.618
1988 0.015 0.14 0.104 1.57 1.82 0.016 0.074
1989 0.008 0.08 0.070 1.35 1.43 0.014 0.064
1990 −0.105 −0.85 −0.041 −0.55 0.71 0.018 0.031
1991 0.242 2.27** 0.201 2.91*** 0.34 0.017 0.039
1992 0.028 0.37 0.092 2.13** 2.17 0.008 0.051
1993 0.178 2.01** 0.084 2.40** 7.12*** 0.011 0.046
1994 0.013 0.20 0.021 0.48 0.03 0.025 0.082
1995 0.030 0.47 0.111 3.09*** 5.07** 0.011 0.039
1996 0.255 3.72*** 0.087 1.48 8.23*** 0.028 0.399
1997 0.163 1.56 0.113 1.43 0.39 0.049 0.490
1998 −0.187 −1.18 0.052 0.48 4.95** 0.078 0.427
1999 −0.117 −0.65 0.243 2.86*** 17.97*** 0.012 0.131
2000 0.133 0.88 −0.038 −0.29 1.72 0.029 0.336
2001 0.163 1.01 0.072 0.67 0.73 0.074 0.476
2002 0.124 0.73 −0.054 −0.49 2.55 0.116 0.456
2003 0.231 2.71*** 0.231 3.00*** 0.00 0.137 0.581
2004 0.130 1.54 0.056 1.03 1.89 0.107 0.482
2005 0.003 0.06 0.018 0.41 0.11 0.213 0.646
2006 0.088 1.74* 0.043 0.95 1.01 0.187 0.696
2007 −0.087 −1.08 −0.019 −0.31 1.19 0.291 0.689
2008 −0.047 −0.29 −0.127 −0.81 0.26 0.384 0.796
2009 0.130 1.42 0.183 1.73* 0.25 0.425 0.826
2010 0.121 1.48 0.098 1.27 0.09 0.397 0.810
2011 0.011 0.13 −0.008 −0.09 0.05 0.509 0.874
2012 0.123 1.87* 0.063 1.12 1.15 0.244 0.677
2013 0.049 0.76 0.136 3.02*** 3.76* 0.235 0.561
For each year, the risk premium relative to the excess return on the CRSP Value-Weighted Portfolio
(MKT) is estimated from the set of all REITs and the set of all US common stocks. Premia are listed
as percent daily returns. A Wald test is used to test for equal premia. REIT R2 is the average R-
Square from regressing each REIT on MKT, and PC R2 is the R-square from regressing the MKT
on the first eight principal components estimated from REIT returns

begin to increase after 2002 and reach a maximum of 0.509 in 2011. With such
small R-squares early in the sample, relative to our choice of common factors, the
REIT market must be primarily influenced by local factors up until 2002.
The second test determines whether MKT lies in the REIT factor space. We
follow Bai and Ng [2] who argue that a proposed economic factor should lie in the
232 D. W. Blackburn and N. K. Chidambaran

factor space spanned by the set of principal components.5 For each year, we select
the first eight principal components to represent the REIT factor space. The choice
of eight principal components is determined using the information criteria of [1].
MKT is regressed on the eight principal components and the R-square is reported
in column 8 of Table 1, PC R2. In their study, Bai and Ng [2] state that when an
economic factor lies in the factor space, we should observe a canonical correlation
(square root of R-square in this case) near one. We find many R-squares below 10%
in the early part of the sample. R-square does eventually reach 70% in year 2007.
During the early part of the sample, it is clear that MKT does not lie in the REIT
market’s factor space.
These tests show that despite failing to reject equal premia, there is absolutely
no evidence that MKT is a REIT factor for the vast majority of the sample period,
especially prior to 2002. Hence, the GMM tests are misspecified in the early part of
the sample and the failure to reject integration is meaningless.
Our analysis above suggests that the critical issue in market integration is in
determining the appropriate model for stock and REIT returns and there are several
questions to ask with respect to the factor model. Do the two markets even share
a common factor or are the two markets completely segmented? If the markets do
share a common factor, what are the reasonable proxies for the common factors?
Our proposed technique, presented in the next section, addresses these questions
prior to estimating and testing risk premia.

4 Two-Step Approach to Testing Integration

The first step in our two-step testing approach is to determine the factor model
for returns in each of the markets. As is standard in the investigation of multiple
markets, we allow returns of stocks and REITs to follow different linear factor
models. Returns may depend on either factors common to both markets—common
factors—or factors that are unique to a particular market—local factors.
For the i = 1, 2, . . . , M stocks and j = 1, 2, . . . , N REITs, the return generating
functions for stocks and REITs are given by
g h
Stock : RitS = Et−1 [RitS ] + l=1 βil Flt +
S C
k=1 γik Fkt + it
S S S
g l (1)
REIT : RjRt = Et−1 [RjRt ] + l=1 βj l Flt +
R C
k=1 γj k Fkt + j t .
R R R

where g, h, and l are the number of common factors, local stock factors, and local
REIT factors, respectively. For stocks (REITs), RitS (RjRt ) are the asset returns,
E[RitS ] (E[RjRt ]) are the expected returns, βilS (βjRl ) are the sensitivities to the

5 See also [23].

Measuring Market Integration: US Stock and REIT Markets 233

common factors FltC , and γikS (γjRk ) are the factor sensitivities to the local factors
FktS (FktR ). We assume that all local and common factors are mutually orthogonal.
Expected returns are given by
g h
Stock : Et−1 [RitS ] = λS0,t + l=1 βil λlt +
S C
k=1 γik λkt + it
S S S
g l (2)
REIT : Et−1 [RjRt ] = λR 0,t + l=1 βj l λlt +
R C
k=1 γj k λkt + j t .
R R R

where λS0,t (λR0,t ) is the riskless rate of return at time t for the stock (REIT) market.
λlt are the risk premia for the g common sources of risk, and λSkt (λR
C
kt ) are the risk
premia for the h (l) local sources of risk. In many studies of market integration it
must be tested whether the riskless rates are equal across markets. For this study,
since REITs and stocks both trade in the USA, it is reasonable to assume the same
riskless rate applies to both markets, λS0,t = λR 0,t .
Perfectly integrated markets require that the same expected return model applies
to securities in all markets. Expected returns are determined only by common
factors and not local factors, and the risk premia on the common factors must be
equal across markets. The presence of local risk factors, therefore, is sufficient
to show markets are not perfectly integrated. The literature does discuss partial
integration. For example, [7] have an equilibrium model that allows for mild or
partial segmentation, where both common and local factors are present.
We next turn to determining the number of common factors g and proxies for the
common factors. As noted before, once the correct model and appropriate proxies
for the common factors are identified, we can test for equal premia using GMM.

4.1 Existence of Common Factors

The approach we use to identify the appropriate model for each market builds
on [1, 2]. It is well known that the factor space for a particular market can
be consistently constructed using principal components [23]. Therefore, we first
determine the factor structure and factor proxies for each market separately using [1]
to find the number of principal components required to explain the returns of assets
in each market—g + h for equities and g + l for REITs. Canonical correlation
analysis is used to compare the factor spaces spanned by each market’s principal
components in order to identify the number common factors g. Canonical variates
associated with sufficiently large canonical correlations are related to proposed
economic factors to identify economically meaningful sources of risk. A test for
equal risk premia is conducted over periods when g > 0 using the identified factor
proxies.6

6 Studies
that have used a similar approach are [9] in their comparison NYSE and Nasdaq stocks,
and Blackburn and Chidambaran [5] who study international return comovement derived from
common factors.
234 D. W. Blackburn and N. K. Chidambaran

We first identify the number of factors (principal components) required to explain

the returns in each market individually. Using [1] on daily data over each quarterly
we find that the number of factors varies from quarter to quarter.7 For stocks, the
number of factors ranges from one to four while for REITs, the number of factors
ranges from five to eight. In every quarter, the number of REIT factors is strictly
larger than the number of stock factors implying l < h.8 At this point it is clear
that the two markets cannot be perfectly integrated since REITs always have h > 0
local factors. Nonetheless, it is still interesting to investigate further to determine
the extent to which the two markets are partially integrated.
To simplify the analysis going forward, we hold the number of principal
components constant for each market. We use eight principal components for
REITs and four for stocks. These choices have support in the literature. Liow and
Webb [14] use seven principal components to describe REIT returns—equal to the
number of eigenvalues greater than one, and the four-factor Fama-French-Carhart
model is commonly used for equities in the asset pricing literature. In addition to
using the first four principal components to proxy for stock factors, we also use the
four Fama-French-Carhart factors.9 Since the use of the four-factor equity model
has strong support in the literature, it is reasonable to assume that these factors are
prime candidates for being common factors.
While it is the case that there are many quarters when the number of factors
is less than four for stocks and eight for REITs, the use of too many principal
components does not present a problem. What is essential is that the factor space
be completely spanned by the set of principal components. Additional principal
components simply add noise. For the cases when the true number of factors is less
than what we assume, the true factor space will be a subspace of the one empirically
estimated.
Canonical correlation analysis is used to determine the number of common
factors g from the set of g + h stock latent factors and the g + l REIT latent
factors by finding the linear combination of the stock factors that has the maximum
correlation with a linear combination of REIT factors [2]. The pair of factors
with maximum correlation are called first canonical variates and the correlation
between them is the first canonical correlation. The process is then repeated to
determine a pair of second canonical variates that are characterized as having
the maximum correlation conditional on being orthogonal to the first canonical
variates. The correlation between the pair of second canonical variates is called the

7 Thisanalysis was also performed on an annual basis. Results are similar.

8 It
is common to use variables other than the Fama-French factors in studying REIT returns. For
example, [17] use the excess returns on the value-weighted market portfolio, the difference between
the 1-month T-bill rate and inflation, the 1-month T-bill rate relative to its past 12-month moving
average, and the dividend yield (on an equally weighted market portfolio).
9 While it would be ideal to also include a standard set of economically motivated REIT factors,

unlike the literature on explaining the cross-section of stock returns, the REIT literature has not yet
agreed on such standard set of factors.
Measuring Market Integration: US Stock and REIT Markets 235

second canonical correlation, and so on. Essentially, canonical correlation analysis

determines whether and to what degree the two factor spaces overlap.
Other researchers have also used a principal component approach to identify
factor models while studying integration. Instead of estimating principal compo-
nents from returns in each market separately as we do here, the common approach
is to combine returns of assets from all markets into one large data set and then
calculate the principal components of the combined set of returns (see, e.g., [3, 20]).
This traditional approach presents several problems that can lead to incorrectly
identifying a strong local factor in one market as a global factor. First, when markets
do not share any common factors the first set of principal components will be
comprised of some rotation of local factors. Second, suppose, as is the case here,
that the two markets are of unequal size, with market one having a large number of
assets while market two having only few securities. If market one has a particularly
strong local factor relative to a weak common factor, then principal components will
identify those factors that explain the greatest proportion of return variation—the
strong local factor. In both cases, the first set of principal components that represent
local factors are assumed to be proxies for common factors. Our approach avoids
these pitfalls by first identifying the factor structure for each market separately and
then comparing the two factor spaces.

4.1.1 Benchmark Canonical Correlation Level

An important issue in canonical correlation analysis not rigorously addressed in the

literature is the magnitude of the canonical correlation required to claim existence
of one or more common factor. Previous research has tended to arbitrarily select
a benchmark value [2, 9]. Since our approach requires making strong statements
about the existence of common factors, we need to be more precise. To do so, we
perform a simulation to determine the appropriate benchmark canonical correlation
level for the stock and REIT data.
From the estimated four and eight principal components for stocks and REITs,
respectively, we follow the canonical correlation methodology to calculate the first
four canonical variates for both stocks and REITs. We begin with one common
factor. The following steps are then repeated for two, three, and four common
factors. For one common factor, each stock return is regressed on the first stock
canonical variate and each REIT return is regressed on the first REIT canonical
variate to determine their respective alphas, betas, and errors:

Stock : RitS = αiS + βiS VtS + itS

(3)
REIT : RjRt = αjR + βjR VtR + jRt .

where βiS is the loading of stock i on the first stock canonical variate VtS and
βjR is the factor loading on the first REIT canonical variate VtR . The errors are
236 D. W. Blackburn and N. K. Chidambaran

represented by itS for stocks and jRt for REITs. We do not require any particular
level of canonical correlation between canonical variates in this step.
We now create simulated returns with the feature of having a common factor. We
do this by replacing the canonical variates VtS and VtR in Eq. (3) with a proxy for
the common factor FtC :

Stock : R̂itS = αiS + βiS FtC + itS

. (4)
REIT : R̂jRt = αjR + βjR FtC + jRt .

The alphas, betas, and errors remain the same as in Eq. (3) only the factor is changed.
Constructing the simulated returns in this way preserves many of the statistical
features of the true stock returns such as heteroskedasticity, serially correlated
errors, and cross-correlated errors. However, the simulated returns now share a
common factor. The proxy used for the common factor is derived from the sum
of the stock and REIT canonical variates:

VtS + VtR
FtC = " (5)
V ar(VtS ) + V ar(VtR ) + 2Cov(VtS , VtR )

where the denominator normalizes the factor FtC to have unit variance.10 This proxy
for the common factor maintains many of the desirable statistical features observed
in the canonical variates.
Using the simulated returns R̂itS and R̂jRt , principal components are estimated
and canonical correlation analysis is used to determine the simulated canonical
correlations when common factors exist and the number of common factors is
known. Since we have embedded common factors into the simulated returns, we
should expect canonical correlations much closer to one; however, estimation error
in calculating principal components caused by finite samples, heteroskedasticity,
and autocorrelation may cause the first canonical correlation to be large, but less
than one. We perform this simulation for each quarter of our sample period using
the same set of returns—same time-series length and cross-section—as used in the
study. The benchmark levels are unique to this particular study.
Results are plotted in Fig. 1. Figure 1a is the plot of the first four canonical
correlations when there is only one common factor. The bold, solid line is the first
canonical correlation, and as expected, is close to one with a time-series average, as
shown in Table 2-Panel A, of 0.954. The standard error, 0.045, is listed below the
mean. All other canonical correlations are much smaller with time-series averages
of 0.521 for the second canonical correlation (light, solid line), 0.367 for the third
canonical correlation (dashed line), and 0.215 for the fourth canonical correlation

10 We also used the stock canonical variates and REIT canonical variates as common factor proxies

and the results did not change.

Measuring Market Integration: US Stock and REIT Markets 237

a b
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
1985-1
1986-1
1987-1
1988-1
1989-1
1990-1
1991-1
1992-1
1993-1
1994-1
1995-1
1996-1
1997-1
1998-1
1999-1
2000-1
2001-1
2002-1
2003-1
2004-1
2005-1
2006-1
2007-1
2008-1
2009-1
2010-1
2011-1
2012-1
2013-1

1985-1
1986-1
1987-1
1988-1
1989-1
1990-1
1991-1
1992-1
1993-1
1994-1
1995-1
1996-1
1997-1
1998-1
1999-1
2000-1
2001-1
2002-1
2003-1
2004-1
2005-1
2006-1
2007-1
2008-1
2009-1
2010-1
2011-1
2012-1
2013-1
c d
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
1985-1
1986-1
1987-1
1988-1
1989-1
1990-1
1991-1
1992-1
1993-1
1994-1
1995-1
1996-1
1997-1
1998-1
1999-1
2000-1
2001-1
2002-1
2003-1
2004-1
2005-1
2006-1
2007-1
2008-1
2009-1
2010-1
2011-1
2012-1
2013-1

1985-1
1986-1
1987-1
1988-1
1989-1
1990-1
1991-1
1992-1
1993-1
1994-1
1995-1
1996-1
1997-1
1998-1
1999-1
2000-1
2001-1
2002-1
2003-1
2004-1
2005-1
2006-1
2007-1
2008-1
2009-1
2010-1
2011-1
2012-1
2013-1
Fig. 1 Simulated canonical correlations. The first four canonical correlations between REIT and
stock principal components are estimated each quarter from 1985 to 2013 using simulated returns.
The simulated returns are modeled to include one, two, three, and four factors common to both
markets. The bold solid line is the first canonical correlation, the thin solid line is the second
canonical correlation, the dashed line is the third, and the dotted line is the fourth canonical
correlation

(dotted line). This is exactly what is expected. With only one common factor, it
is expected that we find one canonical correlation close to one with the remaining
much lower.
In Fig. 1b, we plot the time-series of canonical correlation when we simulate
returns with two common factors. We notice two different affects. First, the second
canonical correlation, with a time-series average of 0.896, is much larger than in
the one common factor case when its average was 0.521. Second, the average first
canonical correlation is slightly larger and has noticeably less variance. Meanwhile,
the time-series average for the third and fourth canonical correlations is nearly
unchanged. The presence of two common factors increases the signal-to-noise ratio
in the system thus making it easier to identify common factors.
Figure 1c for three common factors and Fig. 1d for four common factors tell
similar stories. Common factors are associated with high canonical correlations.
238 D. W. Blackburn and N. K. Chidambaran

Table 2 Mean canonical correlations

Panel A: average canonical correlations from simulated returns
CC 1 CC 2 CC 3 CC 4
One common factor 0.954 0.521 0.367 0.215
0.045 0.135 0.114 0.094
Two common factors 0.967 0.896 0.368 0.216
0.032 0.037 0.114 0.091
Three common factors 0.972 0.928 0.875 0.216
0.027 0.026 0.034 0.091
Four common factors 0.977 0.941 0.909 0.864
0.022 0.024 0.025 0.035
Panel B: average canonical correlations from actual returns
Eight REIT PCs vs Four stock PCs 0.752 0.514 0.371 0.216
Eight REIT PCs vs FFC four-factors 0.731 0.452 0.316 0.199
We use canonical correlation analysis [2] to measure the presence of common factors between the
US stock and REIT markets. In Panel A, we simulate return data to have one to four common
factors to determine a canonical correlation level expected when common factors exist. The
standard error is listed below each mean canonical correlation. In Panel B, we calculate the time-
series mean canonical correlation from actual stock and REIT returns each quarter from 1985
to 2013. For REITs, we use the first eight principal components to represent REIT factors. Two
different proxies for stock factors are used—the first four principal components (PC) and the four
Fama-French-Carhart factors (FFC)

From Table 2-Panel A, we see the dramatic jump in average canonical correlation
from 0.216 to 0.864 when moving from three to four common factors. These results
clearly demonstrate that for our particular system, common factors are associated
with high canonical correlations. The benchmarks we use in this paper to claim
statistical evidence for the presence of one or more common factors is the 95% lower
confidence level below the time series canonical correlation mean. The benchmark
threshold levels for the first, second, third, and fourth canonical correlations, used
in this paper are 0.87, 0.82, 0.81, and 0.79, respectively.

4.1.2 Evidence for a Common Factor

The canonical correlation method is now used on the real stock and REIT
return data to find regimes when a common factor exists. Results are plotted in
Fig. 2a–d. Each figure shows two time-series of canonical correlations. The solid
line is the canonical correlation estimated from the comparison of eight REIT
principal components and four stock principal components. The dotted line is
estimated from eight REIT principal components and the four Fama-French-Carhart
factors. The two dotted horizontal lines represent the simulated time series mean
canonical correlations and its 95% lower confidence. We say that a common factor
exists when the canonical correlation rises above the 95% lower confidence interval
(the lower of the two horizontal dotted lines).
Measuring Market Integration: US Stock and REIT Markets 239

a b
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
1985
1987
1989
1991
1993
1995
1997
1999
2001
2003
2005
2007
2009
2011
2013

1985
1987
1989
1991
1993
1995
1997
1999
2001
2003
2005
2007
2009
2011
2013
c d
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
1985
1987
1989
1991
1993
1995
1997
1999
2001
2003
2005
2007
2009
2011
2013
1985
1987
1989
1991
1993
1995
1997
1999
2001
2003
2005
2007
2009
2011
2013

Fig. 2 Canonical correlations. The first four canonical correlations between the principal compo-
nents of daily REIT returns and proxies for stock factors are plotted for each quarter from 1985 to
2013 [2]. Two proxies for stock factors are considered—principal components from stock returns
(given by the solid line), and the three Fama-French plus Carhart factors (given by the dotted line).
The first canonical correlations is given in (a), second canonical correlations in (b), third in (c),
and the fourth in (d). The two dotted horizontal lines represent the simulated canonical correlation
time series mean and its 95% lower confidence interval

Figure 2a depicts the quarterly first canonical correlations values. We first notice
that the choice of stock factors, whether the first four principal components or the
four Fama-French-Carhart factors, does not greatly affect our results. The solid and
dotted curves track each other closely. Second, the time series average canonical
correlation is lower than the 95% lower confidence benchmark of 0.87. Table 2-
Panel B lists a time-series average first canonical correlation of 0.752; however,
there is substantial time variation in the canonical correlation suggesting significant
variation in integration between the two markets. The first canonical correlation
initially increases from 0.6 to 0.97 in the fourth quarter of 1987, concurrent with the
1987 market crash. It then trends downward until 1993 when it begins a long trend
upward until 2012 before it drops down again.
The time series of canonical correlations describes the integration process from
markets having completely separate return models with only local factors to models
comprised of both local and common factors. Using 0.87 as the threshold required
for a common factor, we observe the process as it moves toward and away from
the threshold. With the exception of the fourth quarter of 1987, the first canonical
correlation remains below 0.87 until 1998. After 2000, the canonical correlation
remains close to the benchmark level fluctuating above and below the 0.87 line.
Finally, in 2005 the canonical correlation breaches the threshold and for the most
240 D. W. Blackburn and N. K. Chidambaran

part remains above 0.87. The canonical correlation drops back down in the last
two quarters of 2013. From 2005 to 2012, the canonical correlation is above the
threshold 29 out of 32 quarters providing strong support for the existence of a single
factor over this time period.
Figure 2b–d shows plots of the second, third and fourth canonical correlations,
respectively. The second canonical correlation rises above its threshold level of
0.82 three times but each time it drops immediately back down. Spikes tend to
be associated with market crashes when returns of all assets tend to be highly
correlated. The third and fourth canonical correlations never rise above their
respective benchmarks.
Our evidence shows that only one common factor exists and only during the
latter part of the sample. Both choices of stock factors, the latent factors and the
Fama-French-Carhart factors, support this result. Conservatively, we can claim the
existence of a common factor between 2005 and 2012; but, more liberally, there
is some support for a common factor over the longer period from 2000 to 2013.
We find no evidence for any significant common factor prior to this thus calling
into question the research that found evidence for stock and REIT integration using
samples from the 1980s and 1990s.

4.1.3 Economic Identity of Common Factors

Having identified a regime over which a common factor exists, we relate the
single statistically determined latent factor with the Fama-French-Carhart factors
to identify an economically meaningful proxy for the common factor. While the
literature has posited a large number of possible risk factors, it has also consistently
agreed on the use of the four-factor model comprised of the MKT, SMB, HML, and
UMD. Therefore, it is reasonable to assume that since these factors perform well
in explaining stock returns that any common factors are likely to be selected from
this list. In the end, it is an empirical matter as to how well the latent factor can be
explained by these four factors.
For the statistical proxy for the common factor, we use F C from Eq. (5)
constructed each quarter from 1985 to 2013. We look over the entire time period
in order to contrast the regime when integration is believed to be possible with
the regime over which integration is believed to be impossible. Figure 3a plots the
R-square from regressing F C on MKT each quarter and Fig. 3b is the R-square
from regressing F C on all four Fama-French-Carhart factors. In both plots, we
find a very noisy and inconsistent relationship between the statistically determined
common factor and the economic factors. It is not until 2005 that the R-squares
are consistently high. This is particularly true for the R-squares in Fig. 3b where
R-square is on average 0.93 over 2005–2013. For the results using MKT, R-square
is 0.82 on average over the same time period. The market factor, MKT, is clearly
the primary driver of the high R-square in Fig. 3b. In unreported results, we show
that SMB accounts for much of the remaining difference in R-square while HML
and UMD contribute very little. As previously mentioned, the value-weighted MKT
Measuring Market Integration: US Stock and REIT Markets 241

a b
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
1985 1989 1993 1997 2001 2005 2009 2013 1985 1989 1993 1997 2001 2005 2009 2013

Fig. 3 Economic factors. We relate the statistically determined common factor shared by both
the REIT and equity markets to four commonly used factors—the excess CRSP value-weighted
portfolio, the Fama-French hedge portfolios SMB and HML, and Carhart’s UMD portfolio. The
time-series of the R-squares from regressing the statistical common factor on MKT (a) and on all
four factors (b) is plotted. Regressions are performed each quarter using daily data from 1985 to
2013

portfolio comprised of stocks and REITs, among other assets, has strong economic
and theoretical support. We feel that we are on both strong statistical and economic
footing by selecting MKT as our proxy for the common factor towards the end of
our sample period.

4.1.4 Test for Equal Risk Premia

As our final step, we test for equal risk premia over the identified regime using
the previously described GMM approach. In Table 3-Panel A, we list the estimated
risk premia for both the stock and REIT markets and the result of the Wald test
for the more conservative regime 2005–2012 and the more liberal regime 2000–
2013. In both cases, we are unable to reject differences in risk premia. Moreover,
the daily percent risk premia are indeed very similar. For the 2005–2012 regime,
the daily REIT and stock premia are 0.032% and 0.037%, respectively, equivalent
to 0.72% and 0.81% per month. For the 2000–2013 regime, the daily premia are
0.057% and 0.058%, equivalent to 1.24% and 1.27% monthly, for REITs and stock,
respectively. The highly similar premia provide strong evidence supporting our
proposed approach to studying integration.
To contrast these results, we take one last view over the entire sample period
1985–2013. In Panel B of Table 3, risk premia are estimated over 5-year rolling
windows. An interesting result emerges. Rejection of equal premia appears cyclical.
This is clearly seen in Fig. 4 where we plot the daily risk premia from Table 3-Panel
B. The solid line represents the REIT premia and the dotted line represents stock
premia. The stock risk premia (dotted line) varies through time but the variation is
rather stable—no large jumps or swings. The REIT premia, on the other hand, are
highly cyclical and characterized by large swings. The premia is at its minimum
in 1989–1993 but is at its maximum over the period 1993–1997. The grey shaded
regions in Fig. 4 indicate the time periods when statistical tests reject equal premia.
242 D. W. Blackburn and N. K. Chidambaran

Table 3 Mean canonical correlations

Panel A: test for equal risk premia
REIT REIT Stock Stock
Premium t-stat Premium t-stat Wald P-value
2005–2012 0.033 1.12 0.037 1.17 0.02 0.89
2000–2013 0.057 2.37** 0.058 2.28** 0.00 0.96
Panel B: test for equal risk premia: 5-year rolling window
1985–1989 0.026 0.75 0.066 1.58 0.91 0.340
1986–1990 −0.012 −0.32 0.033 0.75 1.03 0.311
1987–1991 0.003 0.07 0.090 1.90* 3.34 0.068*
1988–1992 −0.013 −0.29 0.153 3.92*** 18.00 0.001***
1989–1993 −0.041 −0.72 0.126 3.68*** 23.80 0.001***
1990–1994 0.094 2.25** 0.148 4.47*** 2.65 0.103
1991–1995 0.160 5.53*** 0.162 5.74*** 0.01 0.928
1992–1996 0.180 6.06*** 0.159 6.69*** 0.73 0.394
1993–1997 0.214 5.80*** 0.137 4.77*** 7.10 0.008***
1994–1998 0.079 1.95* 0.109 3.16*** 0.74 0.390
1995–1999 0.074 1.66* 0.146 3.61*** 3.19 0.074*
1996–2000 0.057 1.12 0.110 2.17** 1.09 0.295
1997–2001 0.029 0.53 0.109 1.97** 2.09 0.149
1998–2002 0.028 0.48 0.071 1.21 0.55 0.457
1999–2003 0.134 2.67*** 0.114 2.06** 0.12 0.727
2000–2004 0.153 3.07*** 0.067 1.29 2.81 0.094*
2001–2005 0.104 2.48** 0.074 1.78* 0.52 0.472
2002–2006 0.093 2.58*** 0.068 1.97** 0.53 0.465
2003–2007 0.038 1.31 0.068 2.53** 1.25 0.263
2004–2008 −0.004 −0.10 −0.002 −0.05 0.00 0.960
2005–2009 0.015 0.38 0.028 0.65 0.09 0.764
2006–2010 0.032 0.78 0.044 0.99 0.08 0.784
2007–2011 0.024 0.56 0.031 0.66 0.02 0.886
2008–2012 0.052 1.22 0.048 1.03 0.01 0.935
2009–2013 0.075 2.36** 0.096 2.71*** 0.33 0.563
Risk premia relative to the excess return on the CRSP Value-Weighted Portfolio (MKT) is
estimated from the set of all REITs and the set of all US common stock using GMM. Premia
are listed as percent daily returns. A Wald test is used to test for equal premia. T-statistics testing
the hypothesis of a significant risk premium is provided next to each premium. Panel A measures
the risk premia over the regime determined from the canonical correlation methodology. Panel B
estimates risk premia using a 5-year rolling window over the entire sample period

Rejection is nearly always associated with the peaks and troughs of the REIT
premia. As the REIT premia move through the cycle and cross over the stock premia,
statistical tests are unable to reject equal premia. This is not because markets are
integrated, but because of the time-varying, cyclical nature of the REIT premia.
Measuring Market Integration: US Stock and REIT Markets 243

0.0025

0.002

0.0015

0.001

0.0005

-0.0005
1985-1989
1986-1990
1987-1991
1988-1992
1989-1993
1990-1994
1991-1995
1992-1996
1993-1997
1994-1998
1995-1999
1996-2000
1997-2001
1998-2002

2000-2004
1999-2003

2001-2005
2002-2006
2003-2007
2004-2008
2005-2009
2006-2010
2007-2011
2008-2012
2009-2013
Fig. 4 Stock and REIT risk premia by year. Risk premia relative to the CRSP value-weighted
market portfolio for REITs (solid line) and stocks (dotted line) are estimated over 5-year rolling
windows using GMM. Premia are estimated using daily data and are plotted as daily premia. The
grey shaded regions represent the intervals when equality of risk premia is rejected

As additional support for our approach, Table 3-Panel B and Fig. 4 clearly show
the two premia converging in the later part of the sample, beginning around 2003.
After this point, the stock and REIT risk premia are nearly equal, tracking each other
closely as the rolling window advances through time. This is the precise regime
the canonical correlation methodology identifies as having a common factor—the
regime over which integration of the two markets is a possibility.

4.2 Measures of Integration

Due to the complications involved in testing for integration, a number of papers

instead indirectly test integration by examining its implications, such as changes
in diversification benefits. In this section, we discuss these alternative tests and
relate them to time-varying canonical correlations underlying our procedure. As
in the case of the results of GMM tests, the results of these alternate tests are
also difficult to interpret. Specifically, while such measures are very informative
for understanding the changing opportunities and investment environment, they are
difficult to interpret as measures of integration.
244 D. W. Blackburn and N. K. Chidambaran

a b
0.7 0.5
0.45
0.6
0.4
0.5 0.35
0.3
0.4
0.25
0.3
0.2

0.2 0.15
0.1
0.1
0.05
0 0
1985-1
1986-1
1987-1
1988-1
1989-1
1990-1
1991-1
1992-1
1993-1
1994-1
1995-1
1996-1
1997-1
1998-1
1999-1
2000-1
2001-1
2002-1
2003-1
2004-1
2005-1
2006-1
2007-1
2008-1
2009-1
2010-1
2011-1
2012-1
2013-1

1985-1
1986-1
1987-1
1988-1
1989-1
1990-1
1991-1
1992-1
1993-1
1994-1
1995-1
1996-1
1997-1
1998-1
1999-1
2000-1
2001-1
2002-1
2003-1
2004-1
2005-1
2006-1
2007-1
2008-1
2009-1
2010-1
2011-1
2012-1
2013-1
Fig. 5 Measures of integration. We calculate the R-square measure of integration [20] and the
comovement measure of Bekaert et al. [3] using two different models—a single factor model
comprised of a factor derived from canonical correlation analysis (solid line) and a four-factor
model that includes the Fama-French-Carhart factors (dotted line). Measurements are taken using
daily REIT and stock return data over each quarter from 1985 to 2013

In a recent study, Pukthuanthong and Roll [20] argue that R-square is a “sensible
intuitive quantitative measure of financial market integration”, in the sense that if
R-square is small, “the country is dominated by local or regional influences. But if
a group of countries is highly susceptible to the same global influences, there is a
high degree of integration”. (p 214–215) We compute the R-square by averaging the
individual R-squares obtained from regressing each REIT on proxies for common
factors. A plot of the R-square measured each quarter using daily return data is
shown in Fig. 5a. The solid line uses F C from Eq. (5) as the proxy for the common
factor while the dotted line assumes the Fama-French-Carhart factors. Both curves
tell a similar story. The factors explain a very small proportion of REIT returns
throughout the late 1980s and 1990s. At the end of the 1990s and throughout the
2000s, both measurements begin to trend upward, explaining greater proportion of
the REIT returns. There is a decrease in R-square at the end of the sample.
Comovement of asset returns across markets is another commonly used measure.
Bekaert et al. [3] state, “If the increase in covariance is due to increased exposure
to the world markets, as opposed to an increase in factor volatilities, the change
in covariance is much more likely to be associated with the process of global
market integration”. (p2597). We compute the comovement between each REIT
and each stock relative to our common factors and the results are plotted in Fig. 5b.
Specifically, from Eq. (1) the covariance between the returns on stock i and REIT
j as
# $ # $
cov RiS , RjR = βiS ΣC βjR + cov iS , jR
Measuring Market Integration: US Stock and REIT Markets 245

where ΣC is the covariance matrix of some set of common factors.11 For M stocks
and N REITs, (equal-weighted) comovement is defined as

1  S
M N
ΓC = βi ΣC βjR . (6)
MN
i=1 j =1

From Eq. (6), comovement either increases due to increased factor volatility or by
increased factor loadings. Long trends in comovement persisting across multiple
periods are likely caused by increased factor loadings and therefore are consistent
with increasing integration. The time variation in comovement observed in Fig. 5b
is similar to R-square. REITs exhibit low comovement with stocks until early 2000
when it then begins a steady upward trend. The long last trend is most likely due to
increased betas as opposed to a persistent upward trend in factor volatility.
Both measures are consistent with our previous findings. There is not a strong
relationship between markets in the early part of the sample and the strongest
relationship is observed in the 2000s. The plots also demonstrate large time variation
in the relationships between the two markets. It is challenging, however, to make
any statement about integration from these two measures. Unlike our canonical
correlation approach, threshold values do not exist that describe regimes over which
integration is possible. Further, markets can be integrated and have moderate or
low levels of R-square and comovement.12 If markets are integrated at R-square or
comovement values of say 10%, then what is the interpretation when an increase
from 11% to 12% is observed? It cannot be the case that markets are even more
integrated than before. Similarly, if markets are not integrated and the measure
decreases, then it is not useful to say that markets are now less integrated than
they were in the previous period. Hence, in order to interpret these measures of
integration, one must know whether markets are already integrated and at what point
the markets change from being segmented to integrated—defeating the purpose of
the measures.
Two measures that do not suffer from these problems are the regime-switching
model of Bekaert and Harvey [4] and the canonical correlation methodology
proposed here. Bekaert and Harvey [4] measure of time varying integration allows
for easy interpretation by measuring integration on a zero to one scale with values
of zero indicating complete segmentation and one indicating complete integration.
Changes in this value are readily interpreted as either trending toward one extreme
state or the other. Similarly, the time-varying canonical correlations used in this
study clearly identify regimes over which common factors exist, a necessary
requirement for integration. Time variation in canonical correlations can be used

11 Local factors are assumed to be orthogonal across markets and therefore do not affect covariance.
12 Pukthuanthong and Roll [20] argue this point as a critique of comovement as a measure of
integration. This can also be true for R-square in markets characterized by high idiosyncratic risk.
246 D. W. Blackburn and N. K. Chidambaran

to track the process from the state when the two markets having two completely
different models for returns to the state when there is one market comprised of only
common factors.

5 Conclusion

Financial market integration requires that asset returns across distinct markets are
determined by a common set of factors and the risk premia for these factors be
equal across markets. The dual requirement makes econometric testing for market
integration challenging, in particular a failure to accept or reject integration could
be driven by an incorrect factor model. The approach we have presented in this
paper, based on factor analysis and canonical correlation analysis, overcomes the
joint hypothesis problem. Our approach is to first determine the correct factor
model in each market and determine whether markets share a common factor. We
subsequently develop economic proxies for the shared common factor and test for
the equality of risk premia conditional on a common factor being present for the
data period.
We illustrate the issues involved and the efficacy of our approach using the US
stock and REIT markets over the period from 1985 to 2013. We demonstrate that
traditional GMM tests are often times not able to reject segmentation thus implying
integration. Using canonical correlation analysis on the principal components
constructed from each market separately, we are able to identify the factor structure
for each market and show that often times this is because markets do not share a
common factor. By knowing when common factors exist, we are able to increase
the power of our testing methodology by testing only over those regimes when
integration is possible. Outside of these periods, time varying risk premia, noise
in the data, and changes in integration over time can lead to spurious conclusions.
We find that the REIT and stock markets share a common factor over the 2005–
2012 time period. Relative to the early part of the sample, risk premia over the
2005–2012 regime are indeed similar. The inability to reject equal risk premia in the
early part of the sample is due to the REIT premia exhibiting a strong cyclical time
varying pattern that is much different than the much more stable stock risk premia.
Our results strongly support the necessity of first identifying the correct model of
returns, both common and local factors, prior to testing for equal premia. We also
relate our results to other measures used to imply time variation in integration, such
as R-Square from asset pricing regressions and comovement. These measures do
provide confirming evidence but are much more difficult to interpret.
We note that U. S. REIT markets are relatively small, especially when compared
to the US stock markets. Further, tests have to be run each year as market integration
can vary from year to year. These data features reduce the power of standard
econometric techniques, such as GMM or Fama-Macbeth regressions that require
a large amount of data. Researchers have tried to increase the power of these tests
by increasing the data period or by making strong assumptions such as constant
Measuring Market Integration: US Stock and REIT Markets 247

betas and risk premia—assumptions that are not valid when market integration is
changing over time. Our methodology clearly works in small data sets and across
datasets of unequal sizes. We believe, therefore, our approach is generally applicable
and can be used in other settings where market capitalizations vary greatly or when a
market has a fewer number of assets, e.g. emerging equity and fixed income markets.

Acknowledgements We thank Mark Flannery, Ren-Raw Chen, Shu-Heng Chen, An Yan, Yuewu
Xu, Andre de Souza, seminar participants at Fordham University, the Indian School of Business
and participants at the 1st Conference on Recent Developments in Financial Econometrics and
Applications for their comments and suggestions.

Appendix

We briefly describe the essential elements of Canonical Correlations for identifying

common factors in this appendix.
Let the data generating processes for the two dataset be as given below.

Xkt = β k Ft + δ k Hkt + tk (7)

where there are r c common factors and r k set-specific factors leading to κk = r c +r k

total factors for sets k = 1, 2. Our goal is to determine r c , r 1 and r 2 . We determine
the number of principal components, κ1 and κ2 , required to span the factor space for
each individual set using [1]. We denote the first κ1 and κ2 principal components as
P1 and P2 , respectively.
We next use canonical correlation analysis to relate the two sets of principal
components P1 and P2 in order to determine the dimension of the subspace at
the intersection of the spaces spanned by the two sets. Originally developed by
Hotelling (1936), the idea behind canonical correlation analysis is to find matrices
that can be used to rotate two sets of factors so that their columns are ordered in
terms of decreasing pairwise correlation.13 Since the common factors are the only
factors shared across the two sets, this procedure separates the common factors from
the set-specific factors.
Let set X1 have κ1 factors and set X2 have κ2 factors, and let rmax c = min (κ1 ,κ2 )
be the maximum possible number of common factors. Canonical correlation
analysis finds two T × rmax c matrices U1 and U2 , called canonical variates, such
th
that the i columns of U1 and U2 , U1 (i) and U2 (i) have maximum correlation.

13 Canonical correlations have been used in prior literature on factor analysis. For example, Bai and

Ng [2] used the canonical correlation idea to test the equivalence of the space spanned by latent
factors and by some observable time series. Our use of canonical correlations is very much in the
flavour of Bai and Ng [2], but for very different purposes. We show that the methodology can be
used to determine common and unique factors in panels of data that have intersecting factors and
is able to identify the true factor structure in multilevel data.
248 D. W. Blackburn and N. K. Chidambaran

When there are r c common factors, then the pairwise correlation between the
i th canonical variates from each set will be large for i = 1, .., r c and small for
i = r c + 1, . . . , rmax
c .

Define the covariance matrix

 
 Σ% % Σ % %
= P1 P1 P1 P2
Σ%P2%P1 Σ%P2%P2

We search for vectors α1 and α2 such that the linear combinations %U1 (1) = α1% P1
and %U2 (1) = α2 %
 P2 have maximum correlation, and such that % U1 (1) and %U2 (1)
have zero mean and unit variance. We use the hat symbol to emphasize that the
principal components and rotations of the principal components are estimates of the
true factors. The objective function to be maximized is

1 #  $ 1 # $!
ρ̃1 = Max α1 Σ% P2 α2
P1% − λ1 α1 Σ%
P1%

P1 α1 − 1 − λ2 α2 Σ% P2 α2 − 1
P2% ,
2 2

where λ1 and λ2 are Lagrange multipliers. The objective function is maximized

when vectors α1 and α2 satisfy the first order conditions,
  ! !
−λ1 Σ%P1% Σ% P1% α1 0
P1 P2 = .
Σ% %
P2 P1 −λ 2 Σ%
P2%P2 α2 0

The maximum correlation is

ρ̃1 = λ1 = λ2 = α1 Σ% P2 α2 .
P1%

The correlation ρ̃1 is said to be the first canonical correlation, and %

U1 (1) and %U2 (1)
are referred to as first canonical variates.
The analysis can be repeated to find the second canonical variates % U1 (2) and
%
U2 (2) that are orthogonal to % U1 (1) and %U2 (1) , respectively, and have the second
canonical correlation ρ &2 ≤ ρ̃1 . We apply the process recursively to yield a collection
of canonical correlations ρ &1 ≥ ρ &2 ≥ . . . ≥ ρ &r̃max
c with the canonical variates
%
U1,T ×rmax
c and % U2,T ×rmax
c . The squared canonical correlations obtained above are

the eigenvalues of
−1 −1
Σ% ΣFx %
Fx %
Fx % Fy Σ% ΣFy %
Fy %
Fy % Fx . (8)

The properties of the squared canonical correlations have been well studied.
Anderson (1984) shows that if P1 and P2 are observed (not estimated) and are
normally distributed, then
Measuring Market Integration: US Stock and REIT Markets 249

√ ' 2 (
&i − ρi2
T ρ d
zi = −→ N(0, 1) (9)
ρi (1 − ρ
2& &i2 )

where ρ&i denotes the sample canonical correlation and ρi denotes the benchmark
correlation. Muirhead and Waternaux (1980) extend this result to show that if P1
and P2 are observed (not estimated) and are elliptically distributed, then
√ ' 2 (
1 T ρ &i − ρi2 d
zi = −→ N(0, 1) (10)
(1 + ν/3) 2&
ρi (1 − ρ
&i2 )

where ν is excess kurtosis. Bai and Ng [2] further extend these results to show that
if one set is estimated while the second set is observed, then Eqs. (9) and (10) still
hold.

References

1. Bai, J., & Ng, S. (2002). Determining the number of factors in approximate factor models.
Econometrics, 70, 191–221.
2. Bai, J., & Ng, S. (2006). Evaluating latent and observed factors in macroeconomics and finance.
Journal of Econometrics, 131, 507–537.
3. Bekaert, G., Hodrick, R., & Zhang, X. (2009). International stock return comovement. Journal
of Finance, 64, 2591–2626.
4. Bekeart, G., & Harvey, C. (1995). Time-varying world market integration. Journal of Finance,
50, 403–444.
5. Blackburn, D., & Chidambaran, N. (2014). Is the world stock return comovement changing?
In Working Paper, Fordham University.
6. Carrieri, F., Errunza, V., & Hogan, K. (2007). Characterizing world market integration through
time. Journal of Financial and Quantitative Analysis, 42, 915–940.
7. Errunza, V., & Losq, E. (1985). International asset pricing under mild segmentation: Theory
and test. Journal of Finance, 40, 105–124.
8. Fama, E., & MacBeth, J. (1973). Risk, return, and equilibrium: Empirical tests. Journal of
Political Economy, 81, 607–636.
9. Goyal, A., Perignon, C. & Villa, C. (2008). How common are common return factors across
NYSE and Nasdaq? Journal of Financial Economics, 90, 252–271.
10. Gultekin, M., Gultekin, N. B. & Penati, A. (1989). Capital controls and international capital
market segmentation: The evidence from the Japanese and American stock markets. Journal of
Finance, 44, 849–869.
11. Gyourko, J., & Keim, D. (1992). What does the stock market tell us about real estate returns.
Journal of the American Real Estate and Urban Economics Association, 20, 457–485.
12. Jagannathan, R., & Wang, Z. (1998). An asymptotic theory for estimating beta-pricing models
using cross-sectional regression. Journal of Finance, 53, 1285–1309.
13. Ling, D., & Naranjo, A. (1999). The integration of commercial real estate markets and stock
markets. Real Estate Economics, 27, 483–515.
14. Liow, K. H., & Webb, J. (2009). Common factors in international securitized real estate
markets. Review of Financial Economics, 18, 80–89.
15. Liu, C., Hartzell, D., Greig, W., & Grissom, T. (1990). The integration of the real estate
market and the stock market: Some preliminary evidence. Journal of Real Estate Finance and
Economics, 3, 261–282.
250 D. W. Blackburn and N. K. Chidambaran

16. Liu, C., & Mei, J. (1992). The predictability of returns on equity REITs and their co-movement
with other assets. Journal of Real Estate Finance and Economics, 5, 401–418.
17. Mei, J., & Saunders, A. (1997). Have U.S. financial institutions’ real estate investments
exhibited trend-chasing behavior? Review of Economics and Statistics, 79, 248–258.
18. Miles, M., Cole, R., & Guilkey, D. (1990). A different look at commercial real estate returns.
Real Estate Economics, 18, 403–430.
19. Newey, W., & West, K. (1987). A simple, positive semi-definite, heteroskedasticity and
autocorrelation consistent covariance matrix. Econometrica, 55, 703–708.
20. Pukthuanthong, K., & Roll, R. (2009). Global market integration: An alternative measure and
its application. Journal of Financial Economics, 94, 214–232.
21. Schnare, A., & Struyk, R. (1976). Segmentation in urban housing markets. Journal of Urban
Economics, 3, 146–166.
22. Shanken, J. (1992). On the estimation of beta pricing models. Review of Financial Studies, 5,
1–34.
23. Stock, J., & Watson, M. (2002). Forecasting using principal components from a large number
of predictors. Journal of the American Statistical Association, 97, 1167–1179, 32.
24. Stulz, R. (1981). A model of international asset pricing. Journal of Financial Economics, 9,
383–406.
25. Titman, S., & Warga, A. (1986). Risk and the performance of real estate investment trusts: A
multiple index approach. Real Estate Economics, 14, 414–431.
Supercomputer Technologies in Social
Sciences: Existing Experience and Future
Perspectives

Valery L. Makarov and Albert R. Bakhtizin

Abstract This work contains a brief excursus on the application of supercomputer

technologies in social sciences, primarily, in part of the technical implementation
of large-scale agent-based model (ABM). In this chapter, we will consider the
experience of scientists and practical experts in the launch of agent-based model
on supercomputers. On the example of agent model developed by us of the social
system in Russia, we will analyze the stages and methods of effective projection
of a computable core of a multi-agent system on the architecture of the acting
supercomputer.

Keywords Agent-based models · Parallel calculations · Supercomputer

technologies

1 Introduction

Computer modeling is the broadest, most interesting, and intensely developing area
of research and is in demand today in many spheres of human activity. The agent-
based approach to modeling is universal and convenient for practical researchers
and experts because of its visualization, but at the same time sets high requirements
for computing resources. It is obvious that significant computing capacities are
necessary for direct modeling of sufficiently long-term social processes in a given
country or on the planet as a whole.
Due to exponential growth of data volumes, the upcoming trend in agent-
based model (ABM) development is the ABM development using supercomputer
technologies (including those based on geoinformation systems). This direction
is developing rapidly today, and it is already the subject of discussions at world
congresses dedicated to ABM.

V. L. Makarov () · A. R. Bakhtizin

CEMI RAS, Moscow, Russia
e-mail: makarov@cemi.rssi.ru

© Springer Nature Switzerland AG 2018 251

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_13
252 V. L. Makarov and A. R. Bakhtizin

The relevance of using supercomputer technologies to develop ABM is explained

by the fact that an ordinary personal computer no longer has enough memory
to fit the number of programming environment items corresponding to, e.g., the
population of the Earth, or even some densely populated countries. The launch of
original models in specialized environments for ABM design with the number of
agents exceeding several million already exceeds the memory amount in a personal
computer.
The same situation is observed in terms of performance. Computing the state
of a large-scale system with nontrivial behavior logic and interaction between
agents requires substantial computing resources comparable to the demands of
computational methods used in mathematical physics with the same number of
computational cells. However, unlike the movements of particles, the social agents’
behavior comprises an element of chance, which makes necessary an additional
series of calculations to estimate the probability distribution of the key characteris-
tics in the final stage of the modeled environment.
The factors listed above stipulate the need for large-scale experiments using
supercomputer versions of the models in which the agents’ population is distributed
over a variety of supercomputer nodes and the calculations are performed in parallel.
This, in return, requires adaptation of the models built in traditional development
environments to supercomputer architecture.
Similarly to the supercomputer programs used to address numerous physical
tasks, the potential for parallelizing multi-agent systems is in using the locality of
agent interaction. Most of the interactions in the model, just like in real life, occur
between subjects located near each other. This allows parallelizing “by space,” i.e.,
ensuring a more uniform distribution of the agents’ population over supercomputer
nodes given their geographic proximity. Thus, breaking down the territory occupied
by the agents into the so-called blocks provides the basic capacity for parallelizing
the task. This is the approach most commonly used in practice for cases where
distribution of elements in an environment being modeled. Whether we model
agents in multi-agent systems or individual computational cells with averaged
parameters of the physical environment, they should meet the special localization
principle: the connections and exchange of data happen mostly for elements with
close coordinates and are processed almost instantly within each computational
node.
Supercomputers allow to enormously increase the number of agents and other
quantitative characteristics (network nodes, and the size of territory) of models,
originally developed for use on ordinary desktop computers. For this reason,
supercomputer modeling is a logical and desirable step for those simplified mod-
els, which have already passed practical approbation on conventional computers.
Unfortunately, the specific architecture of modern computers does not guarantee that
the software of a computer model will immediately work on a supercomputer. The
paralleling of the computable core is required as a minimum as well as frequently
its deep optimization. In the absence of these adjustments, the use of expensive
supercomputer calculation will most likely not pay off.
Supercomputer Technologies in Social Sciences 253

2 Experience of Some Scientists and Practical Experts

In September of 2006, a project on the development of a large-scale ABM of the

European economy—EURACE, i.e., Europe ACE (Agent-based Computational
Economics), was launched, with a very large number of autonomous agents,
interacting within the socioeconomic system [4]. Economists and programmers
from eight research centers in Italy, France, Germany, Great Britain, and Turkey
are involved in the project, including an advisor from Columbia University, USA,
the Nobel Prize winner Joseph Stiglitz.
According to the developers, virtually all existing ABMs either cover only a
single industry or a relatively restricted geographical area and accordingly, small
populations of agents, while the EURACE presents the entire European Union, so
the scope and complexity of this model is unique, and its numerical computation
requires the use of supercomputers as well as special software.
The information about 268 regions in 27 countries was used to fill the model with
necessary data, including some geoinformation maps.
In the model, there are three types of agents: households (up to 107 ), enterprises
(up to 105 ), and banks (up to 102 ). They all have a geographical reference, and are
also linked to each other through social networks, business relationships, etc.
EURACE was implemented using a flexible scalable environment for simulating
agent-based model—FLAME (Flexible Large-scale Agent Modeling Environ-
ment), developed by Simon Coakley and Mike Holcombe1 initially to simulate the
growth of cells under different conditions. With the help of the developed model,
several experiments were conducted in order to study the labor market. Without
going into detail about the obtained numerical results, we note that, according to the
authors, the main conclusion of the research is that the macroeconomic measures
of two regions with similar conditions (resources, economic development, etc.)
during a long period (10 years and more) may vary significantly, due to an initial
heterogeneity of the agents.2
In ABM EpiSims, developed by researchers from the Virginia Institute of bioin-
formatics (Virginia Bioinformatics Institute), the movement of agents is studied as
well as their contacts within an environment as close as possible to reality and
containing roads, buildings, and other infrastructure objects [10]. To develop this
model, a large array of data was necessary, including information about the health
of individual people, their age, income, ethnicity, etc.
The original goal of the research was to construct an ABM of high dimension
to be launched on the supercomputer, which could be used to study the spreading
of diseases in society. However, afterwards, in the course of work, another task
was also being resolved, regarding the creation of specialized ABM++ software,
which allows to carry out the development of ABM in the C++ language, and also

1 For more thorough information, see www.flame.ac.uk.

2 More thorough information can be found on the website of the project: www.eurace.org.
254 V. L. Makarov and A. R. Bakhtizin

contains functions, facilitating the allocation of the program code in use among
the cluster nodes of the supercomputer. Apart from that, ABM++ provides for
the possibility of dynamic redistribution of currents of calculations as well as the
synchronization of ongoing events.
ABM++, the first version of which appeared in 2009, is the result of the
modernization of the instrument, developed in 1990–2005 in the Los Alamos
National Laboratory during the process of constructing large-scale ABMs (EpiSims,
TRANSIMS, and MobiCom).
Specialists of another research team from the same Bioinformatics Institute of
Virginia created an instrument for the study of the particularities of the spreading of
infectious diseases within various groups of society—EpiFast, among the assets of
which is the scalability and high speed of performance. For example, the simulation
of social activity of the population of Greater Los Angeles Area (agglomerations
with a population of over 17 million people) with 900 million connections between
people on a cluster with 96 dual-core processors POWER5 took less than 5 min.
Such fairly high productivity is provided by the original mechanism of paralleling
presented by the authors [2, 7].
Classic standard models of spread of epidemics were mostly based on the
use of differential equations; however, this tool complicates the consideration of
connections between separate agents and their numerous individual particularities.
ABM allows to overcome such shortcomings. In 1996, Joshua Epstein and Robert
Axtell published a description of one of the first ABMs, in which they reviewed the
process of the spread of epidemics [6]. Agent models, which differ from each other
in their reaction to the disease, which depends on the state of their immune system,
are spread out over a particular territory. At that, in this model, agents, the number
of which constitutes a mere few thousand, demonstrate fairly primitive behavior.
Later on, under the supervision of Joshua Epstein and Jon Parker at the Center
on Social and Economic Dynamics at Brookings, one of the largest ABMs was
constructed, which included data about the entire population of the USA, that is
around 300 million agents [9]. This model has several advantages. First of all,
it allows to predict the consequences of the spread of diseases of various types.
Second of all, it focuses on the support of two environments for calculations: one
environment consists of clusters with an installed 64-bit version of Linux, and the
other of servers with quad-core processors and an installed Windows system (in
this regard, Java was chosen as the language of the programming, although the
developers did not indicate which particular version of Java they used). Third of
all, the model is capable of supporting from a few hundred million to six billion
agents.
The model in question (the US National Model) includes 300 million agents,
which move around the map of the country in accordance with the mobility plan of
4000 × 4000 dimensions, specified with the help of a gravity model. A simulation
experiment was conducted on the US National Model, imitating the 300-day long
process of spreading a disease, which is characterized by a 96-h incubation period
and a 48-h infection period. In the course of the study, among other things, it was
determined that the spreading of the disease was declining, after 65% of the infected
Supercomputer Technologies in Social Sciences 255

got better and obtained immunity. This model has repeatedly been used by the
specialists of the Johns Hopkins University as well as by the US Department of
National Security, for research, dedicated to the strategy of rapid response to various
types of epidemics [5].
In 2009, a second version of the US National Model was created, which included
6.5 billion agents, whose actions were specified taking into consideration the
statistical data available. This version of the model was used to imitate the spreading
of the A(H1N1/09) virus all over the planet.
Previously, this kind of model was developed by the Los Alamos National
Laboratory (USA), and the results of the work with this model were published
on April 10, 2006 [1]. One of the most powerful computers which existed at the
time known by the name of “Pink,” which consisted of two 1024 processors with a
2.4 GHz frequency and a 2 GB memory each was used for the technical realization
of the model. This large-scale model, composed of 281 million agents, was used to
study scenarios of the spreading of various viruses, including the H5N1, and took
into consideration several possible operational interventions such as vaccinations,
closing of schools, and introducing of quarantines in some territories.
Researchers at Argonne National Laboratory have been successfully using
a new modeling paradigm agent-based modeling and simulation (ABMS)—to
address challenges and gain valuable insights in such key areas as energy, biology,
economics, and social sciences. To maximize potential, they are developing a next-
generation ABMS system that can be extended to exascale computing environments
to achieve breakthrough results in science, engineering, and policy analysis.
Argonne researchers have developed and used large-scale agent-based model
to provide important information to policymakers that would not be available
using other modeling approaches. One outstanding example—Electricity Markets
Complex Adaptive Systems (EMCAS)—was used to model the Illinois electric
power industry under deregulation conditions in an effort to anticipate the likely
effects of deregulation on electricity prices and reliability.3
Interdisciplinary project on modeling of technological, social, and economic
systems of the world was launched in 2012 and has involved scientists from
almost all developed countries. Its implementation period is 10 years and the
initial funding is 1 billion Euros.4 Project managers emphasize the use of advanced
information technologies (first of all, the agent-based model on the basis of
geographic information systems). According to one of the leaders of the project—
Dirk Helbing, despite the relevancy of developing such multilevel systems and the
existence of some of their components, an integrated product is still missing due to
institutional barriers and lack of resources. In this regard, FuturICT promises to
become the first of its kind.

3 Electricity
Market Complex Adaptive System (EMCAS), Argonne National Laboratory, http://
www.dis.anl.gov/projects/emcas.html.
4 www.futurict.eu.
256 V. L. Makarov and A. R. Bakhtizin

FuturICT is organized as a network of national centers, each of which is

represented by several scientific institutions within the same country. In addition
to this network, there is a problem-oriented network, which uses the main national
science centers for solving specific issues. Thus, FuturICT integrates organizations
on institutional and problem-solving level.
Every scientific community at the country level has a certain degree of autonomy,
but also a set of obligations. Developers offer the FuturICT platform, which includes
the three following components:
1. the nervous system of the planet (Planetary Nervous System);
2. simulator of a living planet (Living Earth Simulator);
3. global unified platform (Global Participatory Platform).
A set of models, forming a “simulator of a living planet,” through “observatories”
will allow for the detection of crises and finding solutions for mitigation. These
models will be verified and calibrated using data collected in real time using the
“nervous system of the planet.” Ultimately, decision-makers will interact with the
global unified platform, which will combine the results of the first two parts of
FuturICT.
Various “observatories” (financial, economic, energy, and transport) are arranged
in four groups of main directions of research in the field of public relations, tech-
nological and economic development as well as in monitoring of the environmental
state.
“The nervous system of the planet” can be represented in the form of a global
network of sensors that collect information on socioeconomic, technological, and
ecological systems of the world. For its construction, the project coordinators of
FuturICT work closely with the team of Sandy Pentland from the Massachusetts
Institute of technology (MIT) in order to “connect the sensors to modern gadgets.”
In the framework of the “simulator of the living planet,” an open software
platform will be implemented, which, according to the initiators of the project,
will remind the well-known App Store in the sense that scientists, developers, and
interested people will be able to upload and download other models related to
the various parts of the planet. The basic modeling approach will be based on an
agent-based paradigm. In the future, unified model components are expected to be
implemented with the use of supercomputer technologies.
“Global unified platform,” will also serve as an open platform for discussion of
the forecasts of development of the world in regard to citizens, government officials,
and the business community obtained using FuturICT.
There are several tools for high-performance computing for ABM.
Microsoft Axum is a domain-specific concurrent programming language, based
on the Actor model that was under active development by Microsoft between 2009
and 2011. It is an object-oriented language based on the .NET Common Language
Runtime using a C-like syntax which, being a domain-specific language, is intended
for the development of portions of a software application that is well suited to
concurrency. But, it contains enough general-purpose constructs that one need not
Supercomputer Technologies in Social Sciences 257

switch to a general-purpose programming language (like C#) for the sequential parts
of the concurrent components.5
The main idiom of programming in Axum is an Agent (or an Actor), which is an
isolated entity that executes in parallel with other Agents. In Axum parlance, this is
referred to as the agents executing in separate isolation domains; objects instantiated
within a domain cannot be directly accessed from another.
Agents are loosely coupled (i.e., the number of dependencies between agents
is minimal) and do not share resources like memory (unlike the shared memory
model of C# and similar languages); instead, a message passing model is used. To
coordinate agents or having an agent request the resources of another, an explicit
message must be sent to the agent. Axum provides Channels to facilitate this.
The Axum project reached the state of a prototype with working Microsoft Visual
Studio integration. Microsoft had made a CTP of Axum available to the public, but
this has since been removed. Although Microsoft decided not to turn Axum into a
project, some of the ideas behind Axum are used in TPL Dataflow in .Net 4.5.
Repast for High-Performance Computing (Repast HPC) 2.1, released on 8 May
2015, is a next-generation agent-based modeling and simulation (ABMS) toolkit for
high-performance distributed computing platforms.
Repast HPC is based on the principles and concepts developed in the Repast
Simphony toolkit. Repast HPC is written in C++ using MPI for parallel operations.
It also makes extensive use of the boost (http://boost.org) library.
Repast HPC is written in cross-platform C++. It can be used on workstations,
clusters, and supercomputers running Apple Mac OS X, Linux, or Unix. Portable
models can be written in either standard or Logo-style C++.
Repast HPC is intended for users with:
• Basic C++ expertise
• Access to high-performance computers
• A simulation amenable to a parallel computation. Simulations that consist of
many local interactions are typically good candidates.
Models can be written in C++ or with a Logo-style C++.6
CyberGIS Toolkit is a suite of loosely coupled open-source geospatial software
components that provide computationally scalable spatial analysis and modeling
capabilities enabled by advanced cyberinfrastructure. CyberGIS Toolkit represents
a deep approach to CyberGIS software integration research and development and
is one of the three key pillars of the CyberGIS software environment, along with
CyberGIS Gateway and GISolve Middleware.7

5 https://www.microsoft.com/en-us/download/details.aspx?id=21024.
6 More thorough information on the S/W can be found in the User Manual [3]; a new version of
the 1.0.1. package (dated March 5, 2012) can be downloaded from the following website: http://
repast.sourceforge.net/repast$_$hpc.html.
7 http://cybergis.cigi.uiuc.edu/cyberGISwiki/doku.php/ct.
258 V. L. Makarov and A. R. Bakhtizin

The integration approach to building CyberGIS Toolkit is focused on developing

and leveraging innovative computational strategies needed to solve computing-
and data-intensive geospatial problems by exploiting high-end cyberinfrastructure
resources such as supercomputing resources provided by the Extreme Science and
Engineering Discovery Environment and high-throughput computing resources on
the Open Science Grid.
A rigorous process of software engineering and computational intensity analysis
is applied to integrate an identified software component into the toolkit, including
software building, testing, packaging, scalability and performance analysis, and
deployment. This process includes three major steps:
1. Local build and test by software researchers and developers using continuous
integration software or specified services;
2. Continuous integration testing, portability testing, small-scale scalability testing
on the National Middleware Initiative build and test facility; and
3. XSEDE-based evaluation and testing of software performance, scalability, and
portability. By leveraging the high-performance computing expertise in the
integration team of the NSF CyberGIS Project, large-scale problem-solving
tests are conducted on various supercomputing environments on XSEDE to
identify potential computational bottlenecks and achieve maximum problem-
solving capabilities of each software installation.
Pandora is a novel open-source framework designed to accomplish a simulation
environment corresponding to the above properties. It provides a C++ environment
that automatically splits the run of an ABM in different computer nodes. It allows
the use of several CPUs in order to speed up the decision-making process of agents
using costly AI algorithms. The package has also support for distributed execution
and serialization through HDF5, and several analysis techniques (spatial analysis,
statistics and geostatistics, etc.).8
In addition, pyPandora is a Python interface to the framework, designed to allow
people with minimal programming background a tool to develop prototypes. The
ability to develop ABMs using Python is an important feature for social scientists,
because this programming language is also used in other common modeling tools,
like Geographical Information Systems (i.e., ESRI ArcGIS).
PyPandora allows researchers to create ABMs using a language that they
already know and enables the design of a common framework where one can
combine simulation, spatial analysis, and geostatistics. Finally, the package is
complemented by Cassandra, a visualization tool created to detect spatiotemporal
patterns generated by the simulation. This application allows any user to load the
data generated by a Pandora parallel execution into a single computer and analyze it.
SWAGES, a distributed agent-based Alife simulation and experimentation envi-
ronment that uses automatic dynamic parallelization and distribution of simulations
in heterogeneous computing environments to minimize simulation times.

8 https://www.bsc.es/computer-applications/pandora-hpc-agent-based-modelling-framework.
Supercomputer Technologies in Social Sciences 259

SWAGES allows for multi-language agent definitions, uses a general plug-in

architecture for external physical and graphical engines to augment the integrated
SimWorld simulation environment, and includes extensive data collection and anal-
ysis mechanisms, including filters and scripts for external statistics and visualization
tools.9 Moreover, it provides a very flexible experiment scheduler with a simple,
web-based interface and automatic fault detection and error recovery mechanisms
for running large-scale simulation experiments.
A Hierarchical Parallel simulation framework for spatially explicit Agent-Based
Models (HPABM) is developed to enable computationally intensive agent-based
model for the investigation of large-scale geospatial problems. HPABM allows for
the utilization of high-performance and parallel computing resources to address
computational challenges in agent-based model.10
Within HPABM, an agent-based model is decomposed into a set of sub-models
that function as computational units for parallel computing. Each sub-model is
comprised of a subset of agents and their spatially explicit environments. Sub-
models are aggregated into a group of super-models that represent computing tasks.
HPABM based on the design of super- and sub-models leads to the loose coupling of
agent-based model and underlying parallel computing architectures. The utility of
HPABM in enabling the development of parallel agent-based model was examined
in a case study.
Results of computational experiments indicate that HPABM is scalable for
developing large-scale agent-based model and, thus, demonstrate efficient support
for enhancing the capability of agent-based modeling for large-scale geospatial
simulation.
The growing interest in ABM among the leading players in the IT industry
(Microsoft, Wolfram, ESRI, etc.) definitely shows the relevance of this instrument
and its big future, while exponential growth of overall data volumes related to
human functioning and the need for analytical systems to obtain new-generation
data needed to forecast social processes call for the use of supercomputer technolo-
gies.
There are currently several international associations uniting groups of
researchers from the largest institutes and universities working in this direction.
The most famous ones are: (1) North American Association for Computational
Social and Organizational Sciences (NAACSOS); (2) European Social Simulation
Association (ESSA); and (3) Pacific Asian Association for Agent-Based Approach
in Social Systems Science (PAAA). Each association holds regular conferences on
social modeling in America, Europe, and Asia, respectively. A World Congress on
the topic is also conducted every 2 years.
In Russia, ABMs are a relatively new scientific development, pioneered by the
Central Economics and Mathematics Institute of the Russian Academy of Sciences.

9 http://www.hrilab.org/.
10 https://clas-pages.uncc.edu/wenwu-tang.
260 V. L. Makarov and A. R. Bakhtizin

The content and results of the studies conducted by this institution are described in
the second part of this chapter.

3 Adaptation of Agent-Based Models for the

Supercomputer: Our Approach

3.1 Experiment #1

In March of 2013, the model was launched on the supercomputer Lomonosov which
simulated the socioeconomic system of Russia for the next 50 years. This ABM
is based on the interaction of 100 million agents, which hypothetically represent
the socioeconomic environment of Russia. The behavior of each agent is set by a
number of algorithms, which describe its actions and interaction with other agents
in the real world.
Five people participated in the project: two specialists of the Central Economic
and Mathematical Institute of the Russian Academy of Sciences (V.L. Makarov
and A.R. Bakhtizin) and three researchers of the Moscow State University (V.A.
Vasenin, V.A. Roganov, and I.A. Trifonov) [8]. The data for the modeling was
provided by the Federal Service of State Statistics and by the Russian Monitoring
of the Economic Conditions and Health of the Population. A model for a standard
computer was developed in 2009, and in 2013 it was converted into a supercomputer
version. Below, we will examine the main stages and methods of effective projection
of a computable kernel of a multi-agent system on the architecture of a modern
supercomputer, which we have developed during the process of resolving the issue
in question.

3.1.1 The Problem of Scaling

It is important to understand that the scaling of programs for supercomputers is

a fundamental problem. Although the regular and supercomputer programs carry
out the same functionalities, the target functions of their development are usually
different.
During the initial development of the complex application software, the first and
foremost goal is to try to minimize the costs of programming, personnel training,
enhance the compatibility between platforms, etc., and to leave optimization for
later. This is quite reasonable, since at the early stages, the priority of development
is exclusively the functionality.
However, when the developed software is already being implemented, it is often
discovered that there is not enough productivity for real massive data. And, since
modern supercomputers are not simple computers, but work thousands of times
faster than personal computers, in order to launch the program on a supercomputer
Supercomputer Technologies in Social Sciences 261

it is necessary to introduce significant alterations first. Doing this effectively without

special knowledge and skills is by far not always successfully achieved.
Doing work properly and correctly usually leads to significant increase in
efficiency on the following three levels:
1. multisequencing of calculation;
2. specialization of calculative libraries by task;
3. low-level optimization.

3.1.2 Specialization and Low-Level Optimization

Before seriously talking about the use of supercomputers, the program must be
optimized to the maximum and adapted to the target hardware platform. If this is
not done, the parallel version will merely be a good test for the supercomputer, but
the calculation itself will be highly inefficient.
To use a supercomputer without optimization and adaptation of the program
to the target hardware platform is the same as sending a junior regiment on a
combat mission: first of all, it is necessary to teach the recruits how to properly
carry out their tasks (specialization, and optimization of software), as well as how
to efficiently handle weapons (low-level optimization of software), and only then
considerations of effective use of resources can really enter into account.
In the universal systems of modeling of the AnyLogic type, the procedures
presented are universal. And, a universal code can often be optimized for a particular
family of tasks.

3.1.3 Selection of a Modeling Support System

Certainly, ABM can be programmed without a special environment, in any object-

based language. In addition, the main shortcoming of the existing products for
creating ABM except for RepastHPC is its inability to develop projects that would
run on a computing cluster (i.e., there is no mechanism for paralleling the process
of executing the program code).
However, a more reasonable approach would be to use one of the proven systems
for ABM because of the unified implementation of standard ways of interacting
agents. In this chapter, we will limit our scope to the ADEVS system.11
ADEVS is a set of low-level libraries for discrete modeling done in C++
language. Some of the advantages worth mentioning are the following:
• ease of implementation;
• high efficiency of models;

11 TheADEVS system and its description can be downloaded from here: http://www.ornl.gov/~
1qn/adevs/.
262 V. L. Makarov and A. R. Bakhtizin

• support of basic numerical methods used for modeling;

• built in paralleling of simulation with the help of OpenMP;
• the possibility of using standard means of paralleling;
• fairly rapid development of libraries in the current time;
• Cross-platform software;
• low-level software (current functionality does not impose any restrictions on the
model);
• independence of the compiled code on unusual libraries;
• open-source code.
However, significant shortcomings of this product is a complete lack of means
of presentation and quite complicated modeling, when compared to, for example,
AnyLogic. Therefore, this product cannot be used to build models on the consumer
level, however, is an effective platform for implementing parallel simulations.
The main elements of the program when using the ADEVS library to build an
ABM are the following:
• adevs simulator:: Simulator< X >;
• primitive of adevs agents:: Atomic< X >;
• model of adevs (container of agents ):::: Digraph< V ALU E, P ORT >.
It was decided to develop the supercomputer version of the program described
below, based on the ADEVS system, in view of its advantages listed above. Within
the framework of this project, an MPI version of the ADEVS simulator was created,
as well as a visualization system of the calculation process based on the Qt library—
a cross-platform set of tools for creating software written in the C++ programming
language.
Next, we turn to a brief description of the developed model and the procedures
for its subsequent run on a supercomputer.

3.1.4 The Initial Agent-Based Model

The first stage of development of the ABM described below is constructing a tool
that effectively solves the problem of research on conventional computers as well as
adjusting the parameters of the model. The model is then tested on a conventional
computer with a small number of agents (depending on the complexity of agents,
conventional computers are able to perform calculations at a satisfactory rate and
with good productivity for approximately 20 thousand agents). During the first
stage, the AnyLogic product was used, the technical capabilities of which allowed to
debug the model at a satisfactory speed and to configure its settings. After successful
testing, we proceeded to the second stage of development: to convert the model so
that it could be used on a supercomputer. The model for an ordinary computer was
built in 2009, and it was converted into a supercomputer version in 2013.
Figure 1 shows the operating window of the developed ABM (dots—agents).
During the operation of the system, current information can be obtained on the
Supercomputer Technologies in Social Sciences 263

Fig. 1 Operating window of the developed ABM

socioeconomic situation in all regions of Russia, including the use of cartographic

data, changing in real time depending on the values of the endogenous parameters.
The specification of the agents of the model was carried out taking into consider-
ation the following parameters: age, life expectancy, specialization of parents, place
of work, region of residence, income, and others.
The specification of regions was carried out, taking into consideration the
following parameters: geographic borders, population, number of workers (by type),
GRP, GRP per capita, volume of investments, volume of investments per capita,
average salary, average life expectancy, index of population growth, etc.
Statistics manuals of Rosstat as well as sociological databases of RLMS were
used to fill the model with the necessary data.
Agents in the model are divided into two groups (types) with different repro-
ductive strategies. Agents of the first type follow the traditional strategy, known for
high birth rate, while the second group follows the modern strategy with a much
lower birth rate. The model starts working by initializing its starting environment
conditions and creating a population of agents with personal characteristics (age,
gender, reproductive strategy type, and the number of children sought) assigned so
as to reproduce the age, gender, and social structure of the population in the region
being modeled.
Next, the natural population movement processes—mortality and birth—are
imitated using the age movement methods and probability mechanisms. Agents
die following the mortality rates that are differentiated by age and gender but
corresponding to the actual figures for the entire population. Meanwhile, the
creation of new agents in the model (childbirth) is a result of other agents’ actions.
264 V. L. Makarov and A. R. Bakhtizin

First, individual agents interact to form couples and agree on how many children
they want to have together. Next, the couples agree on when each child will be born,
depending on them belonging to one or another type.
The model was used to conduct experiments forecasting changes in the popula-
tion of agents inhabiting a certain region, the age structure of this population, and the
correlation of agent numbers in various types for main age groups and the population
in general. Experiments were conducted with the following parameter values: total
population—20,000; share of agents with traditional reproductive strategy type—
10%. The outcomes show that the model adequately imitates processes observed
in real life as a reduction in the overall population numbers, as well as its aging—
reduction of the younger population groups and increase in older groups.

3.1.5 Conversion of the Model into a Supercomputer Program

Earlier, we had already discussed the problems of using ABM development

tools for the realization of projects, carried out on the computing clusters of the
supercomputer. Due to the difficulties in separating the computing part from the
presentational part as well as to the realization of the code using a high level of the
JAVA language, the productivity of the implementation of the code is significantly
lower for AnyLogic than for ADEVS. Apart from that, it is extremely difficult to
reprocess the generated code into a concurrently executed program.
Below is the algorithm of the conversion of the AnyLogic model into a
supercomputer program.

Translation of the Model

The models in the AnyLogic project are kept in the format of an XML file,
containing the tree diagram of the parameters necessary for the generation of the
code: classes of agents, parameters, elements of the presentation, and descriptions
of the UML diagrams of the behavior of agents.
During the work of the converter, this tree diagram is translated into code C++ of
the program, calculating this model. The entry of the tree is executed depth wise. At
that, the following key stages are marked, and their combination with the translation
process is carried out.
1. Generating the main parameters. The search for the root of the tree and the
reading of the parameters of daughter nodes, such as the name of the model,
address of assembly, type of model, and type of presentation.
2. Generating classes. Generating classes (more detailed):
(a) configuration of the list of classes;
(b) reading of the main class parameters;
(c) reading of the variables;
(d) reading of the parameters;
Supercomputer Technologies in Social Sciences 265

(e) reading of the functions;

(f) generating a list of functions;
(g) reading the functions code;
(h) conversion of the functions code Java − > C + +;
(i) reading of the figures and elements of control that are being used;
(j) generating the code of initialization of figures and elements of control;
(k) generating constructor, destructor, and visualizer codes;
(l) generating the class structure;
(m) generating header code and source files.
3. Generating the simulator. Search for the peak, containing the information
about the process of simulation (controlling elements, significance of important
constants, elements of presentation, etc.).
4. Generating shared files of the project (main.cpp, mainwindow.h, mainwin-
dow.cpp, etc.).
Import of Incoming Data Data from the geoinformation component of the initial
model (map of Russia), containing all of the necessary information, is imported into
the model as input data.
Generating Classes and the Transformation of the Functions Code When generat-
ing the functions from the tree, the following information is read: name of function,
return type, parameters, and its body.
Based on the list of classes constructed earlier, changes are introduced into the
arguments of the functions, replacing the heavy classes, i.e., all generated classes,
classes of figures, and other classes, which do not form a part of the standard set
with corresponding indicators. The purpose of this is to save memory space and
avoid mistakes when working with it. After that, the titles of the functions are
generated, which are later inserted into the title and source files. In the course of
such reading, the body of the function, by the means of the relevant function (listing
5), is transformed from a Java-based code into an analogical C++ code (this is
possible due to the fairly narrow class of used functions; as for more complicated
functions, manual modification of the translated code is required), after which it is
added to the list of bodies for this class.
In the course of the translation, the need for the transformation of the initial
functions code from the Java language to the C++ often arises. It can be presented
in the form of sequential replacements of constructions, for example:
• Transformation of cycles: Java format.
• Transformation of indicators. Java, unlike C++, does not contain such an
obvious distinction between the object and the object indicator, hence the
structure of the work with them does not differ. That is why, a list of classes is
introduced, in which it is important to use operations with object indicators, and
not with the object itself, and all of the variables of such classes are monitored
with the subsequent replacement of addresses to the objects with corresponding
addresses to the object indicators within the framework of the given function.
266 V. L. Makarov and A. R. Bakhtizin

• The opening of black boxes. In Java, and in the AnyLogic library in particular,
there is a certain number of functions and classes, which do not have analogues
in the C++ itself, nor in the ADEVS library. Due to this fact, additional libraries,
such as, shapes.h and mdb-work.h, had been created, which compensate for the
missing functions.
• During the generating stage of the main parameters of the lists of classes the
name of the main class and the names of the modulated agent classes are
obtained. The procedure of adding an agent into the visibility range of the
simulator is introduced into the code of the main class.
Generating Outside Objects In the process of generating outside objects, a separate
function Main :: initShapes() is created, which contains all of the graphic
information, i.e., the initialization of all figures, the classes of which had also
been implemented in the shapes.h., and is carried out within the framework of the
function. The relevant example is presented in the following code fragment.
Generating Classes and the Code of the Title and Source Files Based on all the data
that has been read and generated, the title and source files of the corresponding class
are created.
Generating Simulation For the generation of simulation, it turned out to be enough
to have the main.cpp, mainwindow.cpp, mainwindow.h files, written beforehand,
in which the templates define the type of the main class and the added title files.
When compiling the initial code, the templates are replaced with the names of the
classes received earlier (at the generating stage). This is enough for the double-
flow simulation, which can later be replaced with a corresponding module for a
multiprocessor simulation.
Additional Attributes At the stage of analyzing the tree (see above), a tree, similar
in structure, is formed for the generation of a C++ code, with the help of which the
necessary attributes of compilation can be set (visualization of certain parts, visual
validation of code recognition, additional flags of assembly etc.), during the stage of
preparation for translation. After that, at receiving the command for transformation,
the final compilation takes place, taking into consideration all of these attributes.
Assembly of the Ready Project For the assembly of the translated project, the
QtCreator is used—cross-platform shareware integrated environment for work with
the Qt framework.
Agent Code With the help of the translator described above, an initial code (except
for the behavior pattern of agent) has been generated from the data of the files of the
AnyLogic project (model.alp and others).
The behavior pattern of the agent must be generated from the diagram of
conditions; however, currently the automation of this process has not yet been
implemented. Therefore, a certain volume of the code had to be added to the
generated code.
Supercomputer Technologies in Social Sciences 267

Fig. 2 Result of the work of the supercomputer program in graphic format

After the introduction of the necessary changes, a cross-platform application,

repeating the main functionality of the given model, was achieved as a result of the
compilation.
Statistics and Visualization of Time Layers Given the noninteractive mode of
the launching of the program on big supercomputers, the collection of data and
visualization were separated (this has to do with the load imbalance on clusters at
various times of the day; as for exclusive access, it is simply impossible). After the
recalculation of the model, the information obtained can once again be visualized,
for example, in the following manner (Fig. 2).
Supercomputers Available for Calculations At the moment of making the calcula-
tions, three supercomputers were available to us (Table 1), which were in the top
five of the supercomputer rating of the Top-50 supercomputers in the CIS countries.
For our calculations, we used two supercomputers—the “Lomonosov” and the
MVS-100K.
268 V. L. Makarov and A. R. Bakhtizin

Table 1 Supercomputers available for research group

Position in Top-50 Supercomputers Nodes CPU Kernels RAM/node TFlops
1 Lomonosov (Moscow 5130 10,260 44,000 12 GB 414
State University)
4 MVS-100K 1256 2332 10,344 8 GB 123
(Interagency
Supercomputing
Center of the RAS)
5 Chebyshev (Moscow 633 1250 5000 8 GB 60
State University)

3.1.6 Results

By using supercomputing technologies and optimizing the program code, we were

able to achieve very high productivity.
The optimization of the code and the use of C++ instead of Java allowed for the
increase in speed of execution of the program. The model was tested in the following
initial conditions: (1) number of agents, 20 thousand; and (2) forecasting period, 50
years. The results of the calculations showed that the count time of the model using
ADEVS amounted to 48 s using one processor, whereas the count time of the model
using AnyLogic and a single processor amounted to 2 min and 32 s. It means that
the development framework was chosen correctly.
As has already been noted above, an ordinary personal computer with high
productivity is able to carry out calculations with satisfactory speed with a total
number of 20 thousand agents, provided that the behavior of each agent is defined
by approximately 20 functions. The average count time of one unit of model time
(1 year) amounted to around 1 min. When dealing with a larger number of agents,
for example, 100 thousand, the computer simply freezes.
Using the 1000 processors of the supercomputer and executing the optimized
code allowed to increase the number of agents to 100 million and the number of
model years to 50. All this enormous volume of research was carried out in a period
of time which approximately equaled 1 min and 30 s (this indicator may slightly
vary depending on the type of processors used).
The results of the modeling showed that if the current tendencies persist, the
population of Russian Siberian and Far-Eastern Federal Districts will significantly
decrease, while the Southern Federal District, on the contrary, will see a significant
increase in population. In addition to that, the model suggests a gradual decrease of
Russian GDP as well as of several other macroeconomic indicators.
The results of the experiments carried out using the developed ABM revealed that
the scaling of the model in itself has certain meaning. For example, when launching
the same version of the model for 50 model years using the same parameters, except
for the number of agents (in the first case, there were 100 million agents, and in the
second, 100 thousand agents), the results received diverged by approximately 4.5%.
Supercomputer Technologies in Social Sciences 269

Fig. 3 Model performance (X axes number of processors, Y axes time in seconds)

It can be assumed that in complex dynamic systems the same parameters (birth
rate, life expectancy, etc.) may produce different results depending on the size of the
community.
Then, we continued to increase the number of processors under the same model
parameters in order to establish a dependency of the time for computation from the
resources involved. Here are the results (Fig. 3).

3.2 Experiment #2

That model version did not have interagent communication that, on the one hand, did
not allow to use the agent approach to the full extent, and on the other, significantly
simplified paralleling of the source code. In the current version, agents communicate
with each other that resulted in changing of the model paralleling technology as well
as other software libraries. Six people participated in the project: three specialists
of the Central Economic and Mathematical Institute of the Russian Academy of
Sciences (V.L. Makarov, A.R. Bakhtizin, and E.D. Sushko) and three researchers of
the Moscow State University (V.A. Vasenin, V.A. Roganov, and V.A. Borisov).
Besides a common paralleling scheme that determines the model source code
translation method, it is necessary to meet the requirements baseline of up-to-date
supercomputers for which the de facto standard is focused on MPI (Message Passing
Interface).
The use of Java standard environment is undesirable, because it is seldom
available in supercomputer. That is why, all specificity peculiar for the source
environment of AnyLogic version and standard Java environment, it shall be adapted
to supercomputer, and communication between computational nodes shall be based
on MPI standard.
270 V. L. Makarov and A. R. Bakhtizin

Already tested during paralleling of demographic model of Russia, the library for
multi-agent modeling ADEVS declared itself quite well. Last ADEVS versions have
some support of Java as well being an advantage too. However, ADEVS designers
have not implemented the parallel operation on supercomputer (except for OpenMP
technology for multiprocessors that required significant finalization in part of MPI
support in our previous work) up to now.
Also, during paralleling of a previous, quite simple model, it was rewritten using
C++ entirely, being superfluous: pre- and post-processing of data, and creation of
the initial status of multi-agent environment are not critical operations in terms of
time.
For a supercomputer, paralleling of an algorithm to compute kernel is usually
enough, i.e., in this case—population status recalculation phase.
Analysis of advanced software technologies showed that integrated tooling for
Java programs execution has been developed for the last time actively. It uses the
so-called AOT (Ahead-Of-Time) compilation.
At the same time, the result of AOT-compiler operation is a usual autonomous
executable module that contains a machine code for a target platform. Such
approach is used in new versions of Android operating system that, in our opinion,
is not accidental—both for embedded systems, and for a supercomputer the code
execution effectiveness is a key factor. Experiments with one such product—AOT-
compiler Avian—made it possible to draw the following conclusions:
• Avian allows obtaining an autonomous executable module in the form of MPI
application for a supercomputer, at the same time an undirected additional code
is implemented using C++ easily, including initialization and reference to MPI
communication library.
• Operating speed of the obtained software module (e.g., in classic game by
Conway “Game of Life”) approximately corresponds to the operating speed of
ADEVS.
• It allowed transferring a considerable part of work on AOT-compiler and to use
C++ for implementation of the most necessary one leaving for ADEVS a niche
of accelerated stages support with complicated interagent communication.

3.2.1 Technology of Parallel Version Obtaining and Launching

A source, developed in AnyLogic environment model description, represents an

XML file with extension .ALP (XML file in standard AnyLogic) that contains
description of agents’ essences of models, their parameters, and rules of agents’
status recalculation during evolution. Except for ALP description, the model is
attached with a data file in Excel format, where numeric parameters are specified
used both at the population formation stage and during recalculation of population
status.
The whole process of creation of a paralleled supercomputer version of the
program is performed in several stages listed below.
Supercomputer Technologies in Social Sciences 271

1. Input ALP file with model description is being read by converter that redesigns an
object representation for all described essences and rules and, having performed
their necessary processing, forms a package of software modules (Java class of
the total volume of several thousand lines) containing all significant information.
At the same time, agents’ variables are transformed into classes’ fields, and rules
of agents’ response to events—into the corresponding methods.
2. Input Excel file with source parameters of the model is converted into the source
Java text (also having the volume of several thousand lines) to provide for the
immediate access to model parameters at every node of a supercomputer. In
other words, external set of model parameters becomes a constituent part of
the executable module for a supercomputer. Formed software modules together
with a developed code of emulation of used AnyLogic environment functions are
compiled and configured into machine code for a target supercomputer.
3. Executable module, at the same time, is completely autonomous, and during
launching it takes several key parameters only in the command line. For example,
launch of the model with a million of agents for 20 years on 12 computational
nodes is executed by the following command:

$ mpirun -np 12 -x Population=1000000 rt.Main

4. Output data, at the same time, is collected from computational nodes in the
process of calculation, and population key characteristics and total calculation
time are preserved in the log and printed upon launch end:

[0]: totalNumberPeople = 990221

[0]: ***Total stages time: 21.173s***

3.2.2 Comparative Results of Models’ Operating Speed

For the first experiments with a parallel version of a simplified demographic model,
they used a fragment of the laboratory computational cluster with 12 computational
kernels and total volume of random access memory 96 GB. With such configuration,
RAM accommodates 60 million of agents easily that allows simulating population
growth dynamics within a small country.
A parallel version was tested on multicore processor as well. Results of mea-
surements of operating time of the model original and parallel version are shown in
Tables 2, 3, and 4.
As you can see in the above tables, operating speed and compatibility in the
number of agents of a supercomputer version is considerably higher than readings
of an original model. One can see as well that increasing the localing of interagent
communication (reducing quarters area), the calculation effectiveness increases.
272 V. L. Makarov and A. R. Bakhtizin

Table 2 Original version, 1 Agents Seconds

computational kernel (8 GB
RAM) 25,000 2.5
50,000 4.7
100,000 17.6
200,000 52.7
400,000 194
1,000,000 766

Table 3 Parallel version, 4 Kernels Quarters Agents Seconds

computational kernels
4 12 1,000,000 50
4 40 1,000,000 30
4 40 2,000,000 75
4 20 4,000,000 344

Table 4 Fragment of Kernels Quarters Agents Seconds

computational cluster (12
computational kernels Core 12 12 1,000,000 21
i7, 96 GB RAM) 12 24 1,000,000 12
12 24 10,000,000 516
12 60 10,000,000 303
12 300 10,000,000 132
12 300 30,000,000 585
12 300 50,000,000 1371
12 300 60,000,000 1833

3.2.3 Conclusions and Development Prospects

of the Technology Developed

The main positive moment of the developed approach to paralleling of models

designed in AnyLogic environment is automation of their supercomputer versions
creation. It simplifies the development considerably, because in most cases, after
insignificant modification of the source model, it does not require finalization of
rules of transformation into executable module for a supercomputer.
This approach is expendable in part of used source language and software–
hardware platform. Besides already successfully tested execution platforms Avian
and ADEVS, we can develop other, lower-level means for agents’ status recalcu-
lation state acceleration, and in prospect consider an issue of using such hardware
accelerators as Xeon Phi and NVidia CUDA.
The used technology of internodal communication by means of active com-
munication technology makes it possible, whenever required, to implement easily
both interactive simulation and interactive visualization of simulation process within
estimated time. However, it is possible only if a supercomputer is available in a burst
mode, for example, if a compact personal supercomputer is used.
Supercomputer Technologies in Social Sciences 273

The main research question remaining is a question of maximum achievable

effectiveness of paralleling in case of mass communication of agents being in
different computational nodes of a supercomputer.
It is quite evident that if every agent communicates actively with all other agents,
the productivity will be low in view of extensional internodal traffic.
Nevertheless, even in such unfavorable case, a supercomputer makes it possible
to accelerate simulation considerably, when a model is launched either repeatedly
(for statistics) or with different parameters. In other extreme case, when almost all
communications are localized in terms of geographic location of agents, effective-
ness of paralleling will be good.
Therefore, effectiveness of the parallel version depends directly on that part
of interagent communication that requires transfer of large data volume between
computational nodes.

Acknowledgements This work was supported by the Russian Science Foundation (grant  14-18-
01968).

References

1. Ambrosiano, N. (2006). Avian flu modeled on supercomputer. Los Alamos National Labora-
tory NewsLetter, 7(8), 32.
2. Bisset, K., Chen, J., Feng, X., Kumar, V. S. A., & Marathe, M. (2009). EpiFast: A fast algorithm
for large scale realistic epidemic simulations on distributed memory systems. In Proceedings
of 23rd ACM International Conference on Supercomputing (ICS’09), Yorktown Heights, New
York (pp. 430–439)
3. Collier, N. (2012). Repast HPC Manual. [Electronic resource] February 23. Access mode:
http://repast.sourceforge.net, free. Screen title. Language. English. (date of request: May 2013)
4. Deissenberg, C., Hoog, S., & van der Herbert, D. (2008, June 24). EURACE: A massively
parallel agent-based model of the European economy. Document de Travail No. 2008 (Vol.
39).
5. Epstein, J. M. (2009, August 6). Modeling to contain pandemics. Nature, 460, 687.
6. Epstein, J. M., & Axtell, R. L. (1996) Growing artificial societies: Social science from the
bottom up. Ch. V. Cambridge, MA: MIT Press.
7. Keith, R. B., Jiangzhuo, C., Xizhou, F., Anil Kumar, V. S., & Madhav, V. M. (2009, June 8–12).
EpiFast: A fast algorithm for large scale realistic epidemic simulations on distributed memory
systems. In ICS09 Proceedings of the 23rd international conference on supercomputing,
Yorktown Heights, New York (pp. 430–439)
8. Makarov, V. L., Bakhtizin, A. R., Vasenin, V. A., Roganov, V. A., & Trifonov, I. A. (2011).
Tools of supercomputer systems used to work with agent-based models. Software Engineering,
2(3), 2–14.
9. Parker, J. (2007). A flexible, large-scale, distributed agent based epidemic model. Center on
social and economic dynamics. Working Paper No. 52
10. Roberts, D. J., Simoni, D. A., & Eubank, S. (2007). A national scale microsimulation of disease
outbreaks. RTI International. Research Triangle Park, NC: Virginia Bioinformatics Institute.
Is Risk Quantifiable?

Sami Al-Suwailem, Francisco A. Doria, and Mahmoud Kamel

Abstract The work of Gödel and Turing, among others, shows that there are
fundamental limits to the possibility of formal quantification of natural and social
phenomena. Both our knowledge and our ignorance are, to a large extent, not
amenable to quantification. Disregard of these limits in the economic sphere might
lead to underestimation of risk and, consequently, to excessive risk-taking. If so, this
would expose markets to undue instability and turbulence. One major lesson of the
Global Financial Crisis, therefore, is to reform economic methodology to expand
beyond formal reasoning.

Keywords Financial instability · Gödel’s incompleteness theorem · Irreducible

uncertainty · Lucas critique · Mispricing risk · Quantifiability of risk ·
Reflexivity · Rice’s theorem · Self-reference

1 Introduction

It is dangerous to think of risk as a number

– William Sharpe [10]

Can we estimate uncertainty? Can we quantify risk? The late finance expert Peter
Bernstein wrote in the introduction to his book, Against the Gods: The Remarkable
Story of Risk [9, pp. 6–7]:

S. Al-Suwailem ()
Islamic Development Bank Group, Jeddah, Saudi Arabia
e-mail: sami@isdb.org
F. A. Doria
Advanced Studies Research Group, PEP/COPPE, Federal University at Rio de Janeiro, Rio de
Janeiro, Brazil
M. Kamel
College of Computer Science, King Abdul-Aziz University, Jeddah, Saudi Arabia

© Springer Nature Switzerland AG 2018 275

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0_14
276 S. Al-Suwailem et al.

The story that I have to tell is marked all the way through by a persistent tension between
those who assert that the best decisions are based on quantification and numbers, determined
by the patterns of the past, and those who base their decisions on more subjective degrees
of belief about the uncertain future. This is a controversy that has never been resolved . . .
The mathematically driven apparatus of modern risk management contains the seeds of
a dehumanizing and self-destructive technology. . . . Our lives teem with numbers, but we
sometimes forget that numbers are only tools. They have no soul; they may indeed become
fetishes.

We argue in this chapter that developments in science and mathematics in the

past century bring valuable insights into this controversy.
Philosopher and mathematician William Byers, in his book The Blind Spot:
Science and The Crisis of Uncertainty [25], makes an interesting case for why we
need to embrace uncertainty. He builds on discoveries in mathematics and science
in the last 100 years to argue that:
• Uncertainty is an inevitable fact of life. It is an irreducible feature of the universe.
• Uncertainty and incompleteness are the price we pay for creativity and freedom.
• No amount of scientific progress will succeed in removing uncertainty from the
world.
• Pretending that scientific progress will eliminate uncertainty is a pernicious
delusion that will have the paradoxical effect of hastening the advent of further
crises.
These aspects probably did not receive enough attention in mainstream literature,
despite the pioneering efforts of many economists (see, e.g., Velupillai et al. [111]).
We deal here with the question at the roots of uncertainty in the sciences which
use mathematics as their main tool. To be specific, by “risk” we mean indeterminacy
related to future economic loss or failure. By “quantifiability” we mean the ability
to model it using formal mathematical systems rich in arithmetic, in order to be able
to derive the quantities sought after. Of course, there are many ways to measure
and quantify historical risk.1 The question more precisely, therefore, is: Can we
systematically quantify uncertainty regarding future economic losses?
We argue that this is not possible.
This is not to say that we can never predict the future, or can never make
quantitative estimates of future uncertainty. What we argue is that, when predicting
the future, we will never be able to escape uncertainty in our predictions. More
important, we will never be able to quantify such uncertainty. We will see that not
even the strictest kind of mathematical rigor can evade uncertainty.

1.1 A Comment: Mainstream Economics and Risk Evaluation

There have been many studies criticizing the mathematization of economics and the
formal approach to economic analysis (e.g. Clower [30], and Blaug [14] and [15]).

1 But see Sect. 3 below.

Is Risk Quantifiable? 277

Many of the critiques of the use of mathematical modeling in economics rely on the
idea that models are approximations of reality. Models cannot capture all what we
know about the world. Economists and investors, therefore, should be less reliant on
models and be little more humble when making forecasts and pricing risks.
While this argument is generally valid, for some it might lead to the opposite
conclusion. The counterargument holds that approximation can always be improved.
There will always be new methods and innovative techniques that help us get more
accurate results, even if we cannot get the exact ones. If we work harder and be little
smarter, we will come closer and closer to the “ultimate truth”, even if we will never
be able to reach it.
So the “approximation argument” would likely lead to the opposite of what was
intended. Rather than discouraging over-reliance on formal models, it encourages
more sophisticated techniques, and larger reliance on computing power.
In this chapter, however, we take a different approach. We examine ideal
conditions for employing formal models and supercomputers, and see to what
extent they can help us predict the future. As we shall see, ideal models and
supercomputers are fundamentally limited in manners that will not be even possibly
approximated. This fundamental uncertainty calls for a radically different approach
for studying and analysing markets and economic phenomena.
The point is: we envisage “predictability” as “computability”, “Turing com-
putability”.2
Ignoring these limitations leads to systematic mispricing of risk, which in
turn encourages unwarranted risk taking. Markets therefore become less stable,
and the economy becomes vulnerable to booms and crashes. Blind faith in risk-
quantification leads to more, not less, risks.

1.2 A Summary of Our Work

We may summarize the gist of this chapter as follows: we want to look at risk,
and risk-evaluation, from the viewpoint of Gödel’s incompleteness phenomenon.
Therefore the next sections of the chapter deal with Gödel incompleteness, and its
rather unexpected consequences for economics. We then apply the concepts we have
introduced to formal models in economics, and conclude that Gödel incompleteness
appears everywhere (and in crucial situations) in mathematical economics. Here, we
arrive at our main question, the quantitative evaluation of risk.
Our discussion is, admittedly, intuitive and nontechnical, as we deliberately
sacrifice rigor for understanding.

2 We will later elaborate on that characterization.

278 S. Al-Suwailem et al.

2 Formal Models

A more detailed presentation of these ideas can be found in Chaitin et al. [28]. Let
us start with the use of formal, axiomatic, mathematical models. Suppose we have
an ideal economic world W , that we would like to investigate.3 The economy W is
characterized by the following properties:
• It is deterministic. No randomness whatsoever exists in this economy.
• It is fully observable. Observation captures the true state of the economy at the
current period t. There is no gap between observation and reality at t.
Now suppose we are able to build a formal mathematical model M of the
economy W , that has the following ideal characteristics:
• M is internally consistent. No two contradictory M-statements can be proved
within M.
• M accurately captures all the knowledge that we have accumulated so far about
W . Any valid statement within M corresponds to a true state of the economy W .
The mathematical model M is therefore sound, i.e. it proves only true statements
about the economy W .
• All statements of M are verifiable. There is no ambiguous or indeterminate M
statement. Any M statement about the state of the economy W at a certain time
t can be verified by directly observing W at t .
• M lies in a language with enough arithmetic. That is both a technical condition
and a pragmatic one. A pragmatic one: we must calculate if we are to quantify
aspects of the economy, and in order to quantify we require arithmetic. A
technical condition: “enough” arithmetic is required for incompleteness, as will
be discussed shortly.
Given the ideal model of the ideal economy, we ask the following question: can
M systematically predict the future states of W ?
We will embed our M in a theory S that has the following characteristics:
• The language of S is the first order classical predicate calculus.
• S has a set of theorems that can be enumerated by a computer program.4
• P A, Peano Arithmetic, is a subtheory of S and describes its arithmetic portion;
M contains the arithmetic portion of S.5

3 We use here “model” in the sense of “mathematical model”, that is, the mathematical depiction
of some phenomenon; we do not use it in the model-theoretic sense.
4 That is, the set of theorems of S is recursively enumerable.
5 This is what we mean by “enough arithmetic”. Actually we can obtain the same results we require

if we add a still weaker arithmetic condition.

Is Risk Quantifiable? 279

2.1 Risk

We informally state a result,6 later to be formulated in a more rigorous way:

M cannot systematically predict (calculate) all future states of an (ideal) economy W .

The proof of this theorem follows from Kurt Gödel’s famous (first) Incomplete-
ness Theorem, published in 1931. In simple words, Gödel’s result states (waving
hands) that any formal system, rich enough to include arithmetics, cannot be both
consistent and complete. To be consistent means there will not be a statement
derived from the system that contradicts any of its statements. To be complete means
that all true interpretations of formal statements in the system are either axioms or
can be derived from the axioms.
Since we have already assumed that model M is consistent, it follows that
M cannot be complete, provided that our theory includes enough arithmetics in
its framework. Since we are investigating “quantification” of the economy, this
condition is readily met. Consequently, there are true M statements about W
that cannot be proved within M. These statements are described as “undecidable
propositions” [37, p. 119]. We call such statements “Gödel sentences”.
One important example of such Gödel sentences is the one that states: “M is
consistent”, i.e., is free from contradictions. As Gödel [51] has shown, this sentence
cannot be proved from within M [42, p. 154].
From the economic point of view, there are many important questions that are
undecidable, including:
• Will the economy be stable in the future? Or will it be chaotic?
• Will markets clear in the future? Would the economy attain equilibrium?
• Will resources be allocated efficiently?
Moreover, if we define:
A risky set is a subset R ⊂ W that satisfies some property PR that represents our
conceptions of risk,

we can ask
• Will the risky set R be decidable?
(A set A ⊂ B is decidable if and only if there is a computer program MA so that
MA (x) = 1 whenever x ∈ A, and 0 otherwise. If there is no such a program, then A
is undecidable.)

The four questions stated above have been shown to be undecidable (see da Costa
and Doria [36] and the references therein). That is, there is no systematic way to
provide a mathematical answer, either in the affirmative or in the negative, to any

6 The theorem that implies such a result is a Rice-like theorem proved by da Costa and Doria in
theorem 1990; see Chaitin et al. [28].
280 S. Al-Suwailem et al.

of these questions. We just don’t know for sure. In fact, any interesting feature of
dynamical systems is undecidable [100].
If we cannot derive or prove Gödel sentences of M, this means we cannot predict
the corresponding states of the economy W . Consequently, the model cannot fully
predict all future states of the economy. This implies that the economy might turn
out to be at a particular state, call it Gödel-event, that is impossible to predict in the
model M.
We can say that this is an indication of the complexity of our theoretical
constructs, where events can occur that cannot be derived mathematically within
M [76, p. 207].
We can—sort of—say that Gödel’s Theorem establishes that mathematics is
inexhaustible [46]. The theorem further shows how a well-structured formal system
can give rise to unexpected and unintended results. As Dyson [39, p. 14] writes,
“Gödel proved that in mathematics, the whole is always greater than the sum of the
parts”.
According to Gödel (1995, p. 309), the theorem:
. . . makes it impossible that someone should set up a certain well-defined system of axioms
and rules and consistently make the following assertion about it: all of these axioms and
rules I perceive (with mathematical certitude) to be correct, and moreover I believe that
they contain all of mathematics. If someone makes such a statement he contradicts himself.

To prove that a particular proposition is mathematically undecidable is not very

easy. It takes a lot of hard work and ingenuity. So, can we assume that a given
mathematical statement is provable or decidable in principle, unless it is proved
to be otherwise? Put differently, can we follow a rule of thumb: all propositions are
decidable unless proved otherwise? Unfortunately, we cannot do so, for two reasons.
One is that undecidability seems to be the rule rather than the exception, as we shall
see later. Second, there is no finite effective procedure that can decide whether a
given statement is provable, in M or in any formal system [12, pp. 107–109]. That
is, we cannot make this assumption because this assumption itself is undecidable.
Undecidability, therefore, is undecidable!

3 Technicalities Galore

Undecidability, in fact, is everywhere. And undecidability and incompleteness are

related: the algorithmical unsolvability of the halting problem for Turing machines7
means that there is no computer program that decides whether the universal Turing
machine stops over a given input. Then, if we formulate Turing machine theory
within our axiomatic theory S, there is a sentence in that theory—“There is a x0
so that the universal Turing machine never stops over it as an input”—which can
neither be proved nor negated within S. For if we proved it, then we would be able

7A simple proof can be found in Chaitin et al. [28].

Is Risk Quantifiable? 281

to concoct a computer program that computed all nonstopping inputs to the universal
machine.
We have a more general result proved by da Costa and one of us (Doria) in 1990
[34]. Still sort of waving hands, let us be given M, and let P be a predicate (roughly,
the formal version of a property) on the objects of M. Let x1 = x2 be terms in M,
that is, the description of objects in M. P (x) can be read as “x has property P ”.
¬P (x), “x doesn’t have P ”. Then:
Given any P so that, for x1 = x2 , P (x1 ) and ¬P (x2 ) hold, then there is a term x3 so that
P (x3 ) is undecidable.

That is, no algorithm can decide whether either P (x3 ) holds or ¬P (x3 ) holds.
(For a complete discussion, which includes the original reference to this result, see
da Costa and Doria [36]).
This also shows that the whole is greater than the sum of the parts. We start with
two decidable properties, and end up with a third that is undecidable.
Actually, any scientific discipline that formulates its theories using mathematics
will face the problem of undecidability, if it is meant to compute values or produce
quantities, that is, if it contains arithmetics.
With more detail: suppose that we know some system S to be at state σ0 at time
t0 . A main task of a scientific discipline is to find some rule ρ to predict the future
state of system ρ(σ0 , t) for t > t0 .
Then, according to the result just quoted, even if we know the exact evolution
function of the system, namely rule ρ, and even if we have absolutely precise
data about the current state of the system, σ0 , this will not be enough for us to
systematically predict an arbitrary future state ρ(σ0 , t) of the system. This is in fact
a quite unexpected result. (A rigorous statement of the result will soon be given.)
No matter how powerful our mathematical model is, no matter how precise our
measurement instruments are, and no matter how comprehensive our data are, we
will not be able to systematically predict the future of any natural or social system.
Although we may be able to predict the future in many cases, we will never be able
to do so for all cases, not even on average.
Wolpert [115] extends the above results to include, in addition to prediction,
both observation and retrieval of past measurements. Wolpert argues that physical
devices that perform observation, prediction, or recollection share an underlying
mathematical structure, and thus call them “inference devices”. Inference devices
therefore can perform computations to measure data, predict or retrieve past
measurements. Based on the work of Gödel and Turing, Wolpert then shows that
there are fundamental limits on inference devices within the universe, such that
these three functions cannot be performed systematically, even in a classical, non-
chaotic world. The results extend to probabilistic inference as well. These limits are
purely logical and have nothing to do with technology or resources. The results are
totally independent of both the details of the laws of physics and the computational
characteristics of the machines. According to Wolpert, these impossibility results
can be viewed as a non-quantum mechanical “uncertainty principle”.
282 S. Al-Suwailem et al.

Hence, not only there are fundamental limits on quantifying the future, there are
also fundamental limits on quantifying the observed reality, and recalling the past
quantities.

3.1 Island of Knowledge

Mathematical models help us understand and quantify aspects of the universe

around us. Scientists clearly have been able to quantify many properties of natural
phenomena. However, Gödel’s theorem shows that there is a price that we have to
pay to get these benefits. While a formal model helps us answer specific questions
about the universe, the same model raises more questions and opens the door
for more puzzles and uncertainties. Moreover, with different models for different
phenomena, we face even more puzzles and uncertainties on how these models link
together. Quantum mechanics and general relativity are obvious examples.
So while formal models help us gain knowledge, they make us aware of how
ignorant we are. As John Wheeler points out: “As the island of our knowledge
grows, so do the shores of our ignorance”. William Dampier writes: “There seems
no limit to research, for as been truly said, the more the sphere of knowledge grows,
the larger becomes the surface of contact with the unknown” (cited in Gleiser [50,
p. 288]). Scientist and mathematicians are well aware of this price, and they are more
than happy to pay. There will be no end to scientific endeavour and exploration.
Popper [82, p. 162] argues that all explanatory science is incompletable; for to be
complete it would have to give an explanatory account of itself, which is impossible.
Accordingly:
We live in a world of emergent evolution; of problems whose solutions, if they are solved,
beget new and deeper problems. Thus we live in a universe of emergent novelty; a novelty
which, as a rule, is not completely reducible to any of the preceding stages.

Hawking [57] builds on Gödel’s Theorem to dismiss a “Theory of Everything”.

He writes:
Some people will be very disappointed if there is not an ultimate theory that can be
formulated as a finite number of principles. I used to belong to that camp, but I have changed
my mind. I’m now glad that our search for understanding will never come to an end, and that
we will always have the challenge of new discovery. Without it, we would stagnate. Gödel’s
theorem ensured there would always be a job for mathematicians. I think M theory [in
quantum mechanics] will do the same for physicists. I’m sure Dirac would have approved.

Mainstream economics, in contrast, pays little attention, if any, to these limita-

tions. For mainstream theoreticians, the Arrow–Debreu model [3] depicts essentially
a complete picture of the economy. We expect no place in this picture for
incompleteness or undecidabilities whatsoever. It is the “Theory of Everything”
Is Risk Quantifiable? 283

in the mainstream economics.8 Limitations arise only in the real world due to
transaction costs and other frictions. To this day, this model is the benchmark
of Neoclassical economic theory. There is therefore not much to explore or to
discover in such a world. In fact, entrepreneurship, creativity and innovation, the
engine of economic growth, have no place in Neoclassical theory. Not surprisingly,
mainstream economic models have a poor record of predicting major events, not the
least of which is the Global Financial Crisis (see, e.g., Colander et al. [31]).

3.2 A Brief Side Comment: Black Swans?

A Gödel sentence describes a Gödel event: a state of the economy that cannot be
predicted by M. Nonetheless, after being realized, the state is fully consistent with
our stock of knowledge and therefore with M.
This might sound similar to Nassim Taleb’s idea of “Black Swan”. A Black Swan
is an event that:
• Lies outside regular expectations;
• Has an extreme impact, and
• Is retrospectively, but not prospectively, explainable and predictable.
(See Taleb [104, pp. xvii–xviii]). If “retrospective but not prospective” pre-
dictability is broadly understood to mean consistency with M but not provability in
it, then a Black Swan event might be viewed as an instance of a Gödel statement
about the world. However, if a Black Swan event is a “fat tail event”,9 then it
is still predictable. In contrast, an event corresponding to a Gödel sentence is
fundamentally unpredictable.
This suggests the following characterization:
An event is predictable if there is a finite set of inputs and a computer program that computes
it.

We thus equate “prediction” to “effective procedure”.

3.3 More Side Comments: Ambiguity

We have assumed that each statement of our M model corresponds unambiguously

to a certain state of the economy. But this is not true. When it comes to interpretation,
mathematical concepts are “relative” [7]. They can be interpreted to mean different
things, and these different interpretations are not always isomorphic, so to say.

8 Yet the Arrow–Debreu theory is undecidable; that follows from a theorem by Tsuji et al. [107].
9 For a summary discussion, see http://http://www.fattails.ca.
284 S. Al-Suwailem et al.

A formal system may have unintended interpretations despite the system being
intended to have a particular interpretation [38, pp. 180–187]. This may be seen
as another example of an emergent phenomenon where the whole goes beyond the
parts.
According to DeLong [38, p. 185], Gödel’s incompleteness theorem and the
work of the Thoralf Skolem (see Gray [55]) show that there exists an infinite
number of interpretations (or models, now in the model–theoretic sense) of any
formal system. In general, no formal system is necessarily categorical, i.e., has
only isomorphic models (in the model–theoretic sense). Hence, mathematical reality
cannot be unambiguously incorporated in axiomatic systems [62, pp. 271–272].
This means that even if an M-statement is predictable (provable), we cannot
tell exactly what would be the corresponding interpretation, seen as a state of the
economy. Unintended interpretations would arise beyond the intended setup of the
model. This ambiguity adds another layer of uncertainty that defeats the best efforts
to quantify our knowledge.

4 Technicalities, II

But there is still more to come: how many Gödel sentences can there be? According
to Gödel [52, p. 272], there are “infinitely many undecidable propositions in any
such axiomatic system”. This is obvious: they must be infinite; if they were finite in
number, or could be generated by a finite procedure, we would just add them to our
axiom system.
Calude et al. [26] prove this result formally, and show that the set of Gödel
sentences is very large, and that provable (i.e. predictable) statements are actually
rare (see also Yanofsky [117, p. 330]). Almost all true sentences are undecidable
[103, p. 17]. Almost all languages are undecidable (Lewis [69, p. 1]; Erickson [40,
pp. 5, 7]). Most numbers are “unknowable”, i.e. there is no way to exhibit them
([24], pp. 170–171).
We give here a brief discussion, based on an argument of A. Bovykin.10
Besides Calude’s paper there are several folklore-like arguments that show that
undecidability is the rule, in formal systems. One of them is based on Chaitin’s
 number, and the fact that it is a normal number.11 Briefly, one concludes that
while the set of well-formed formulae (wff ) of a theory with associated halting
probability  grows as x, the set of provable wff grows as log x. (There is an
intriguing and still unexplored connection with the Prime Number Theorem here.)

10 Personal communication to FAD.

11 In binary form, zeros and ones are evenly distributed.
Is Risk Quantifiable? 285

Yet another quite simple argument that leads to the same conclusion is due to A.
Bovykin. It applies to the theories we are dealing with here:
• We deal with theories that have a recursively enumerable set of theorems. The
sentences in such theories are supposed to be a recursive subset of the set of all
words in the theories’ alphabet, and therefore they can be coded by a 1–1 map
onto ω, the set of all natural numbers.
• The set of theorems of one such theory, say T , is therefore coded by a recursively
enumerable subset of ω.
As a consequence there is a Turing machine MT that lists all theorems of T . Or,
a machine MT which proves all theorems of T (each calculation is a proof).
• The calculation/proof is a sequence of machine configurations.
• Now consider all feasible, even if “illegal,” machine configurations at each step
in a calculation. Their number is, say, n, with n ≥ 2. For a k-step computation
there will be nk possible such legal and illegal sequences, of which only one will
be a computation of MT .
• That is to say, just one in nk possible sequences of machine configurations, or
proof sequences.
This clarifies why among the set of all such strings the set of all proofs is very
small. The argument is sketchy, but its rigorous version can be easily recovered out
of the sketchy presentation above.
In general, the describable and decidable is countably infinite, but the inde-
scribable and undecidable may even be uncountably infinite. And the uncountably
infinite is much, much larger than the countably infinite [117, pp. 80, 345].

4.1 Is There a Limit to Uncertainty?

Quantifying a phenomenon requires a mathematical model of it based on available

information. Since the ideal model M by assumption captures all the knowledge—
available to us—about the world, it follows that the uncertainty that arises from M
cannot be formally modeled, and therefore cannot be quantified.
Suppose we create a “meta model” M2 that somehow quantifies Gödel sentences
arising from the model M. Then the new model will have its own Gödel sentences
which cannot be quantified neither in M nor M2 . If we create yet another meta–meta
model, say M3 , then another set of Gödel sentences will arise, ad infinitum.
For every formal model, there will be (many, many) undecidable propositions and
unprovable statements. If we try to model these statements using a more powerful
formal model, many more undecidable statements will arise. With each attempt to
quantify uncertainty, higher uncertainty inevitably emerges. We will never be able
to tame such uncertainty. Uncertainty is truly fundamental, and it breaches the outer
limits of reason. It follows that uncertainty is beyond systematic quantification.
A frequently cited distinction between “risk” and “uncertainty” is based, mainly,
on Knight [63]. The distinction holds that “risk” is quantifiable indeterminacy, while
286 S. Al-Suwailem et al.

“uncertainty” is not quantifiable. From the previous discussion, this distinction

appears to be less concrete. Since “quantification” must be based on formal
mathematical modeling, then uncertainty, as reflected in infinite undecidable Gödel
statements that creep up from the model, inevitably arises. This “irreducible
uncertainty” evades quantification by all means.

4.2 Rationality

So far we have been discussing the economy at a macro level. Let us now zoom in
to economic agents and see if they also face uncertainty.
One of the most controversial assumptions about economic agents in mainstream
neoclassical economics is the assumption of rationality. There has been, and still
ongoing, a long debate on this assumption, but we shall not address it here; we will
simply state a few conditions and ask our agents to satisfy it. We then proceed to
see what, if any, kind of uncertainty agents face. In particular, we assume that ideal
agents R in the economy W enjoy the following ideal characteristics:
• Agents are well organized and very systematic in planning and executing their
plans. They are as systematic as digital computers. In other words, an agent is a
“Turing machine”, i.e. an ideal computer that is capable of performing any task
that can be structured in a clear sequence of steps, i.e. every economic task is
computable.
• Agents have (a potentially) unlimited memory and ultra-fast computing power.
They face no resource constraints on their computing capabilities.
(That identification can be taken as a definition of rationality, if needed. Putnam
[85] and Velupillai [110] show the formal equivalence of economic rationality to
Turing machines.)
• Agents act deterministically, and are not subject to random errors.
• Agents have full information about the current and past states of the economy W .
• Agents are honest and truthful (this is to avoid artificial uncertainty due to
deliberate misrepresentation).
To be practical, an agent would decide his economic choice based on an ideal
plan P : a set of instructions of how to make an economic choice. For example, a
consumption plan would have a clear set of instructions on how to screen goods
and services, how to evaluate each, how to choose, over how many periods, etc.
The same applies to production and investment. Agents formulate plans identical
to software programs that specify each step needed to perform consumption,
production and investment.
R agents need not be utility or profit maximizers. They need not be optimizers of
any sort. But they are well organized and systematic that they formulate their plans
in a clear sequence of instructions, just as a computer program.
Now we ask the following question: would R agents face any uncertainty in their
course of economic activity?
Is Risk Quantifiable? 287

We informally state another result, whose rigorous formulation will be given

below:
Ideal agents R cannot systematically predict the outcome of their ideal plans P .

The proof of this theorem follows quite trivially from Alan Turing’s ((year?))
famous theorem on the “halting problem”. In words, it says that there is no program
(or a Turing machine) that can always decide whether a certain program P r will
ever halt or not. A program halts when it completes executing its instructions, given
a certain input. It will fail to halt if it is stuck in an infinite loop where it fails to
complete the execution of its instructions.
A generalization of Turing’s theorem is provided by Rice [88]. Rice’s Theorem
states that there is no program that can decide systematically whether a given
program P r has a certain non-trivial property y. A non-trivial property is a property
that may be satisfied by some programs but not by others.
A good example is found in computer viruses, which are softwares that damage
the performance of the operating system. Is there a program that can always decide
whether a certain program is (or has) a virus or not? In other words, could there be
an “ideal” anti-virus program?
According to Turing and Rice, the answer is no [28]. There is no ideal anti-virus
program. Nor there is an ideal debugger program. The reason is intuitively simple.
Suppose there is a program that is believed to be an ideal anti-virus. How to know
that it is in fact the ultimate anti-virus? To verify this property, we need to design
another program to test our “ideal anti-virus”. Then the same problem arises for the
testing program, ad infinitum. We will never be able to be fully sure if our software
can catch all viruses, or spot all bugs. Risk is inevitable. The claim of inventing the
“ultimate anti-virus” is no less fraudulent than that of a “perpetual motion machine”.
For our economic agents, this means that agents cannot always know in advance
what will be the full output of ideal plans P .
For example, for a given production or investment plan, we ask the following
question: Is there a general procedure that can tell us whether the plan has no
“bugs” that could lead to loss or adverse outcomes? Rice’s Theorem says no. There
is no general procedure or program that can perform this function. Accordingly, an
investor contemplating to invest in a given business opportunity is unable to know
for sure what would be the output of the business plan, even in a fully deterministic
world, free from cheating and lying, and where agents have full information and
enjoy super-computing capabilities. Agents can never be sure that they will not lose
money or face inadvertent outcomes. Risk is unavoidable.
Again, this risk is not quantifiable. The argument is the same as before. If we were
able to design a program to compute the risks associated with a given investment or
production plan, how can we verify that the program has no “bug”? We will never
be able systematically to do so.
(More on Rice’s Theorem at the end of this section.)
288 S. Al-Suwailem et al.

4.3 Is Knowledge Quantifiable?

We arrive here at an important question: can we quantify our knowledge? Gödel’s

theorem makes it clear that there are statements that are true but not provable in
prescribed formal systems, and that we cannot simply add those statements to our
axioms, as new undecidable statements immediately creep up. If the formal system
is consistent, then we know that the statement is true, but it is not provable. As
Franzen [46, p. 240] points out, Gödel’s theorem “is not an obstacle to knowledge,
but an obstacle to the formalization of knowledge”. Does that mean that truth
extends beyond deductive logic? If so, it is not quantifiable.
Curiously enough, cognitive scientists and neuroscientists agree that “we know
more than we can say” [54, p. 99]. Human beings are equipped with sensors that are
independent of verbal or even visual communications. The flow of such signals is
a transfer of information via non-expressible means. Detecting and analysing these
signals is what amounts to “social intelligence”. The human brain has regions for
handling social intelligence that are different from those for handling deductive and
logical reasoning [70].
The knowledge accumulated via non-deductive, informal, means is what is called
“tacit knowledge”, as Polanyi [81] describes it (from Latin tacere, to shut up).
Knowing how to ride a bicycle or to drive a car, for example, is tacit knowledge that
cannot be fully expressed or articulated. Much of our valuable knowledge that guide
our daily and social activities are of this sort of inarticulate or irreducible knowledge
(see Lavoie [67]; and Collins [33]). According to Mintzberg [77, pp. 258–260], “soft
information”, i.e. information not communicated through formal means, account for
70–80% of information used for decision making.
As already pointed out, what we can describe and formalize is countably infinite,
but the indescribable may even be uncountably infinite, which is much, much larger
than the countably infinite.
This implies that the problem of informational asymmetry is particularly relevant
only for formal or expressible knowledge. However, if tacit knowledge is taken
into account, the problem could be reduced to a greater extent, although it will
never disappear. Main stream economics pays little attention to non-formal or
tacit knowledge. This “epistemic deficit”, as Velupillai [112, p. xix] describes it,
only reinforces the problem of informational asymmetry. Formal requirements of
disclosure, for example, will help only up to a point. Beyond that point, we need to
enhance social relations to improve social intelligence and communication of tacit
knowledge.

4.4 Back to Our Goal: Technicalities Galore; Rice’s Theorem

and the Rice–da Costa–Doria Result

It is now the moment to go technical again, and add some more mathematical rigor
to our presentation.
Is Risk Quantifiable? 289

We sketch here a very brief primer on Rice’s Theorem.

Suppose that the theory S we deal with has also the ι symbol (briefly, ιx Q(x)
reads as “the object x so that Q(x) is true”). Let P be a predicate symbol so that
there are two terms ξ = ζ and S  P (ξ ) and S  ¬P (ζ ) (that is, P (ξ ) and
not-P (ζ ) are provable in S; we call such P , nontrivial predicates). Then, for the
term:

η = ιx [(x = ξ ∧ α) ∨ (x = ζ ∧ ¬α)],

where α is an undecidable sentence in S:

Proposition 1 S  P (η) and S  ¬P (η).

This shows that incompleteness is found everywhere within theories like S,

that is, which include PA, have as its underlying language the first order classical
predicate calculus, and have a recursively enumerable set of theorems.
Notice that this implies Rice’s Theorem [73, 89] in computer science. Briefly:
suppose that there is an algorithm that settles P (n), for each n ∈ ω. Then by the
representation theorem we may internalize that algorithm into S, and obtain a proof
of P (n) for arbitrary n ∈ ω, a contradiction given Proposition 1.
In 1990 (published in 1991), da Costa and Doria extended Rice’s Theorem
to encompass any theory that includes “enough arithmetic” (let’s leave it vague;
we must only be able to make the usual arithmetical computations) and which is
consistent, has a recursively enumerable set of theorems and a model (in the sense
of interpretation) with standard arithmetic. Let us add predicate symbols to S, if
needed (we will then deal with a conservative extension of S). Then,

Proposition 2 If P is nontrivial, then P is undecidable, and there is a term ζ so

that S  P (ζ ) and S  ¬P (ζ ).

So each non-trivial P leads to an undecidable sentence P (ζ ).

Such results explain why undecidability and incompleteness are everywhere to
be found in theories like S, which are the usual kind we deal with. Moreover we can
show that P (ζ ) can be as “hard” as we wish within the arithmetical hierarchy [89].
It is as if mathematical questions were very very difficult, but for a small set of
sentences. . . Of course this nasty situation spills out to applied sciences that use
mathematics as a main tool [35].

4.5 Know Thyself!

Popper [82, pp. 68–77] was among the early writers who considered the fact that a
Turing machine faces an inherent prediction problem of its own calculations. Popper
argues that, even if we have a perfect machine of the famous Laplace’s demon type
[66], as long as this machine takes time to make predictions, it will be impossible for
290 S. Al-Suwailem et al.

the machine to predict its own prediction. That is, the outcome of its calculation will
arrive either at the same time as its primary prediction, or most likely, afterwards.
Thus, a predictor cannot predict the future growth of its own knowledge (p. 71).
Put differently, a Turing machine, i.e. an ideal computer, cannot systematically
predict the result of its own computation [72]. This can be seen as an innocent
extension of Rice’s Theorem above, but it sheds some additional light here. An
ideal agent cannot systematically predict what will be the basket of goods and
services that he or she will choose given its own consumption plan. Similarly,
an agent cannot predict the inputs that will be used for production, or the asset
portfolio that he or she will invested in. Furthermore, suppose agents choose their
consumption or investment plans from a large set of predetermined plans. Could an
agent systematically predict what plan it will choose? The answer, again, is no, for
the same reasons.
Can an ideal agent predict the choice of other agents?
An ideal agent cannot systematically predict his or her choice, nor the choice of any other
ideal agent.

Since an ideal agent is an ideal computer, it will be able to simulate the decision
process of any other ideal agent, assuming the decision process or plan is explicitly
provided. So, if an agent cannot systematically predict his, or her, own choice, it
follows that an agent cannot systematically predict the choice of any other ideal
agent. Since all agents are ideal computers, then no one will be able to systematically
predict the choice of any other. As Berto [12, p. 188] points out, “If we are indeed
just Turing machines, then we cannot know exactly which Turing machines we are”.
Paul Benacerraf, moreover, writes: “If am a Turing machine, then I am barred by
my very nature from obeying Socrates’ profound philosophical injunction: Know
thyself!” (cited in Berto [12]).12
Surprisingly, this uncertainty arises exactly because agents are assumed to
possess maximal rationality. As Lloyd [71, p. 36] rightly points out:
Ironically, it is customary to assign our own unpredictable behavior and that of others to
irrationality: were we to behave rationally, we reason, the world would be more predictable.
In fact, it is just when we behave rationally, moving logically like a computer from step
to step, that our behavior becomes provably unpredictable. Rationality combines with the
capacity of self reference to make our actions intrinsically paradoxical and uncertain.

5 Applications, I: Self-reference

Gödel’s original example of an undecidable statement arose out of self-reference.

Both Gödel and Turing rely on self-reference to prove their theorems [19, 20]. (Even
if self-reference isn’t essential to prove incompleteness [28], it remains an important
tool in the construction of Gödel-like examples.)

12 Actually an injunction engraved at the Apollonic oracle at Delphos.

Is Risk Quantifiable? 291

However, self-reference as such will not always lead to nontrivial results.

The statement “This statement is true” does not lead to a paradox, while “This
statement is false” does.13 More important, many concepts and definitions essential
to mathematics involve self-reference (like the “maximum” and the “least upper
bound”; DeLong [38, p. 84]).
Self-reference is valuable because it allows language to extend beyond the limits
of its constituents. It permits the whole to exceed the parts. In physical systems,
positive or reinforcing feed-back, which is a form of self-reference, is necessary for
emergence and self-organization. In social systems, self-reference allows (social)
institutions to be created, which in turn help reduce uncertainty and coordinate
expectations (see below). It is also required for “shared understanding”, including
mutual belief, common knowledge, and public information, all of which rest on
self-reference [5, 6].
Self-reference arises from meta-statements: statements about themselves. What
Gödel ingeniously discovered was that formal models can be reflected inside
themselves [19]. A formal model of consumer’s choice, therefore, will have
a meta-statement about its own statements. Accordingly, the model describing
consumer’s preferences will have meta-preferences, specifying the desirability of
such preferences [114]. For example, one might have a preference for smoking, but
he or she might not like having this particular preference.
Meta-preferences play the role of values and morality that guide consumer’s
choice [41, p. 41]. Neoclassical economics claims to be “value-neutral”, and
therefore ignores morality and ethics. But, in the light of Gödel’s theorem, this claim
is questionable. Meta-statements arise naturally in any (sufficiently rich) formal
model, and these statements have the properties of values and morality for models
of choice. Simply to ignore meta-statements does not deny their existence and their
impact.

5.1 . . . and Vicious Self-reference

For self-reference to be “harmful” or vicious, it needs to be “complete” [102,

p. 111]. This arises when we try to encapsulate the undescribable within the
describable, which is much larger as pointed out earlier. This will necessarily
involve contradictions, i.e. the description will be inconsistent. Then to avoid
contradiction, the description must be incomplete.
Another way to look at it is that vicious self-reference, as in the Liar paradox
for example, has no stable truth value: from one perspective it is true, from the
other it is false. We can therefore say that paradoxes are sort of “logically unstable”
assertions (see Kremer [64]; Bolander [20]). Vicious self-reference thus creates an

13 Yetone must be careful here, as the so-called “reflection principles” may be interpreted as
sentences of the form “X is true”.
292 S. Al-Suwailem et al.

infinite loop which, if implemented on a physical machine, will exhaust its resources
and results in a crash (see Stuart [101, p. 263]). Normal self-reference, in contrast,
is a commonplace part of any well-designed program.
Vicious self-reference may occur in various areas of market activity. Financial
markets are susceptible to harmful self-reference when they are dominated by
speculation. As Keynes [59, p. 156] notes, speculators devote their intelligence to
“anticipating what average opinion expects the average opinion to be”. It is a “battle
of wits”, and the objective is to “outwit the crowd”. Speculators “spent their time
chasing one another’s tails”, as Krugman [65] remarks. Simon [94, 95] considers
the possibility of “outguessing” among economic agents to be “the permanent and
ineradicable scandal of economic theory”. He points out that “the whole concept of
rationality became irremediably ill-defined when the possibility of outguessing was
introduced”, and that a different framework and methodology must be adopted to
understand economic behaviour in such conditions (cited in Rubinstein [91, p. 188]).
Soros [96, 97] develops a theory of “reflexivity”, which echoes to some extent
the concerns of Simon above. Reflexivity implies that agents’ views influence the
course of events, and the course of events influences the agents’ views. The influence
is continuous and circular; that is what turns it into a feedback loop. But positive
feedback loops cannot go on forever. As divergence between perceptions and reality
widens, this leads to a climax which sets in motion a positive feedback process in
the opposite direction. “Such initially self-reinforcing but eventually self-defeating
boom-bust processes or bubbles are characteristic of financial markets” [97].
Reflexivity therefore undermines agents’ perception and understanding of market
dynamics, and at the same time ensures that agents’ actions lead to unintended
consequences [96, p. 2]. “Financial markets are inherently unstable” [96, pp. 14,
333].
Economists are aware of how agents’ beliefs and behaviour actually changes the
course of events. The “Lucas Critique” implies that a model may not be stable if
it is used to recommend actions or policies that are not accounted for in the model
itself (see Savin and Whitman [92]). This is also referred to as “Goodhart’s Law”
[29]. Lucas’ critique therefore was a strong critique to models for policy-making
purposes. But there is no reason why the same critique would not apply to models
of financial markets. In fact, the Critique is even more relevant to financial markets
than policy making given the intensity of the “guessing game” played therein. This
became apparent in the years leading to the Global Financial Crisis. Rajan et al.
[86] point out that models used to predict risk and probability of defaults are subject
to Lucas Critique: Given the estimated probability of default, lenders will take on
additional risks, rendering the initial estimates invalid. The authors rightly title their
paper: “Failure of models that predict failures”.
The problem of vicious self-reference therefore makes models unstable and
lead to false predictions. The failure of such models to recognize the possibility
of vicious self-reference contributed to additional risk-taking and consequently,
market instability. Physicist and fund manager Bouchaud [22, p. 1181] points out:
“Ironically, it was the very use of a crash-free model that helped to trigger a crash”.
But, how does a vicious circle arise?
Is Risk Quantifiable? 293

There have been many attempts to characterize the nature of paradoxes and
vicious self-reference [20, 83, 84, 116].
Bertrand Russell provided a general characterization for paradoxes: “In each
contradiction something is said about all cases of some kind, and from what is said
a new case seems to be generated, which both is and is not of the same kind. . . ”
(cited in Berto [12, p. 37]). Priest [84, ch. 9] capitalizes on the work of Russell, and
provides a general scheme of many kinds of paradox. Let  be a set with property
φ, such that:
1. Define:  = {y : φ(y)}.
2. If x is a subset of , then:
a. δ(x) ∈ x.
b. δ(x) ∈ .
The function δ is the “diagonalizer”. It plays the role of the diagonalization
technique to produce non-x members.
As long as x is a proper subset of , no contradiction arises. However, when
x = , while other conditions still hold, we end up with a paradox: by condition
(a) above, δ() ∈ , while condition (b) requires that δ() ∈ . This is essentially
Russell’s Paradox.
Intuitively, trying to treat the part as the whole is the heart of the paradox. There
must be limits on the subset x to avoid contradiction and paradox.
From an economic perspective, loosely speaking, the financial market is sup-
posed to be a proper subset, x, of the economy, ω, that facilitates trade and
production. However, when financial claims generate profits purely from other
financial claims, the financial market becomes decoupled from the economy, and
so x becomes a subset of itself, i.e. x becomes . As finance becomes detached
from the real economy, we end up with vicious circles of self-reference.
Money and language share many features: both are abstract social conventions,
as Leonard [68] points out. Each is form, not substance. The two are good examples
of complex systems [1]. As we can extrapolate from the work of Gödel and many
others, language cannot define itself; it is dependent on an external reference [102,
p. 111]. The same is true for money. Money and pure financial claims, therefore,
cannot stand on their own; the financial sector must be part of the real economy.

5.2 Quantifying Vicious Self-reference or Reflexivity

According to Sornette and Cauwels [99], most of the financial crashes have
fundamentally an endogenous, or internal, origin and that exogenous, or external,
shocks only serve as triggering factors. Detecting these internal factors however
is not always very obvious. Filimonov and Sornette [45] use techniques originally
introduced to model the clustered occurrences of earthquakes, to measure endogene-
ity of price changes in financial markets. In particular, they aim to quantify how
294 S. Al-Suwailem et al.

much of price changes are due to endogenous feedback processes, as opposed to

exogenous news. They apply the techniques to the E-mini S&P 500 futures contracts
traded in the Chicago Mercantile Exchange from 1998 to 2010. They find that the
level of endogeneity has increased significantly from 1998 to 2010: 30% of price
changes were endogenous in 1998, while more than 70% were endogenous since
2007. At the peak, more than 95% of the trading was due to endogenous triggering
effects rather than genuine news. The measure of reflexivity provides a measure
of endogeneity that is independent of the rate of activity, order size, volume or
volatility.
Filimonov et al. [44] quantify the relative importance of short-term endogeneity
for financial markets (financial indices, future commodity markets) from mid-2000s
to October 2012. They find an overall increase of the reflexivity index since the
mid-2000s to October 2012, which implies that at least 60–70% of financial price
changes are now due to self-generated activities rather than novel information,
compared to 20–30% earlier.
These results suggest that price dynamics in financial markets are mostly endoge-
nous and driven by positive feedback mechanisms involving investors’ anticipations
that lead to self-fulfilling prophecies, as described qualitatively by Soros’ concept
of reflexivity.
It has been observed that while the volatility of the real economy has been
declining (in industrial countries) since World War II, volatility of financial markets
has been on the rise [2]. Bookstaber [17] comments on this phenomenon saying:
The fact that total risk of the financial markets has grown in spite of a marked decline in
exogenous economic risk to the country is a key symptom of the design flaws within the
system. Risk should be diminishing, but it isn’t.

6 Applications, II: Mispricing Risk

Systematically ignoring irreducible uncertainty might lead to underestimation of

risk, and therefore, to a systematic downward bias in pricing risk.
This can be seen if we realize that ignored uncertainty is not necessarily
symmetric across possibilities of loss and gain. The error in mispricing may not
necessarily cancel out through averaging. Assuming that ignored uncertainty is
symmetric across all possible arrangements of economic factors, it follows that
chances of loss will be underestimated more than chances of gain. The reason is that
there are many more arrangements that end up in loss than in gain. Gain, like order
in physical systems, requires very special arrangements, while loss, like entropy, has
many more possibilities. Consequently, systematically ignoring the limits of formal
models might lead to underestimation of possibilities of loss more than those of
gain. Hence, risk might be systematically underpriced.
Is Risk Quantifiable? 295

With underpricing of risk, agents would take on more risk than they should.
Accordingly, markets become more exposed to instabilities and turbulences. As
early as 1996, this was visible to people like Bernstein [9]. He points out that despite
the “ingenious” tools created to manage risks, “volatilities seem to be proliferating
rather than diminishing” (p. 329).

6.1 Shortsightedness

There is another reason for mispricing risk that can lead to greater risks. Theory
of finance, as Bernstein [10] also points out, revolves around the notion that risk
is equivalent to volatility, measured usually by beta, standard deviation, and related
quantities. These quantities by nature measure risk in the short run, not the long run.
This causes investors to focus more on the shorter run. Major and extreme events,
however, usually build up over a longer horizon. This means that signs of danger
might not be detectable by shortsighted measures of risk. By the time the signs are
detected, it is too late to prevent or escape the crash.
With shortsightedness, rationality of individual agents may lead to “collective
irrationality”. In the short run, it is reasonable to assume that certain variables are
exogenous and are thus not affected by the decisions of market players. But in the
medium to long-run, these variables are endogenous. By ignoring the endogeneity
of these variables, agents ignore the feedback of their collective actions. This
feedback however might invalidate the short-term estimates of risk that agents are
using. When everyone is excessively focusing on the short-run, they collectively
invalidate their individual estimates, which is one way to see the rationale of Lucas
Critique discussed earlier. Hence, agents will appear to be individually rational,
when collectively they might not be so.
Former president of the European Central Bank, Trichet [106], argues that
excessive focus on short-term performance resulted in “excessive risk-taking and,
particularly, an underestimation of low probability risks stemming from excessive
leverage and concentration”. He therefore calls for “a paradigm change” to over-
come the shortsightedness dominating the financial sector.

6.2 Emotions and Herd Behaviour

Emotions play a major role in financial markets, leading to mispricing of risk.

Psychoanalyst David Tuckett conducted in 2007 detailed research interviews with
52 experienced asset managers, working in financial centres around the globe [108].
The interviewed managers collectively controlled assets of more than $500 billion
in value. Tucker points out that it was immediately clear to him that financial
assets are fundamentally different from ordinary goods and services (p. xvi).
Financial assets, he argues, are abstract objects whose values are largely dependent
296 S. Al-Suwailem et al.

on expectations of traders of their future values (pp. 20–25). This created an

environment for decision-making that is completely different from that in other
markets. It was an environment in which there is both inherent uncertainty and
inherent emotional conflict (p. 19). Traders’ behaviour therefore is predominantly
influenced by emotions and imaginations. These emotions are continuously in flux,
which adds to the volatility of markets. Traders seek protection by developing
groupfeel and shared positions (p. 173). This explains the well-documented herd
behaviour in financial markets that plays a major role in contagion and instability
[49, 93, 98]. In short, logical indeterminacy arising from self-reference makes
emotions and herding dominate financial markets, causing higher volatility and
instability that cannot be accounted for by formal models.
Coleman [32, p. 202] warns that “The real risk to an organization is in the
unanticipated or unexpected–exactly what quantitative measures capture least well”.
Quantitative tools alone, Coleman explains, are no substitute for judgment, wisdom,
and knowledge (p. 206); with any risk measure, one must use caution, applying
judgement and common sense (p. 137). David Einhorn, founder of a prominent
hedge fund, wrote that VaR was “relatively useless as a risk-management tool and
potentially catastrophic when its use creates a false sense of security among senior
managers and watchdogs” (cited in Nocera [79]).
Overall, ignoring the limits of quantitative measures of risk is dangerous, as
emphasized by Sharpe and Bernstein [10, 11], among others [74]. This leads to the
principle of Peter Bernstein: a shift away from risk measurement to risk manage-
ment [16]. Risk management requires principles, rules, policies, and procedures that
guide and organize financial activities. This is elaborated further in the following
section.

7 Applications, III: Dealing with Uncertainty

If uncertainty is irreducible and fundamentally unquantifiable, even statistically,

what to do about it? In an uncertain environment, the relevant strategy is one that
does not rely on predicting the future. “Non-predictive strategies” refer to strategies
that are not dependent on the exact course the future takes. Such strategies call
for co-creating the future rather than trying to predict it and then act accordingly
[105, 113]. Non-predictive strategies might imply following a predefined set of rules
to guide behaviour without being dependent on predicting the future.
Heiner [58] argues that genuine uncertainty arises due to the gap between
agents’ competency and environment’s complexity. In a complex environment,
agents overcome this gap by following simple rules (or “coarse behavior rule”;
see Bookstaber and Langsam [18]) to reduce the complexity they face. As the gap
between agents’ competency and environment’s complexity widens, agents tend to
adhere to more rigid and less flexible rules, so that their behaviour becomes stable
and more predictable. This can be verified by comparing humans with different
kinds of animals: as the competency of an animal becomes more limited, its
Is Risk Quantifiable? 297

behaviour becomes more predictable. Thus, the larger the gap between competency
and complexity, the less flexible will be animal’s behaviour, and thus the more
predictable it becomes. The theorems of Gödel and Turing show that the gap
between agents’ competency and environment’s complexity can never be closed;
in fact, the gap in some respects is infinite. The surprising result of this analysis, as
Heiner [58, p. 571] points out, is that:
. . . genuine uncertainty, far from being un–analyzable or irrelevant to understanding behav-
ior, is the very source of the empirical regularities that we have sought to explain by
excluding such uncertainty. This means that the conceptual basis for most of our existing
models is seriously flawed.

North [80] capitalizes on Heiner’s work, and argues that institutions, as con-
straints on behaviour, develop to reduce uncertainty and improve coordination
among agents in complex environments. Institutions provide stable structure to
every day life that guide human interactions. Irreducible uncertainty therefore
induces predictable behaviour as people adopt rules, norms and conventions to
minimize irreducible uncertainty (see Rosser [90]).
Former Federal Reserve chairman, Ben Bernanke, makes the same point with
respect to policy (cited in Kirman [61]):
. . . it is not realistic to think that human beings can fully anticipate all possible interactions
and complex developments. The best approach for dealing with this uncertainty is to make
sure that the system is fundamentally resilient and that we have as many failsafes and back-
up arrangements as possible.

The same point is emphasized by former governor of Bank of England, King

[60, p. 120]: “rules of thumb – technically known as heuristics – are better seen as
rational ways to cope with an unknowable future”.
Constraints in principle could help in transforming an undecidable problem into a
decidable one. Bergstra and Middleburg [8, p. 179] argue that it is common practice
in computer science to impose restrictions on design space to gain advantages and
flexibility. They point out that:
In computer architecture, the limitation of instruction sets has been a significant help for
developing faster machines using RISC (Reduced Instruction Set Computing) architectures.
Fast programming, as opposed to fast execution of programs, is often done by means
of scripting languages which lack the expressive power of full-blown program notations.
Replacing predicate logic by propositional calculus has made many formalizations decid-
able and for that reason implementable and the resulting computational complexity has been
proved to be manageable in practice on many occasions. New banking regulations in con-
ventional finance resulting from the financial crisis 2008/2009 have similar characteristics.
By making the financial system less expressive, it may become more stable and on the long
run more effective. Indeed, it seems to be intrinsic to conventional finance that seemingly
artificial restrictions are a necessity for its proper functioning.

Another approach to avoid undecidability is to rely more on qualitative rather

than quantitative measures. This can be viewed as another kind of restrictions that
aim to limit uncertainty. According to Barrow [4, pp. 222, 227], if we have a logical
theory that deals with numbers using only “greater than” or “less than”, without
referring to absolute numbers (e.g. Presburger Arithmetic), the theory would be
298 S. Al-Suwailem et al.

complete. So, if we restrict our models to qualitative properties, we might face less
uncertainty. Feyerabend [43, pp. xx, 34] points out that scientific approach does
not necessarily require quantification. Quantification works in some cases, fails in
others; for example, it ran into difficulties in one of the apparently most quantitative
of all sciences, celestial mechanics, and was replaced by qualitative considerations.
Frydman and Goldberg [47, 48] develop a qualitative approach to economics,
“Imperfect Knowledge Economics”, which emphasizes non-routine change and
inherently imperfect knowledge as the foundations of economic modelling. The
approach rejects quantitative predictions and aims only for qualitative predictions of
market outcomes. Historically, several leading economists, including J.M. Keynes,
were sceptical of quantitative predictions of economic phenomena (see Blaug [13,
pp. 71–79]).
Certain restrictions therefore are needed to limit uncertainty. The fathers of free
market economy, Adam Smith, John Stuart Mill and Alfred Marshall, all realized
that banking and finance needs to be regulated in contrast to markets of the real
economy [27, pp. 35–36, 163].
In fact, the financial sector usually is among the most heavily regulated sectors
in the economy, as Mishkin [78, pp. 42–46] points out. Historically, financial
crises are associated with deregulation and liberalization of financial markets. Tight
regulation of the banking sector following the Word War II suppressed banking
crises almost completely during 1950s and 1960s [21]. According to Reinhart and
Rogoff [87], there were only 31 banking crises worldwide during the period 1930–
1969, but about 167 during the period 1970–2007. The authors argue that financial
liberalization has been clearly associated with financial crises.
Deregulation has been visibly clear in the years leading to the Global Financial
Crisis. The famous Glass-Steagal Act has been effectively repealed in 1999 by
the Gramm-Leach-Bliley Act. Derivatives were exempted from gaming (gambling)
regulations in 2000 by the Commodity Futures Modernization Act (see Marks [75]).
Within a few years, the world witnessed “the largest credit bubble in history”, as
Krugman [65] describes it.
Sornette and Cauwels [99] argue that deregulation that started (in the US)
approximately 30 years ago marks a change of regime from one where growth is
based on productivity gains to one where growth is based on debt explosion and
financial gains. The authors call such regime “perpetual money machine” system,
that was consistently accompanied by bubbles and crashes. “We need to go back
to a financial system and debt levels that are in balance with the real economy”
(p. 23).
In summary, the above discussion shows the need and rationale in principle for
institutional constraints of the financial sector. The overall objective is to tightly link
financial activities with real, productive activities. Regulations, as such, might not
be very helpful. Rather than putting too much emphasis on regulations per se, more
attention should be directed towards good governance and values that build trust and
integrity needed to safeguard the system (see Mainelli and Giffords [74]).
Is Risk Quantifiable? 299

8 Conclusion

In his lecture at Trinity University in 2001 celebrating his Nobel prize, Robert Lucas
[23] recalls:
I loved the [Samuelson’s] Foundations. Like so many others in my cohort, I internalized
its view that if I couldn’t formulate a problem in economic theory mathematically, I didn’t
know what I was doing. I came to the position that mathematical analysis is not one of
many ways of doing economic theory: It is the only way. Economic theory is mathematical
analysis. Everything else is just pictures and talk.

The basic message of this chapter is that mathematics cannot be the only way. Our
arguments are based on mathematical theorems established over the last century.
Mathematics is certainly of a great value to the progress of science and accumulation
of knowledge. But, for natural and social sciences, mathematics is a tool; it is “a
good servant but a bad master”, as Harcourt [56, p. 70] points out. The master
should be the wisdom that integrates logic, intuition, emotions and values, to guide
decision and behaviour to achieve common good. The Global Financial Crisis shows
how overconfidence in mathematical models, combined with greed and lack of
principles, can have devastating consequences to the economy and the society as
a whole.
Uncertainty is unavoidable, not even statistically. Risk, therefore, is not in
principle quantifiable. This would have a substantial impact not only on how to
formulate economic theories, but also on how to conduct business and to finance
enterprises. This chapter has been an attempt to highlight the limits of reason, and
how such limits affect our abilities to predict the future and to quantify risk. It is
hoped that this contributes to a better reform of economics and, subsequently, to
better welfare of the society.

Acknowledgements We are grateful to the editors and an anonymous referee for constructive
comments and suggestions that greatly improved the readability of this text. FAD: wishes to
acknowledge research grant no. 4339819902073398 from CNPq/Brazil, and the support of the
Production Engineering Program, COPPE/UFRJ, Brazil. SA: wishes to acknowledge the valuable
discussions with the co-authors, particularly FAD. The views expressed in this chapter do not
necessarily represent the views of the Islamic Development Bank Group.

References

1. Al-Suwailem, S. (2010). Behavioural complexity. Journal of Economic Surveys, 25, 481–506.

2. Al-Suwailem, S. (2014). Complexity and endogenous instability. Research in International
Business and Finance, 30, 393–410.
3. Arrow, K., & Debreu, G. (1954). The existence of an equilibrium for a competitive economy.
Econometrica, 22, 265–89.
4. Barrow, J. (1998). Impossibility: The limits of science and the science of limits. Oxford:
Oxford University Press.
300 S. Al-Suwailem et al.

5. Barwise, J. (1989). The situation in logic. Stanford, CA: Center for the Study of Language
and Information.
6. Barwise, J., & Moss, L. (1996). Vicious circles. Stanford, CA: Center for the Study of
Language and Information.
7. Bays, T. (2014). Skolem’s paradox. In E. Zalta (Ed.), Stanford Encyclopedia of Philosophy.
Stanford, CA: Stanford University, https://plato.stanford.edu/.
8. Bergstra, J. A., & Middelburg, C. A. (2011). Preliminaries to an investigation of reduced
product set finance. Journal of King Abdulaziz University: Islamic Economics, 24, 175–210.
9. Bernstein, P. (1996). Against the Gods: The remarkable story of risk. New York: Wiley.
10. Bernstein, P. (2007). Capital ideas evolving. New York: Wiley.
11. Bernstein, P. (2007). Capital ideas: Past, present, and future. In Lecture delivered at CFA
Institute Annual Conference.
12. Berto, F. (2009). There’s something about Gödel: The complete guide to the incompleteness
theorem. Hoboken: Wiley-Blackwell.
13. Blaug, M. (1992). The methodology of economics. Cambridge: Cambridge University Press.
14. Blaug, M. (1999). The disease of formalism, or what happened to neoclassical economics
after war. In R. Backhouse & J. Creedy (Eds.), From classical economics to the theory of the
firm (pp. 257–80). Northampton, MA: Edward Elgar.
15. Blaug, M. (2002). Is there really progress in economics? In S. Boehm, C. Gehrke, H. Kurz,
& R. Sturm (Eds.), Is there progress in economics? European Society for the History of
Economic Thought (pp. 21–41). Northampton, MA: Edward Elgar.
16. Blythe, S. (2007). Risk is not a number. http://www.advisor.ca/investments/alternative-
investments/risk-is-not-a-number-26881.
17. Bookstaber, R. (2007). A demon of our own design. New York: Wiley.
18. Bookstaber, R., & Langsam, J. (1985). On the optimality of coarse behavior rules. Journal of
Theoretical Biology, 116, 161–193.
19. Bolander, T. (2002). Self-reference and logic. # News, 1, 9–44. http://phinews.ruc.dk/.
20. Bolander, T. (2014). Self-reference. In E. Zalta (Ed.), Stanford Encyclopedia of Philosophy.
Stanford, CA: Stanford University, https://plato.stanford.edu/.
21. Bordo, M., Eichengreen, B., Klingebiel, D., & Martinez-Peria, M. S. (2001). Is the crisis
problem growing more severe? Economic Policy, 32, 53–82.
22. Bouchaud, J. P. (2008). Economics needs a scientific revolution. Nature, 455, 1181.
23. Breit, W., & Hirsch, B. (Eds.). (2009). Lives of the laureates (5th ed.). Cambridge, MA: MIT
Press.
24. Byers, W. (2007) How mathematicians think: Using ambiguity, contradiction, and paradox
to create mathematics. Princeton: Princeton University Press.
25. Byers, W. (2011). The blind spot: Science and the crisis of uncertainty. Princeton: Princeton
University Press.
26. Calude, C., Jurgensen, H., & Zimand, M. (1994). Is independence an exception? Applied
Mathematical Computing, 66, 63–76.
27. Cassidy, J. (2009). How markets fail: The logic of economic calamities. New York: Farrar,
Straus, and Giroux.
28. Chaitin, G., Doria, F. A., & da Costa, N. C. A. (2011). Gödel’s way: Exploits into an
undecidable world. Boca Raton, FL: CRC Press.
29. Chrystal, K. A., & Mizen, P. D. (2003). Goodhart’s law: Its origins, meaning and implications
for monetary policy. In P. Mizen (Ed.), Central banking, monetary theory and practice: Essays
in honour of Charles Goodhart (Vol. 1). Northampton, MA: Edward Elgar.
30. Clower, R. (1995). Axiomatics in economics. Southern Economic Journal, 62, 307–319.
31. Colander, D., Goldberg, M., Haas, A., Juselius, K., Kirman, A., Lux, T., & Sloth, B. (2009).
The financial crisis and the systemic failure of the economics profession. Critical Review, 21,
249–67.
32. Coleman, T. (2011). A practical guide to risk management. Charottesville, VA: CFA Institute.
33. Collins, H. (2010). Tacit and explicit knowledge. Chicago: University of Chicago Press.
Is Risk Quantifiable? 301

34. da Costa, N. C. A., & Doria, F. A. (1991). Undecidability and incompleteness in classical
mechanics. International Journal of Theoretical Physics, 30, 1041–1073.
35. da Costa, N. C. A., & Doria, F. A. (1992). On the incompleteness of axiomatized models for
the empirical sciences. Philosophica, 50, 73–100.
36. da Costa, N. C. A., & Doria, F. A. (2005). Computing the future. In K. Velupillai (Ed.),
Computability, Complexity and Constructivity in Economic Analysis. Malden, MA: Blackwell
Publishers.
37. Davis, M. (2000). Engines of logic. New York: W.W. Norton.
38. DeLong, H. (1970). A profile of mathematical logic. Reading, MA: Addison-Wesley.
39. Dyson, F. (2006). The scientist as rebel. New York: New York Review Books.
40. Erickson, J. (2015). Models of computation. University of Illinois, Urbana-Champaign. http://
jeffe.cs.illinois.edu
41. Etzioni, A. (1988). The moral dimension: Toward a new economics. New York: The Free
Press.
42. Ferreirós, J. (2008). The crisis in the foundations in mathematics. In T. Gowers (Ed.), The
Princeton companion to mathematics (pp. 142–156). Princeton: Princeton University Press.
43. Feyerabend, P. (2010). Against the method (4th ed.). London: Verso.
44. Filimonov, V., Bicchetti, D., Maystre, N., & Sornette, D. (2013). Quantification of the high
level of endogeneity and of structural regime shifts in commodity markets. The Journal of
International Money and Finance, 42, 174–192.
45. Filimonov, V., & Sornette, D. (2012). Quantifying reflexivity in financial markets: Towards a
prediction of flash crashes. Physical Review E, 85(5), 056108.
46. Franzen, T. (2004). Inexhaustibility. Boca Raton: CRC Press.
47. Frydman, R., & Goldberg, M. (2007). Imperfect knowledge economics: Exchange rates and
risk. Princeton: Princeton University Press.
48. Frydman, R., & Goldberg, M. (2011). Beyond mechanical markets: Asset price swings, risk,
and the role of state. Princeton: Princeton University Press.
49. FSA. (2009). The Turner review. London: Financial Services Authority.
50. Gleiser, M. (2014). The island of knowledge: The limits of science and the search for meaning.
New York: Basic Books.
51. Gödel, K. (1931). Über formal unentscheidbare Sätze der Principia Mathematica und
verwandter Systeme I (On formally undecidable propositions of Principia Mathematica and
related systems I). Monatshefte für Mathematik und Physik, 38, 173–198.
52. Gödel, K. (1947). What is Cantor’s continuum problem? American Mathematical Monthly,
54, 515–525.
53. Gödel, K. (1995). Some basic theorems on the foundations of mathematics and their
implications. In S. Feferman (Ed.), Kurt Gödel Collected Works, Volume III: Unpublished
essays and lectures (pp. 304–323). Oxford: Oxford University Press.
54. Goleman, D. (2006). Social intelligence. New York: Bantam Books.
55. Gray, J. (2008). Thoralf Skolem. In T. Gowers (Ed.), The Princeton companion to mathemat-
ics (pp. 806–807). Princeton: Princeton University Press.
56. Harcourt, G. (2003). A good servant but a bad master. In E. Fullbrook (Ed.), The crisis in
economics. London: Routledge.
57. Hawking, S. (2002). Gödel and the end of the universe. www.hawking.org.uk.
58. Heiner, R. (1983). The origin of predictable behaviour. American Economic Review, 73, 560–
595.
59. Keynes, J. M. (1936/1953). The general theory of employment, interest and money. San Diego:
Harcourt Brace Jovanovich.
60. King, M. (2016). The end of alchemy. New York: W.W. Norton.
61. Kirman, A. (2012). The economy and the economic theory in crisis. In D. Coyle (Ed.),
What’s the use of economics? Teaching the dismal science after the crisis. London: London
Publishing Partnership.
62. Kline, M. (1980). Mathematics: The loss of certainty. Oxford: Oxford University Press.
63. Knight, F. (1921). Risk, uncertainty, and profit. New York: Houghton-Mifflin.
302 S. Al-Suwailem et al.

64. Kremer, P. (2014). The revision theory of truth. In E. Zalta (Ed.), Stanford Encyclopedia of
Philosophy. Stanford, CA: Stanford University, https://plato.stanford.edu/.
65. Krugman, P. (2009, September 2). How did economists get it so wrong? New York Times.
66. Laplace, P. S. (1902/1951). A philosophical essay on probabilities. Translated into English
from the original French 6th ed. by F.W. Truscott & F.L. Emory. New York: Dover
Publications.
67. Lavoie, D. (1986). The market as a procedure for discovery and conveyance of inarticulate
knowledge. Comparative Economic Studies, 28, 1–19.
68. Leonard, R. (1994). Money and language. In J. L. Di Gaetani (Ed.), Money: Lure, lore, and
literature (pp. 3–13). London: Greenwood Press.
69. Lewis, H. (2013). Lecture notes. Harvard University, http://lewis.seas.harvard.edu/files/
harrylewis/files/section8sols.pdf.
70. Lieberman, M. (2013). Social: Why our brains are wired to connect. New York: Crown
Publishers.
71. Lloyd, S. (2006). Programming the universe. New York: Vintage Books.
72. Lloyd, S. (2013). A Turing test for free will. Cambridge, MA: MIT Press. http://arxiv.org/abs/
1310.3225v1.
73. Machtey, M., & Young, P. (1979). An introduction to the general theory of algorithms. New
York: North–Holland.
74. Mainelli, M., & Giffords, B. (2009). The road to long finance: A systems view of the credit
scrunch. London: Centre for the Study of Financial Innovation. Heron, Dawson & Sawyer.
www.csfi.org.uk.
75. Marks, R. (2013). Learning lessons? The global financial crisis five years on. Journal and
Proceeding of the Royal Society of New South Wales, 146, 3–16, A1–A43.
76. McCauley, J. (2012). Equilibrium versus market efficiency: Randomness versus complexity in
finance markets. In S. Zambelli & D. George (Eds.), Nonlinearity, complexity and randomness
in economics (pp. 203–210). Malden, MA: Wiley-Blackwell.
77. Mintzberg, H. (1994). The rise and fall of strategic planning. New York: The Free Press.
78. Mishkin, F. (2007). The economics of money, banking, and financial markets (Alternate
edition). New York: Pearson.
79. Nocera, J. (2009, January 2). Risk mismanagement. The New York Times.
80. North, D. (1990). Institutions, institutional change and economic performance. Cambridge,
MA: Cambridge University Press.
81. Polanyi, M. (1966). The tacit dimension. New York: Doubleday.
82. Popper, K. (1956/1988). The open universe: An argument for indeterminism. London:
Routledge.
83. Priest, G. (1994). The structure of the paradoxes of self-reference. Mind, 103, 25–34.
84. Priest, G. (2000). Beyond the limits of thought (2nd ed.). Oxford: Oxford University Press.
85. Putnam, H. (1967). Psychophysical predicates. In W. Capitan & D. Merrill (Eds.), Art, mind,
and religion. Pittsburgh: University of Pittsburgh Press.
86. Rajan, U., Seru, A., & Vig, V. (2015). The failure of models that predict failures: Distance,
incentives, and defaults. Journal of Financial Economics, 115, 237–260.
87. Reinhart, C., & Rogoff, K. (2009). This time is different: Eight centuries of financial folly.
Princeton: Princeton University Press.
88. Rice, H. G. (1953). Classes of recursively enumerable sets and their decision problems.
Transactions of the American Mathematical Society, 74, 358–366.
89. Rogers, H., Jr. (1967). Theory of recursive functions and effective computability. New York:
McGraw-Hill.
90. Rosser, J. B. (2001) Alternative Keynesian and post Keynesian perspectives on uncertainty
and expectations. Journal of Post Keynesian Economics, 23, 545–66.
91. Rubinstein, A. (1998). Modeling bounded rationality. Cambridge, MA: MIT Press.
92. Savin, N., & Whitman, C. (1992). Lucas critique. In P. Newman, M. Milgate, & J. Eatwell
(Eds.), New Palgrave dictionary of money and finance. New York: Palgrave.
93. Shiller, R. (2009). Irrational exuberance. Princeton: Princeton University Press.
Is Risk Quantifiable? 303

94. Simon, H. (1978). Rationality as process and as product of thought. American Economic
Review, 68, 1–16.
95. Simon, H. (1978/1992). Rational decision-making in business organizations. In A. Lindbeck
(Ed.), Nobel Lectures, Economics 1969–1980 (pp. 343–371). Singapore: World Scientific
Publishing Co. https://www.nobelprize.org.
96. Soros, G. (2003). Alchemy of finance. Hoboken: Wiley.
97. Soros, G. (2009, October 26). The general theory of reflexivity. Financial Times.
98. Sornette, D. (2003). Why stock markets crash: Critical events in complex financial systems.
Princeton: Princeton University Press.
99. Sornette, D., & Cauwels, P. (2012). The illusion of the perpetual money machine. Notenstein
Academy White Paper Series. Zurich: Swiss Federal Institute of Technology. http://arxiv.org/
pdf/1212.2833.pdf.
100. Stewart, I. (1991). Deciding the undecidable. Nature, 352, 664–665.
101. Stuart, T. (2013). Understanding computation. Beijing: O’Reilly.
102. Svozil, K. (1993). Randomness and undecidability in physics. Singapore: World Scientific.
103. Svozil, K. (1996). Undecidability everywhere? In A. Karlqvist & J.L. Casti (Eds.), Boundaries
and barriers: On the limits to scientific knowledge. Reading, MA: Addison-Wesley, pp. 215–
237. http://arxiv.org/abs/chao-dyn/9509023v1.
104. Taleb, N. (2007). The black swan. New York: Random House.
105. Taleb, N. (2012). Antifragile. New York: Random House.
106. Trichet, J.-C. (2009). A paradigm change for the global financial system. BIS Review, 4/2009.
107. Tsuji, M., Da Costa, N. C. A., & Doria, F. A. (1998). The incompleteness of theories of games.
Journal of Philosophical Logic, 27, 553–568.
108. Tuckett, D. (2011). Minding the market: An emotional finance view of financial instability.
Basingstoke: Palgrave MacMillan.
109. Turing, A. (1936). On computable numbers, with an application to the Entscheidungsproblem.
Proceedings of the London Mathematical Society, Series, 2(42), 230–265.
110. Velupillai, K. V. (2000). Computable economics. Oxford: Oxford University Press.
111. Velupillai, K. V. (2012). Introduction. In K. V. Velupillai et al. (Eds.), Computable economics.
International Library of Critical Writings in Economics. Northampton, MA: Edward Elgar.
112. Velupillai, K. V., Zambelli, S., & Kinsella, S. (Eds.). (2012). Computable economics.
International Library of Critical Writings in Economics. Northampton, MA: Edward Elgar.
113. Wiltbank, R., Dew, N., Read, S., & Sarasvathy, S. D. (2006). What to do next? A case for
non-predictive strategy. Strategic Management Journal, 27, 981–998.
114. Winrich, J. S. (1984). Self-reference and the incomplete structure of neoclassical economics.
Journal of Economic Issues, 18, 987–1005.
115. Wolpert, D. (2008). Physical limits of inference. Physica, 237, 1257–1281.
116. Yanofsky, N. (2003). A universal approach to self-referential paradoxes, incompleteness and
fixed points. The Bulletin of Symbolic Logic, 9, 362–386.
117. Yanofsky, N. (2013). The outer limits of reason: What science, mathematics, and logic cannot
tell us. Cambridge, MA: MIT Press.
Index

A Consumer search, 210, 211, 214, 215

Adaptive expectations, 154 Consumption set, 118
Agent-based artificial stock market, 71 Continuous double auction, 42, 79
Agent-based macroeconomic model, 205, 206, Coordination device, 120
215 Coordination failure, 118
Agent-based model, 18, 19, 21, 33, 74, 91, CRSP value weighted portfolio, 231
207, 251, 253, 255, 259
Agent-based Scarf economy, 115, 127, 138
D
Agriculture, 92
Dark pools, 17–19
Allocative efficiency, 44
dark pool usage, 18–20, 22, 25, 29, 30, 33
Anderson-Darling (AD) test, 105
midpoint dark pool, 21, 32, 33
Artificial stock market, 76, 89, 181
midprice, 22, 25, 32
ATS, see automated trading systems
Decentralized market, 146
Automated trading systems, 39
Delta profit, 44
Average Imitation Frequency, 135–137
Disequilibrium, 114, 118, 136, 138
Double coincidence of wants, 117
DSGE, 141
B
Dual learning, 164, 171, 173
Barter, 144
Behavioral economics, 193
Bio-fuels, 91 E
Black swan, 283 Economic factors, 240, 241
Boom-and-bust, 196 Economic growth, 206
Brownian motion, 194 Emergence, 282, 284
Business cycle, 205, 206, 212, 215, 216 (the whole is greater than the parts), 280,
281, 291
Emotions, 295
C Endogenous dynamics, 194, 198
Canonical correlation, 223, 224, 229, 235, 242 Entropy, 294
Canonical correlation analysis, 247 Equal risk premia, 224
Capital asset pricing model, CAPM, 230 Equilibrium, 194
Common agricultural policy, 92 Experimental economics, 41
Common factor, 233 Exploitation, 134, 138
Comovement, 225 Exploration, 122, 127, 131, 134, 136–138

© Springer Nature Switzerland AG 2018 305

S.-H. Chen et al. (eds.), Complex Systems Modeling and Simulation
in Economics and Finance, Springer Proceedings in Complexity,
https://doi.org/10.1007/978-3-319-99624-0
306 Index

F L
Factor model, 224 Law of Effect, 121
Fama-French-Carhart, 234 Learning, 136
Flash crash, 36 individual learning, 114, 118, 119, 121,
Forward premium puzzle, 161–166, 169–171, 125, 138
173, 174, 177 meta-learning, 114, 115, 121, 122, 124,
Fractures, 36, 40, 41, 61, 63, 65 125, 127, 131, 133
social learning, 114, 118, 120, 121, 125,
137, 138
G Leontief payoff function, 114, 116
Gödel–event, 280, 283 Leptokurtosis, 195
Gödel sentences, 279, 283–285 Life cycle assessment (LCA), 92
Genetic programming, 71, 76 Limited information, 161, 163, 164, 166–171,
Geometric Brownian motion, 197 174, 177
Gibbs-Boltzmann distribution, 121 Liquidity traders, 21, 22
Global financial crisis, 275, 283, 292, Lloyd’s algorithm, 134
297–299 Local interaction, 136
GMM, 226 Lucas critique, 292
Guru, 191 Luxembourg, 92

H M
Halting problem, 280, 287 Market efficiency, 162–165, 167, 168, 174
Hedonic, 205, 207–209, 211–213, 215 Market factor, 230
Herd behavior, 181 Market fraction, 125, 137
Herding, 196 Market integration, 224
Heterogeneous agent models, 195 Market orders, 19, 25, 30, 32, 33
Heterogeneous agents, 114, 115, 127–129 Market quality, 19, 20
Heteroskedasticity, 195 market volatility, 20, 25, 28–33
Heuristics, 297 spreads, 18–20, 32, 33
HFT, see high-frequency trading Market share, 205, 206, 211, 215, 216
High-frequency trading, 39, 71, 72 Market volatility, 18, 19
Hyperbolic tangent function, 119 Mean Absolute Percentage Error, 126, 127,
136
Meta-preferences, 291
I Minsky, 202
Imitation, 120, 125, 131, 133, 134, 136 Mispricing risk, 277, 294
Incompleteness, 276, 277, 279, 282, 289 Multiple equilibria, 143, 151
Informational asymmetry, 288
Information criteria, 232
Innovation, 125, 131, 133, 134, 136 N
product, 205–207, 209–212, 214–216 Neoclassical economics, 38
Instability, 292, 296 Network structures, 181
Institutions, 297, 298 Non-convergence, 118
Integration measures, 243, 244 Non-Markovian processes, 194
Intensity of choice, 115, 122, 124–137 Non-tâtonnement, 114
Interaction effects, 125, 127, 136 Normalized Imitation Frequency, 131, 133,
Irreducible uncertainty, 276, 286, 294, 296, 134
297 Numerairé, 131

K O
K-means clustering, 134 Order-driven market, 21
Kolmogorov-Smirnov (KS) test, 105 continuous double-auction market, 18
Index 307

limit order book, 19 Risk premia, 230

market impact, 18, 23 Risk premia test, 241
price improvement, 18, 20, 25 Robot phase transition, 39, 40, 50, 53
evidence for, 60, 64
RPT, see robot phase transition
P R-Square, 225
Paradoxes, 291, 293
Persistent learning, 164, 167, 170, 174
Perverse incentives, 196 S
Positive feedback, 194 Saturation condition, 117
Positive mathematical programming (PMP), 92 Scarf economy, 115, 118, 123, 127
Power law of practice, 134 Self-reference, 290, 291, 293
Predictability, 277, 279–281, 283 Separation of timescales, 197
Presburger arithmetic, 297 Shortsightedness, 295
Price discovery, 98 Smith’s alpha, 43
Price stickiness, 146 Social network, 181
Price volatility, 181 Soros, 202
Principal components, 224 Stochastic differential equation, 194
Private information, 146 Stock markets, 224
Private prices, 116 Subjective prices, 116, 117, 119
Production set, 118 Supercomputer technologies, 251, 252, 256,
Product quality, 205, 206, 211, 213, 214, 217 259
Product space, 205, 215 Symmetric rational expectations equilibrium,
Product variety, 206 150
Profit dispersion, 44

T
Q Tacit knowledge, 288
Qualitative predictions, 298 Tâtonnement, 114, 115
Quantifiability Time-decay function, 119
of knowledge, 288 Turing machine, 280, 285–287, 289, 290
of natural and social phenomena, 275
of risk, 277, 287, 296
of the future, 282
U
of uncertainty, see irreducible uncertainty
Uncertainty principle, 281
Undecidability, 280, 285, 297
R
Rational expectations equilibrium, 142
Rational expectations hypothesis, 142 V
Rationality, 286, 290, 292, 295 Volume weighted average price, 28
Real business cycles, 143
Recency parameter, 122
Reference point, 122, 123 W
Reflexivity, 292 Walrasian auctioneer, 114
Regulations, 297, 298 Walrasian equilibrium, 142
Reinforcement learning, 114, 121, 122, 131 Walrasian general equilibrium, 115
REIT, 224
Representative agent, 141, 194
Research and development, 207, 209–218 Z
Rice’s theorem, 287, 288, 290 Zero-intelligence, 18, 19, 21, 33