Working Paper in Economics

# 202024

July 2020

Applying a Bayesian Network to

VaR Calculations

Emma Apps

https://www.liverpool.ac.uk/management/people/economics/
© authors
 Applying a Bayesian Network to VaR Calculations

Abstract

This paper focuses considers a methodology for deriving stock returns and VaR through
the application of a Bayesian Network (BN). A network map is specified where the returns
for three stocks are deemed to be conditionally dependent on two factors. The latter are
defined having previously considered literature relating to the financial crisis and risk
contagion. Subsequently, two factors are identified as influencing the individual stock
returns – one relating to liquidity and the other relating to the market. Following application
of the Gaussian Bayesian Network, regressions generate models for the said returns. The
latter are then used to simulate time series of stock returns and those outcomes are
compared to the original data series. The BN specification is found to be a satisfactory
alternative for the modelling of stock returns. Furthermore, the resulting quantiles are
shown to be more prudent estimates in relation to VaR calculations at the 5% level and,
therefore, can result in increases in regulatory capital.
Applying a Bayesian Network to VaR Calculations

1 Introduction

In their survey of 31 quantitative measures of systemic risk, Bisias et al (2012) identify a

research method in relation to Network Analysis in general. Specifically, a small network

of factors is defined as being systemically important in relation to their impact on the

returns of a set of financial entities. Where each factor is regarded as commonly significant

to each entity. Existing research tends to focus on applying such networks in the assessment

of how events spread through a financial system and interconnectedness in general. For

example, simulating how the failure of one bank can trigger the domino effects across many

and whether certain ones are more resilient to the default than others. Indeed, Chan-Lau et

al (2009) and the IMF (2009a) use network models to assess the impact of a failing bank

on others given respective exposures between them. While the specified networks can be

used to quantify VaR losses at the bank level following the original default and subsequent

domino effect, this is rarely discussed. This paper thereby attempts to contribute to existing

literature by applying a Bayesian Network of two factors to determine their impact on the

returns of three UK banking stocks and their three-stock portfolio in terms of VaR.

Identification of the factors isn’t necessarily intuitive but existing literature in relation to

financial linkages and reasons for the spread in financial crises, can be drawn upon. For

example, issues around market liquidity are raised and I suggest that the latter is an

important factor when assessing impacts on stock returns. There are certain market

indicators of the overall health and strength of liquidity among financial institutions, such

as the LIBOR-OIS spread in the UK and the TED-spread in the US. Indeed, Hull and White
(2013) suggest that, despite both spreads being stable and largely ignored pre-financial

crisis, both are now used as the summary indicators of liquidity following their extreme

movements in 2007 and 2008. Subsequently, in order to define a workable network, I begin

with just two factors – firstly the aforementioned liquidity factor and secondly, the

influence of the wider financials’ sector on each stock. In terms of visualising the network,

there are a series of nodes connected to each other by edges – where the latter represent the

relationship between the nodes. Thereafter, the resulting model is used to simulate returns

data for each bank and their three-stock portfolio and quantify their respective 5% and 1%

quantiles - where the latter can be used in a VaR calculation and be reasonably applied as

an alternative to the RiskMetrics approach. The network itself is specified using Bayesian

techniques as presented by Scutari and Denis (2015) and Shonoy and Shonoy (2000).

This paper is divided into several parts. Section 2 highlights the recent literature in relation

to Network Analysis and measuring systemic risk but also general applications of Bayesian

Networks (BN). Section 3 presents the data, identifying each time series and summary

statistics. Application of the BN to this data set in modelling stock returns is presented in

section 4 – including specification of the network, the underlying probability distributions

and tests of conditional independence and model specifications. The process for simulating

stock returns is also discussed. Results are detailed in section 5 – specifically the respective

significance of the partial correlations, the parameters of the BN model specifications and

the comparisons of the simulated summary statistics and quantiles versus those of the actual

historical returns. The paper ends with concluding remarks.

2 Relevant Literature

2.1 Network Analysis

A network rationale has been applied in a diverse range of social and behavioural science

contexts. For example, considering how large corporations differ in the extent to which

they offer support or assistance to local communities in which they have a presence.

Corporate and social responsibility dictates that they should be actively involved in their

communities but how much of that is influenced by the activities of other corporations? A

network can be used to model how such community involvement is influenced by their

interactions and relationships with other corporations. Likewise, in any decision-making

process involving several individuals or groups, a network approach can be used to

understand how individuals within a group influence each other in the decision-making

process. A common underlying theme is how the units within the network interact – they

are not viewed in isolation. According to Faust and Wasserman (1994, pp. 7):

“The network perspective differs in fundamental ways from

standard social and behavioural science research methods.

Rather than focus on attributes of autonomous individual units,

the associations among these attributes, or the usefulness of one

or more attributes for predicting the level of another attribute

are theorised and modelled through a network.”

Such associations and relationships can be witnessed in many other contexts, certainly

within science, finance and economics. Indeed, the interlinkages and interconnectedness
between financial institutions and within financial systems, leading to spreading in crises

are directly relevant (Diao et al 2000). From a scientific perspective, networks are used in

a variety of contexts – engineering, biology, ecology, medicine. For example, they are used

to analyse ecological systems and specifically how the food chains and ecosystems are

connected. In relation to public health, Luke and Harris (2007) present their use in the study

of how diseases are transmitted, specifically HIV and AIDS. Applications in medical and

microbiology contexts are popular – for instance, Barabasi et al (2011) use network-based

methods in genetics to identify molecular linkages and subsequent gene mutations.

Of course, this paper is interested in their application in finance and economics, specifically

in relation to systemic risk. Given the focus on liquidity issues, particularly in the interbank

markets, the work of Chan-Lau et al (2009), is relevant. In using network models, they

highlight the impact of institutional failure when there are exposures within such markets

– where the network illustrates the domino effect between connected banks when exposed

to a failing institution. Furthermore, Bilio et al (2012) go beyond the inter-bank markets in

their application of Granger-causality networks to the study of interconnectedness between

a variety of investor sub-groups, specifically hedge funds, banks, brokers and insurance

companies. Likewise, the IMF (2012) assess linkages within the global over-the-counter

derivative markets and the identification of systemically important financial

intermediaries. In each case, as suggested by Battiston et al (2012), there is no widely

accepted, single methodology to determine the systemically important nodes or factors

within the network – it is very much linked to interpretation and the underlying data set
(financial instrument, market, sub-sector, region). The latter indicates the degree of

qualitative judgement required in defining the network in the first instance. Nevertheless,

Allen and Babus (2009, pp. 367) argue that network analysis can assist our understanding

of financial systems and specifically risk contagion, given the interconnections revealed by

the 2008 financial crisis. Furthermore, aside from defining the network itself, they suggest

that it can then be usefully applied in formulating a regulatory framework for supervising

financial institutions, an objective entwined within this paper. Consistent with Allen and

Babus (2009), Hu at al (2012) allude to the deficiencies of pre-existing methods in

measuring exposures to systemic risk, given the significant widespread losses post 2008.

Accordingly, they too suggest a network-based approach as a more appropriate and

accurate measurement and monitoring process.

Unsurprisingly, there has been an upsurge in interest in research in this area - several

empiricists identify the importance of the use of network analysis. For example, Markose

et al (2012) apply a network to investigate the connections between banks in the Credit

Default Swap market – the latter market being identified as a key determinant of substantial

losses in 2008. In some cases, there is the final realisation that, given their widespread

application in science and medicine, surely analogies can be drawn in finance. For instance,

Haldane and May (2011) apply the dynamics of food webs in an ecological context to

modelling the stability of a given financial system. A leading empiriscist in relation to

network theory, Kimmo Soramaki, has several publications focusing on applications in

finance. For instance, Soramaki et al (2016) simplify complex network structures in order
to filter or highlight the most important determinants of correlations between returns of

European stocks. Earlier studies focus on the interbank payment systems and, specifically,

the creation of a network representing how payments are transferred between financial

institutions (see Soramaki et al (2007)). The latter highlights the key players in such

markets and the degrees of connectedness between them but also makes the point that the

“minor” players in the market are also connected to the more tightly connected core of

major players. Given the financial linkages, the network illustrates the severe impact of

any subsequent disruption to it and the issues arising in transferring and accessing capital

through the interbank markets. This is further explored by Soramaki and Cook (2013) and

Soramaki and Langfield (2016), whereby, following a bank’s failure, the disruption to the

payment network is identified along with systemically important institutions and the

resulting impacts on individual network participants. A common theme, once again, is the

interbank markets. It is clear that, whether referring to literature immediately following the

crisis or more recently, that theme remains - the liquidity issues generating from within

the inter-bank markets. Similar to Chan-Lau et al (2009), Krause and Giansante (2012) also

focus on the exposures within those markets and use a network of connected banks to

model how failure of one spreads through the network. Subsequently, a factor

encompassed within the BN defined in section 4, relates to liquidity – denoted by a

particular spread quoted in the inter-bank markets.

2.2 Network Analysis from a Bayesian Perspective

Bayesian Networks are encompassed within the framework of network analysis and

incorporate graphical theories and conditional dependencies between variables in the

graphical network. Within the literature there are several instances of the application of

BNs to data sets, not necessarily from a finance perspective. Indeed, almost any event

conditional on the probability of a prior event can be analysed using this concept. In

geographical and environmental studies, for example, a BN is used to evaluate flood plains

and the extent of flooding given certain extreme prior events such as changes in sea level

and improvements in coastal defenses (see Narayan et al 2018). They are also applied

within the context of Social Corporate Responsibility in assessing a corporation’s likely

compliance with child labour regulations across their supply chain network. The BN is

used to determine the likelihood of breaches to such regulations using available data on

suppliers, their employee demographics and the frequency of child labour incidents (see

Thoni et al 2018). From a medical research perspective, BNs are also readily applied. For

instance, in assessing links between patients diagnosed with clinical depression and

variability in their heart rates and also in identifying important factors in relation to survival

rates from lung cancer (see Anisa and Lin 2017).

From a risk management perspective, there is ample literature relevant to their application,

particularly in relation to operational risk. Essentially, various factors are identified and

inserted into the BN with estimates made of associated loss distributions resulting from the

various risk factors. According to Cowell et al (2007), such techniques can be applied in

insurance settings when assessing the financial impact of cases of fraud upon the insuring
company and thereby in the setting aside of adequate regulatory capital in relation to such

cases. From a banking perspective, Aquaro et al (2010) present their application in relation

to losses sustained through cases of failed internal processes, human error, IT failures and

certain litigation cases. The factors leading to the losses in each situation become part of

the BN and the associated loss distributions are generated. Clearly, there is some degree of

subjectivity in identifying the break-downs in the internal processes or human interventions

leading to loss making errors but, this is a commonality across all BNs, regardless of the

arena in which they are being applied. In all cases, analogies can certainly be drawn with

VaR and the need to ascertain the loss quantiles from subsequent returns’ distributions.

Indeed, Martin et al (2005) apply BNs to specifically model the severest loss inducing

events, referred to as the long tails or unexpected losses from an operational loss

perspective – similar of course to the 1% and 5% VaR scenarios. Of direct relevance to this

paper is the research of Hager and Andersen (2010) who seek to model loss severity across

all activities of a financial institution and not just from an operational perspective. This is

done through the identification of influencing factors – which I argue can be liquidity and

market based.

Other literature identifies the importance of risk contagion through BN modelling of

default probabilities resulting from financial linkages – for example Giudici and Spelta

(2016) and Chong and Kluppelberg (2017). Furthermore, interconnectedness is also

examined through the effect of exposures within the interbank markets. A BN is applied to

model individual institutional liabilities within that market and the subsequent impact on

other banks in the event that a participant in the network defaults. Gandy and Veraart
(2016) illustrate that the BN can be used to stress test differing assumed levels of inter-

bank liabilities and likelihood of default conditional on another bank defaulting. At the

very least it indicates the importance of the inter-bank markets once more, if not

specifically assessing the impact on bank returns’ distributions. Despite all of the literature

under review, however, there appears to be a lack of focus on application of BNs

specifically in modelling stock returns and losses applied in VaR estimations. This paper

seeks to provide a workable alternative approach to modelling both, whilst also considering

the importance of the entire financials sector, interlinkages and the impact of reducing

liquidity within the inter-bank markets.

3 The Data

In order to produce the appropriate network, and specifically assess the impact of the

chosen factors, the data is gathered for the period 14th December 2000 to 29th June 2012 –

implying 2,914 daily observations for each variable. This timeframe incorporates the

financial crisis but also adequate periods pre and post crisis. The data is sourced from

Bloomberg (excluding the portfolio) and the variables are as follows, where the daily return

and the daily percentage change ensure stationarity:

• Daily returns for Barclays Bank stock;

• Daily returns for Lloyds Bank stock;

• Daily returns for HSBC stock;

• Daily percentage change in the 3-month Sterling LIBOR vs. 3-month

sterling overnight index swap spread (OIS);

 • Daily returns for the MSCI Financials’ Sector Index:

• Daily returns for the three-stock portfolio.

The data set representing the impact of the market on each stock is the MSCI Financials

Sector Index, within which all three banking stocks have a percentage weighting. The

chosen liquidity factor is represented by the daily percentage change in the difference

between the 3-month sterling LIBOR rate and the 3-month sterling overnight indexed swap

rate. Overnight indexed swaps are interest rate swaps whereby a fixed rate of interest is

exchanged for floating and the latter is the average of a daily overnight rate. In deriving

the floating rate payment, the intention is to replicate the aggregate amount of interest that

would be earned from rolling over a sequence of daily loans at an appropriate overnight

rate. Given that we are applying a 3-month time-frame, it implies rolling over a sequence

of daily loans, for 90 days at an overnight rate – where that rate is determined in the UK

by the Bank of England and referred to as SONIA (sterling overnight index average). The

3-month sterling libor rate is the average interest rate at which a selection of banks lend

British pounds to one another for a period of 3 months.

In deriving the daily returns for each bank and the nominated market index, the following

is applied:

𝑝𝑡 −𝑝𝑡−1
𝑟𝑡 = (3.1)
𝑝𝑡−1

Where: 𝑝𝑡 refers to the closing price of the stock or index at

time t.
𝑝𝑡−1 refers to the closing price of the stock or index at time
t-1.
 𝑟𝑡 refers to the daily return of the stock or index at time t.

With regards the three-stock portfolio, we begin with a total initial investment of

£30,000,000, split equally between the stocks – representing an equal weighting of 33.33%

and £10,000,000 invested in each stock in the portfolio. As the prices of the component

stocks change in value, their weights in the portfolio change, as does the notional value of

the portfolio. The daily return on the portfolio is derived as follows:

𝑁𝑜𝑡𝑖𝑜𝑛𝑎𝑙 𝑉𝑎𝑙𝑢𝑒𝑝𝑜𝑟𝑡,𝑡 −𝑁𝑜𝑡𝑖𝑜𝑛𝑎𝑙 𝑉𝑎𝑙𝑢𝑒𝑝𝑜𝑟𝑡,𝑡−1

𝑟𝑝𝑜𝑟𝑡,𝑡 = (3.2)
𝑁𝑜𝑡𝑖𝑜𝑛𝑎𝑙 𝑉𝑎𝑙𝑢𝑒𝑝𝑜𝑟𝑡,𝑡−1

The notional value of the portfolio each day is derived as follows:

𝑁𝑉𝑝𝑜𝑟𝑡,𝑡 = [(1 + 𝑟𝑡,𝐵 ) × 𝑁𝑉𝑡−1,𝐵 ] + [(1 + 𝑟𝑡,𝐻 ) × 𝑁𝑉𝑡−1,𝐻 ] + [(1 + 𝑟𝑡,𝐿 ) × 𝑁𝑉𝑡−1,𝐿 ]
(3.3)

Where: 𝑁𝑉𝑝𝑜𝑟𝑡,𝑡 refers to the notional value of the 3-stock

portfolio at time t.
𝑟𝑡,𝐵 refers to the daily return of Barclays at time t;
𝑁𝑉𝑡−1,𝐵 refers to the notional value of investment in
Barclays at time t-1;
𝑟𝑡,𝐻 refers to the daily return of HSBC at time t;
𝑁𝑉𝑡−1,𝐻 refers to the notional value of investment in HSBC
at time t-1;
𝑟𝑡,𝐿 refers to the daily return of Lloyds at time t;
𝑁𝑉𝑡−1,𝐿 refers to the notional value of investment in Lloyds
at time t-1.

The graphs are presented for each time series of returns’ data in figures 3.1 to 3.5 plus an

indication of how the LIBOR-OIS spread moved in the period under review in figure 3.6.

The former illustrate stationarity in the time series and significant volatility in the 2007-
2008 time-frame of the financial crisis. In relation to the LIBOR-OIS spread, there are

noticeable peaks associated with certain key events. For instance, in September 2007, the

spread reached 85 basis points in response to the Bank of England announcing emergency

funding to rescue Northern Rock and, three months later as the crisis began to unfold,

reached an all-time high of 108 basis points. At its worst, following the insolvency of

Lehman Brothers in the autumn of 2008, the spread was around 300 basis points. The

Augmented Dickey Fuller tests at various lags in table 3.1, indicate the stationarity in the

time series of returns for each variable:

Table 3.1: Augmented Dickey Fuller tests for each variable

LIBOROIS Market Barclays HSBC Lloyds

1 Lag -47.0068* -38.5601* -36.5903* -41.127* -38.3401*

2 lags -40.925* -33.0937* -30.3896* -33.8231* -31.5608*

3 lags -33.7785* -27.1927* -25.4275* -28.0044* -27.0613*

4 lags -29.4624* -26.0475* -23.5451* -25.5346* -26.8292*

5 lags -24.4209* -24.5432* -21.8710* -24.7745* -25.0340*

6 lags -22.0149* -22.1618* -20.8262* -21.6686* -23.5661*

7 lags -20.7559* -19.8397* -19.6280* -19.5955* -21.9056

8 lags -18.6174* -18.4505* -17.2105* -18.2352* -20.2806*

9 lags -17.3829* -17.9787* -17.1425* -17.7670* -18.0233*

10 lags -16.8473* -17.4447* -16.3418* -17.9372* -16.7871*

Note: Critical values of -3.43, -2.86, -2.57. * denotes test statistic < critical values at all level
Figure 3.1: Time Series of Barclays Daily Returns.

Figure 3.2: Time Series of HSBC Daily Returns.

Figure 3.3: Time Series of Lloyds Daily Returns.

Figure 3.4: Time Series of Market Daily Returns.

Figure 3.5: Time Series of Portfolio Daily Returns.

Figure 3.6: Graph of the LIBOR-OIS spread.

In relation to the summary statistics presented in table 3.2, the mean daily returns appear

close to zero and the minimum returns reflect the substantial losses during the financial

crisis.

Table 3.2: Summary Statistics for LIBOROIS % change, Market, Stock and Portfolio Daily Returns.
LIBOROIS Market Barclays HSBC Lloyds Portfolio
Max 141.6667 16.0399 29.2357 15.5148 32.2159 21.2197
Min -67.6692 -9.8446 -24.8464 -18.7788 -33.9479 -16.3841
Median 0.0000 0.0000 -0.0501 0.0000 -0.0504 -0.0076
Mean 0.7612 -0.0223 -0.0076 0.0024 -0.0429 -0.0183

4 Application of a Gaussian Bayesian Network to Continuous Data

4.1 Proposed Network Structure

In applying Bayesian Networks to modelling data, they are useful in the situation where

information is incomplete and uncertainty exists over the key determinants of the

dependent variable. According to Shenoy and Shenoy (2000), there is initially a degree of

qualitative judgement and subjectivity in specifying the factors to include in the graphical

representation of the network. However, in subsequently applying quantitative tests of the

model and simulating posterior data distributions, certain inferences can be made about its

validity. In this instance the proposed network is being applied to model portfolio returns

based on certain inputs or factors added to it. Furthermore, given the simulated posterior

return distribution of the portfolio, a cut-off return is derived for use in a VaR calculation,

where the cut-offs refer to the 1% and 5% quantiles of the said distribution.

In specifying a Gaussian Bayesian Network, I am modelling continuous data sets with the
underlying assumption of multivariate normality. With regards the variables defined in

section 3, I denote them with the following abbreviations:

• Daily returns for Barclays Bank stock→B

• Daily returns for Lloyds Bank stock→L

• Daily returns for HSBC stock→H

• Daily percentage change in the 3-month Sterling LIBOR vs. 3-month

sterling overnight index swap spread (OIS) →S

• Daily returns for the MSCI Financials’ Sector Index→M

• Daily returns for the three-stock portfolio→P

Prior to tests of conditional independence, the suggested relationships between variables

are as follows:

B is directly influenced by S and M, L is directly influenced by S and M, H is directly

influenced by S and M and P is directly influenced by B, L and H. Consequently, the

proposed relationships are defined as follows:

{𝑆, 𝑀} → 𝐵, {𝑆, 𝑀} → 𝐿, {𝑆, 𝑀} → 𝐻, {𝐵, 𝐿, 𝐻} → 𝑃 (4.1)

4.2 Proposed Network Graph and Probability Distribution

Based upon the above suggested relationships between the variables a graphical

representation can be defined – as presented in figure 4.2.1. It is referred to as a directed

acyclic graph (DAG) and contains a series of arcs and nodes. The former reflect the direct
dependencies between variables and the latter reflect the variables within the network. Each

variable or node has its own distribution – for example, ‘B” has a distribution or time series

of daily returns. If an arc exists from one variable to another, the latter variable is dependent

upon the former, otherwise known as the parent. The overall distribution, encompassing

all variables and suggested dependencies, can be depicted as follows:

Pr(𝑆, 𝑀, 𝐵, 𝐿, 𝐻, 𝑃) = Pr(𝑆) Pr(𝑀) Pr (𝐵|𝑆, 𝑀)Pr (𝐻|𝑆, 𝑀)Pr (𝐿|𝑆, 𝑀)Pr (𝑃|𝐵, 𝐻, 𝐿)

Furthermore, the distributions at each node can be expressed as:

𝐵|𝑆 = 𝑠, 𝑀 = 𝑚 𝐻|𝑆 = 𝑠, 𝑀 = 𝑚 𝐿|𝑆 = 𝑠, 𝑀 = 𝑚 𝑃|𝐵 = 𝑏, 𝐻 = ℎ, 𝐿 = 𝑙

where, the distribution at each node is conditional on the values of its parents. Rather than

determining the overall joint probability distribution encompassing all variables from the

outset, the Bayesian Network (BN) approach breaks the distribution into sub-groups and

derives the local distributions at each node. Scutari and Denis (2015) present that

specifying a joint probability distribution is rather difficult and complex given the numbers

of variables and correlations requiring estimation. Therefore, the BN overcomes this

modelling issue through specifying the local distribution at each node conditional on the

values of the parents.

Figure 4.2.1: Proposed DAG of Relationship Between 2 factors, Stock Returns and Portfolio Returns

4.3 Algebraic Representation of the DAG

The conditional relationships for each of the nodes of the three stocks may be specified as

an equation, consistent with the assumptions that 1) every node follows a normal

distribution and 2) the equations represent a Gaussian linear model incorporating an

intercept, with the node’s parents as the explanatory variables. The specifications in this

case, for each factor and stock are as follows:

𝑆~𝑁(𝜇𝑆 , 𝜎𝑆2 ) 2
𝑀~𝑁(𝜇𝑀 , 𝜎𝑀 ) (4.2)

𝐵|𝑆 = 𝑠, 𝑀 = 𝑚~𝑁(𝛼𝐵 + 𝛽1,𝐵 𝑠 + 𝛽2,𝐵 𝑚, 𝜀𝐵2 ) (4.3)

 𝐻|𝑆 = 𝑠, 𝑀 = 𝑚~𝑁(𝛼𝐻 + 𝛽1,𝐻 𝑠 + 𝛽2,𝐻 𝑚, 𝜀𝐻2 ) (4.4)

𝐿|𝑆 = 𝑠, 𝑀 = 𝑚~𝑁(𝛼𝐿 + 𝛽1,𝐿 𝑠 + 𝛽2,𝐿 𝑚, 𝜀𝐿2 ) (4.5)

Where: 𝛼 refers to the intercepts, 𝛽 refers to the regression coefficients for

the parents, S and M and 𝜀 represents the standard deviation of the residuals.

There is no specification for the three-stock portfolio because its subsequent simulated

returns are derived using equations (3.2) and (3.3).

4.4 Testing for Conditional Independence

As each arc in the DAG encompasses a probabilistic dependence, conditional

independence tests can be used to assess whether the data actually supports it. In terms of

hypotheses, for each variable, the following conditional dependencies are being tested:

𝐻0 : 𝐵 𝑖𝑠 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑀|𝑆 versus 𝐻1 : 𝐵 𝑖𝑠 𝑛𝑜𝑡 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑀|𝑆

𝐻0 : 𝐵 𝑖𝑠 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑆|𝑀 versus 𝐻1 : 𝐵 𝑖𝑠 𝑛𝑜𝑡 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑆|𝑀

𝐻0 : 𝐻 𝑖𝑠 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑀|𝑆 versus 𝐻1 : 𝐻 𝑖𝑠 𝑛𝑜𝑡 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑀|𝑆

𝐻0 : 𝐻 𝑖𝑠 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑆|𝑀 versus 𝐻1 : 𝐻 𝑖𝑠 𝑛𝑜𝑡 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑆|𝑀

𝐻0 : 𝐿 𝑖𝑠 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑀|𝑆 versus 𝐻1 : 𝐿 𝑖𝑠 𝑛𝑜𝑡 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑀|𝑆

𝐻0 : 𝐿 𝑖𝑠 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑆|𝑀 versus 𝐻1 : 𝐿 𝑖𝑠 𝑛𝑜𝑡 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑓𝑟𝑜𝑚 𝑆|𝑀

The null hypothesis depicts that B, H or L may be independent from M given S or S given

M. If the null is proven, the Beta coefficients in equations 4.3 to 4.5 are equal to zero. Using

“B” as an example, through the hypotheses, the partial correlation between B and M given
S or S given M, is being tested – denoted by 𝜌𝐵,𝑀|𝑆 or 𝜌𝐵,𝑆|𝑀 . The null holds if 𝜌𝐵,𝑀|𝑆 or

𝜌𝐵,𝑆|𝑀 is not statistically different from zero. In the test, the appropriate distribution is a

student’s t distribution with n – 3 degrees of freedom (where n refers to the total number

of observations in each time series of B, H and L and 3 refers to the number of variables in

the test e.g. B, S and M).

2911
𝑡(𝜌𝐵,𝑀|𝑆 ) = 𝜌𝐵,𝑀|𝑆 √1−𝜌2 (4.6)
𝐵,𝑀|𝑆

The null hypothesis of independence is rejected if the corresponding p-value is less than

the 10%, 5% and 1% degrees of significance.

4.5 Simulating the Returns’ Distributions

Following the independence tests in section 4.4, the parameters of equations 4.3 to 4.5 are

estimated using the maximum likelihood estimator. Each time-series of bank returns, as

the response variables, are regressed on the time-series of the daily percentage change in

the LIBOROIS spread and the daily returns in the market index. Having determined the

parameters of the models proposed by the DAG in section 4.3, they are then used to

simulate sets of random variables for each node, B, H and L. Simulation is performed from

the BN by generating a sample of random values from the joint distribution of the specified

nodes. It is performed following the order implied by the arcs in the DAG – from the

parents first, followed by the children (LIBOROIS and the Market being the parents, the 3

banks being the children). For each node, 2,914 random values are generated – depicting

estimates of the daily returns for each stock. In each case, the simulation is performed on
the basis of both a normal distribution and a student’s t-distribution, using the “rnorm” and

“rt” functions in R-studio.

5 Results

5.1 Tests of Conditional Independence

For each of the banks, inverse correlation matrices are produced, which are required for

the significance tests and generation of p-values. The resulting correlations are presented

in tables 5.1.1 to 5.1.3.

Table 5.1.1: Correlation Matrix for Barclays versus 2 parent nodes

Barclays Returns LIBOROIS Market

Barclays Returns 1.0000 0.0332 0.7818

LIBOROIS 0.0332 1.0000 -0.0827

Market 0.78718 -0.0827 1.0000

Table 5.1.2: Correlation Matrix for HSBC versus 2 parent nodes

HSBC Returns LIBOROIS Market

HSBC Returns 1.0000 0.0306 0.7828

LIBOROIS 0.0306 1.0000 -0.0805

Market 0.7828 -0.0805 1.0000

Table 5.1.3: Correlation Matrix for Lloyds versus 2 parent nodes

Lloyds Returns LIBOROIS Market

Lloyds Returns 1.0000 0.0192 0.6953

LIBOROIS 0.0192 1.0000 -0.0787

Market 0.6953 -0.0787 1.0000

The respective significance tests for the partial correlations are presented in table 5.1.4 In

all cases, the bank returns have a significant positive correlation with the market (M) given

the daily percentage change in the LIBOROIS spread (S) and we can thereby reject the null

hypothesis of independence given the extremely small p-values at all levels of significance.

In relation to the conditional dependence between the bank returns and the LIBOROIS

variable, given the market returns, there is positive correlation but at a low level.

Furthermore, the p-values only indicate significance at the 10% level. However, at that

level of significance the null hypothesis of independence is rejected and we can surmise

that there is a degree of conditional dependence between daily bank returns and the chosen

indicator of liquidity in the financial markets. Thereby, both factors, deemed to be the

parents in the DAG, can subsequently be applied in the modelling of the bank returns.

Table 5.1.4: Significance Tests of Partial Correlations

𝐵~𝑆|𝑀 𝐵~𝑀|𝑆 𝐻~𝑆|𝑀 𝐻~𝑀|𝑆 𝐿~𝑆|𝑀 𝐿~𝑀|𝑆
Pearson’s 0.0332 0.7818 0.0306 0.7828 0.0192 0.6953
Correlation
Degrees of 2911 2911 2911 2911 2911 2911
Freedom
P-Value 0.0729* 0.0000*** 0.0991* 0.0000*** 0.0990* 0.0000***
Note: * denotes significance at 10%, ** significance at 5%, *** significance at 1%.

5.2 Parameters of the BN Model Specification

Following the Gaussian linear regression for each bank, the respective maximum

likelihood estimators are produced and presented in table 5.2.1. Values for the intercepts

(𝛼𝐵 , 𝛼𝐻 𝑎𝑛𝑑 𝛼𝐿 ) and contributions of the parents, as depicted by the Beta coefficients, are

provided.
Table 5.2.1: Parameters of the BN Model for the returns of each bank
𝛼𝐵 𝛼𝐻 𝛼𝐿 𝛽1,𝐵 𝛽2,𝐵 𝛽1,𝐻 𝛽2,𝐻 𝛽1,𝐿 𝛽2,𝐿 𝜀𝐵2 𝜀𝐻2 𝜀𝐿2
0.0175 0.0174 0.019 0.0053 1.304 0.0028 0.769 0.0036 1.197 1.952 1.152 2.322

The contributions from the LIBOROIS variable are small but the value of the spread itself

is also and percentage changes in the daily returns of any stock are not frequently sizeable.

Referring back to equations 4.3 to 4.5, the BN model specifications, following the linear

regression, are as follows:

𝐵|𝑆 = 𝑠, 𝑀 = 𝑚~𝑁(0.0175 + 0.0053𝑠 + 1.304𝑚, 1.952 ) (5.1)

𝐻|𝑆 = 𝑠, 𝑀 = 𝑚~𝑁(0.0174 + 0.0028𝑠 + 0.769𝑚, 1.152 ) (5.2)

𝐿|𝑆 = 𝑠, 𝑀 = 𝑚~𝑁(−0.019 + 0.0036𝑠 + 1.197𝑚, 2.322 ) (5.3)

𝑆~𝑁(0.76, 12.392 ) 𝑀~𝑁(−0.0223, 1.882 )

Equations 5.1, 5.2 and 5.3 are then applied in simulating sets of returns for the three bank

stocks.

5.3 Simulated Data

Following the simulation of time series of returns for each of the three bank stocks and

subsequent three-stock portfolio, a comparison is made between the summary statistics of

the original, actual data sets and the simulations, applying both a normal and student’s t-

distribution. Both are presented in tables 5.3.1 to 5.3.4.

Table 5.3.1: Comparison of Summary Statistics – Actual versus Simulated Returns - Barclays
Barclays Actual Simulated Returns Simulated Returns
Returns (Normal Dist’n) (t-distribution)
Mean -0.00759% -0.02102% 0.06029%
Max 29.2357% 10.45150% 11.02510%
Min -24.8464% -9.75948% -10.64131%
Median -0.05013% 0.01165% 0.06097%
Stdev 3.13367% 3.14959% 3.25881%

Table 5.3.2: Comparison of Summary Statistics – Actual versus Simulated Returns - HSBC
HSBC Actual Simulated Returns Simulated Returns
Returns (Normal Dist’n) (t-distribution)
Mean 0.002432% -0.008251% 0.03320%
Max 15.51481% 6.407118% 7.63505%
Min -18.77880% -6.580565% -7.42987%
Median -0.00000% 0.038593% 0.07663%
Stdev 1.84544% 1.82093% 2.13113%

Table 5.3.3: Comparison of Summary Statistics – Actual versus Simulated Returns - Lloyds
Lloyds Actual Simulated Returns Simulated Returns
Returns (Normal Dist’n) (t-distribution)
Mean -0.04285% -0.09143% -0.07486%
Max 32.21586% 10.26661% 10.93075%
Min -33.94790% -11.76306% -11.21700%
Median -0.007604% -0.02116% 0.000267%
Stdev 2.13653% 1.61464% 1.802019%
Table 5.3.4: Comparison of Summary Statistics – Actual versus Simulated Returns - Portfolio
Portfolio Actual Simulated Returns Simulated Returns
Returns (Normal Dist’n) (t-distribution)
Mean -0.018333% -0.04074% 0.011784%
Max 21.21970% 6.19230% 6.413994%
Min -16.38410% -5.72795% -5.939262%
Median -0.007604% -0.021160% 0.000267%
Stdev 2.13653% 1.61464% 1.802019%

Given that the mean return is expected to be close to zero, in all cases the simulated values

are consistent. Furthermore, the simulated standard deviations match the actuals with

reasonable accuracy. Of greater relevance is the model’s ability to derive meaningful

estimates of minimum returns, given the implications for VaR. For each bank and the

portfolio, the estimated minimum returns are significantly different from the actual values.

However, given the underlying assumption of a Gaussian BN and normality, it will not

necessarily correctly evaluate the tails of the distribution. It is encouraging that, when

applying a t-distribution, the resulting minima are larger than in the normal case for three

of the four data series – consistent with its ability to model tails more effectively. Despite

the clear differences in summary statistics, it is important to consider the relative accuracy

of the model in relation to quantiles. After all, they are used as the cut-off points in relation

to VaR calculations. Given that the most severe maxima or minima values for daily returns

occur so infrequently, they are not necessarily an accurate indicator of the most likely

maximum daily loss. Thereby, a comparison of the quantiles from the actual returns and

the simulated cases are presented in tables 5.3.5 to 5.3.8.

Table 5.3.5: Comparison of Quantiles – Actual versus Simulated - Barclays
Quantile Barclays Simulated Over / Simulated Over /
Actual (Normal Under (t- Under
Dist’n) Estimates distribution) Estimates1
1% -8.786% -7.782% Under -7.479% Under
5% -4.372% -5.174% Over -5.288% Over
10% -3.069% -3.976% Over -4.123% Over
90% 2.939% 3.854% Over 4.287% Over
95% 4.785% 5.035% Over 5.482% Over
99% 8.641% 7.322% Under 7.708% Under

Table 5.3.6: Comparison of Quantiles – Actual versus Simulated - HSBC

Quantile HSBC Simulated Over / Simulated Over /
Actual (Normal Under (t- Under
Dist’n) Estimates distribution) Estimates1
1% -5.279% -4.215% Under -4.718% Under
5% -2.546% -3.065% Over -3.529% Over
10% -1.835% -2.382% Over -2.774% Over
90% 1.849% 2.289% Over 2.752% Over
95% 2.689% 3.008% Over 3.615% Over
99% 5.107% 4.154% Under 5.035% Under

Table 5.3.7: Comparison of Quantiles – Actual versus Simulated - Lloyds

Quantile Lloyds Simulated Over / Simulated Over /
Actual (Normal Under (t- Under
Dist’n) Estimates distribution) Estimates1
1% -8.893% -7.859% Under -7.809% Under
5% -4.408% -5.390% Over -5.523% Over
10% -3.051% -4.204% Over -4.394% Over
90% 2.920% 3.927% Over 4.210% Over
95% 4.468% 5.268% Over 5.318% Over
99% 9.079% 7.487% Under 8.264% Under

Table 5.3.8: Comparison of Quantiles – Actual versus Simulated - Portfolio

Quantile Portfolio Simulated Over / Simulated Over /
Actual (Normal Under (t- Under
Dist’n) Estimates distribution) Estimates1
1% -5.874% -3.881% Under -4.236% Under
5% -3.099% -2.766% Under -2.997% Under
10% -2.187% -2.106% Under -2.312% Over
90% 2.079% 2.027% Under 2.300% Over
95% 3.163% 2.641% Under 2.938% Under
99% 6.423% 3.582% Under 4.318% Under
In all cases, the simulated 1% quantiles from the simulated returns, are less than those

based upon the actual time series of returns. Perhaps not surprising given the underlying

normal distribution assumption. However, the related simulated 5% and 10% quantiles are

larger than those on an actual basis. This implies greater prudence in subsequent VaR

estimates due to the left tail being larger in the simulated cases if the quantiles are used as

the appropriate cut-off. Despite the under-estimations at the 1% level, the simulated results

are still of use in a practical context due to the industry convention of reporting VaRs at

the 5% level for individual stocks. The simulated portfolio quantiles are slightly misleading

given that they are impacted by the respective weights of the component stocks. They are,

nevertheless, comparable to the portfolio actual quantiles at both the 5% and 10% levels.

Application of the t-distribution, allows for a more realistic modelling of the tails.

Consequently, in the simulations, the resulting 5% and 10% quantiles for all stocks are

even more prudent than in the normal case and also at the 1% level for HSBC.

Table 5.3.9 illustrates the absolute percentage increases in the 5% and 10% quantiles

offered by the simulated results. At the 5% level, the increases in the quantile range from

0.5% to just over 1%. From a regulatory perspective, and the setting aside of regulatory

capital based on VaR assessments, an additional 1% would be significant – if we consider

the notional values of stocks and equity portfolios.

Table 5.3.9: Absolute % increase in 5% and 10% quantiles offered by Simulated Data
Barclays HSBC Lloyds Barclays HSBC Lloyds
Normal Normal Normal t-dist’n t-dist’n t-dist’n
Dist’n Dist’n Dist’n
5% 0.802% 0.519% 0.982% 0.916% 0.983% 1.115%
10% 0.907% 0.547% 1.153% 1.054% 0.939% 1.343%
Finally, figures 5.3.1 to 5.3.3 reflect the fitted distributions of the simulated stock returns

according to the underlying assumption of normality.

Figure 5.3.1: Barclays Fitted Simulated Returns Figure 5.3.3: Lloyds Fitted Simulated Returns

Figure 5.3.2: HSBC Fitted Simulated Returns Consistent with the quantiles reflected in

tables 5.3.5 to 5.3.7, the left tails in the

fitted distributions reflect the 10%, 5%

and 1% levels. For example, the 1%

quantile for Lloyds being -7.859% and

reflected in the spread of the left-hand

side of figure 5.3.3.

Figure 5.3.4 reflects comparisons between the original and simulated returns for each stock

on the basis of an assumed Chi-Squared Distribution. In each case, the outcomes are

similar.
Figure 5.3.4: Comparative Graphs of Original versus Simulated Squared Returns and Chi-Squared
Distribution:
5.6 Concluding Remarks

This paper provided a BN approach to modelling stock returns. The data was sourced from

Bloomberg and included time series of daily returns for three UK banks, namely, Barclays,

HSBC and Lloyds. A subsequent portfolio was constructed from the three stocks. Using a

degree of qualitative judgement, a DAG was constructed using the evidence presented in

the literature with regards factors being important in relation to their impact on stock

returns. In this instance, the market and the liquidity factors were selected with the latter

being represented by the 3-month LIBOR versus OIS spread.

The DAG suggested conditional dependencies between the factors and stock returns,

subsequently verified by conditional independence tests and partial correlations. Whilst

low levels of significance were indicated for the liquidity factor, it did still exist and the

linear regression models were specified for the returns of each stock. The latter were

subsequently used to simulate time series of returns. Summary statistics and quantiles were

compared for the actual returns and the simulated returns. Whilst the simulated returns

underestimated minimum values, the quantiles were comparable at the 5% and 10% levels.

The latter suggests that the underlying Gaussian BN (GBN) could be applied in modelling

stock returns and could be further used to estimate quantiles and VaR cut offs. Although it

does assume normality, and may be regarded as over-simplifying the modelling issues, its

comparable estimations are a positive. An objective in this instance was to suggest a

workable alternative to the RiskMetrics approach in deriving VaR. As suggested by Scutari

and Denis (2015), a more complex specification may be preferred but relatively simple
models often perform better. Indeed, the widely used RiskMetrics approach is convenient

to apply and well understood but, I suggest that the GBN is as intuitive and, furthermore,

appears to provide prudent estimates for the quantiles used as the cut-offs in VaR

calculations. Given that losses were underestimated in the 2008 financial crisis applying

VaR techniques of the time, a model resulting in a potential 1% increase in regulatory

capital would be an improvement. Based on a portfolio with a notional value of £1 billion,

it would result in at least an additional £10 million in regulatory capital.

There are, of course, certain limitations with this technique, not least of which is

determining the DAG structure in the first instance. Subsequently, if the structure is

ascertained, there may be issues with data being readily available representing the

components of the DAG – for example sourcing regular data in relation to balance sheet

indicators such as levels of indebtedness or rising levels of delinquencies amongst bank

customers. Furthermore, although the network can be altered or updated for new

components, as the number of variables grows, the simulation methods may produce less

reliable estimations.
 References
Allen, F., and Babus, A., (2009), “Networks in Finance,” Wharton Financial Institutions Center

Working Paper No. 08-07

Anisa, K.N., and Lin, S.W., (2017), “Effect of Socioeconomic Status on Lung Cancer Survival:

A Mediation Analysis Based on Bayesian Network Approach.” 2017 IEEE International

Conference on Industrial Engineering and Engineering Management.

Aquaro, V., Bardoscia, M., Belotti, R., Consiglio, A., De Carlo, F., and Ferri, G., (2010), “A

Bayesian Networks Approach to Operational Risk.” Physica A: Statistical Mechanics and its

Applications, Volume 389, Issue No. 8, pp. 1721-1728.

Barabasi, A.l., Gulbahce, N., and Loscalzo, J., (2011), “Network Medicine: a network-based

approach to human disease.” Nature Reviews Genetics, Vol. 12, pp. 56-68.

Battiston, S., Puliga, M., Kaushik, R., Tasca, P., and Caldarelli, G., (2012), “DebtRank: Too

Central to Fail? Financial Networks, the FED and Systemic Risk.” Scientific Reports, Vol. 2,

Article number: 541.

Bilio, M., Getmansky, M., Lo, A.W., and Pelizzon, L., (2012), “Econometric Measures of

Systemic Risk in the Finance and Insurance Sectors.” Journal of Financial Economics, 104, pp.

535-559.
Bisias, D., Flood, M., and Lo, A.W., (2012). “A Survey of Systemic Risk Analytics.” Annual

Review of Financial Economics, Volume 4, pp. 255-296.

Chan-Lau, J., (2009), “Co-risk measures to assess systemic financial linkages.” IMF Working

Paper.

Chan-Lau, J., Espinosa, M., and Sole, J., (2009), “On the Use of Network Analysis to Assess

Systemic Financial Linkages.” Working Paper. International Monetary Fund. (Forthcoming)

Chong, C., and Kluppelberg, C., (2017), “Contagion in Financial Systems.” Working Paper,

Cornell University Library.

Cowell, R.G., Verrall, R.J and Yoon, Y.K., (2007), “Modelling Operational Risk with Bayesian

Networks.” Journal of Risk and Insurance, December 2007, Vol. 74, Issue. No. 4, pp. 795-

827.

Diao, X., Li, W., and Yeldan, E., (2000), “How the Asian Crisis Affected the World Economy:

A General Equilibrium Perspective.” Federal Reserve Bank of Richmond, Economic Quarterly

Volume 86/2 Spring 2000, pp. 37-38.

Faust, K., and Wasserman, S., (1994, pp. 7), “Social Network Analysis: Methods and

Applications.” Cambridge University Press. ISBN: 0521387078.

Gandy, A., and Veraart, L.A.M., (2017), “A Bayesian Methodology for Systemic Risk

Assessment in Financial Networks.” Management Science, Volume 63, Issue No. 12, pp. 4428-
4446.

Giudici, P., and Spelta, A., (2016), “Graphical Network Models for International Financila

Flows.” Journal of Business and Economic Statistics, Volume 34, Issue No. 1, pp. 128-138.

Hager, D., and Andersen, L.B., (2010), “A Knowledge Based Approach to Loss Severity

Assessment in Financial Institutions Using Bayesian Networks and Loss Determinants.”

European Journal of Operational Research, Volume 207, Issue No. 3, pp. 1635-1644.

Haldane, A.G., and May, R.M., (2011), “Systemic risk in banking ecosystems.” Nature,

International Journal of Science, Issue 469, pp. 351-355.

Hu, D., Zhao, J.L., Hua, Z., and Wong, M.C.S., (2012), “Network-Based Modelling and

Analysis of Systemic Risk in Banking Systems.” MIS Quarterly, Vol. 36, No. 4, pp. 1269-

1291.

Hull, J., and White, A., (2013), “LIBOR vs. OIS: The Derivatives Discounting Dilemma.” The

Journal of Investment Management, Vol. 11, No. 3, pp. 14-27.

International Monetary Fund (2009a), “Assessing the Systemic Implications of Financial

Linkages,” Global Financial Stability Review, pp. 73-110.

International Monetary Fund (2012), “Systemic Risk from Global Financial Derivatives: A

Network Analysis of Contagion and its Mitigation with Super-Spreader Tax.” IMF Working

Paper, (2012).
Krause, A., and Giansante, S., (2012), “Interbank Lending and the spread of bank failures: A

network model of systemic risk.” Journal of Economic Behaviour and Organisation, Vol. 83,

Issue No. 3, pp. 583 – 608.

Luke, D.A., and Harris, J.K., (2007), “Network Analysis in Public Health: History, Methods

and Applications.” Annual Review of Public Health, Vol. 28, pp. 69-93.

Markose, S., Giansante, S., and Shaghaghi, A.R., (2012), “Too Interconnected to fail financial

network of US CDS Market: Topological fragility and systemic risk.” Journal of Economic

Behaviour and Organisation, Vol. 83, Issue 3, pp. 627-646.

Martin, N., Fenton N., and Tailor, M., (2005), “Using Bayesian Networks to Model Expected

and Unexpected Operational Losses.” Risk Analysis, Volume 25, Issue No. 4, pp. 963-972.

Narayan, S., Nicholls, R.J., Clarke, D., and Simmonds, D., (2018), “A Bayesian Network

Model for Assessments of Coastal Inundation Pathways and Probabilities.” Journal of Flood

Risk Management, January 2018, 11:S233-S250.

Scutari, M., and Denis, J.B., (2015), “Bayesian Networks With Examples in R.” CRC Press,

Taylor and Francis Group, ISBN: 9781482225600.

Shenoy, C. and Shenoy, P.P., (2000), “Bayesian Network Models of Portfolio Risk and

Return.” Computational Finance, pp. 87-106, The MIT Press, Cambridge, MA.
Soramaki, K., Bech, M.L., Arnold, J., Glass,R.J., and Beyeler, W.E., (2007), “The Topology

of Interbank Payment Flows.” Physica A:Statistical Mechanics and its Applications, Volume

379, Issue 1, pp. 317-333.

Soramaki, K., Birch, J., and Pantelous, A.A., (2016), “Analysis of Correlation Based Networks

Representing DAX 30 Stock Price Returns.” Volume 47, Issue No. 4, pp. 501-525.

Soramaki, K., and Cook, S., (2013), “SinkRank: An Algorithm for Identifying Systemically

Important Banks in Payment Systems.” Economics, The Open-Access, Open-Assessment E-

Journal, Vol. 7, 2013-28.

Soramaki, K., and Langfield, S., (2016), “Interbank Exposure Networks.” Computational

Economics, Vol. 47, Issue No. 1, pp. 3-17.

Thoni, A., Tjoa., A.M., and Taudes, Q., (2018), “An Information System for Assessing the

Likelihood of Child Labour in Supplier Locations Leveraging Bayesian Networks and Text

Mining.” Journal of Information Systems and e-Business Management, 7th February 2018, pp.

1-34.