CDO Modeling: Techniques, Examples and

Applications
Christian Bluhm
December 10, 2003
Abstract.: Collateralized debt obligatons (CDOs) constitute an important subclass of asset backed securi-
ties. The evaluation of CDOs relies on mathematical modeling and on simulation as well as analytic and
semi-analytic approaches, depending on the underlying asset pool and the cash flow structure of the trans-
action. This paper is an introductory survey on CDO modeling. It starts with a mini course on the use of
CDOs as capital market instruments, explains simulation and analytic approaches for evaluating CDOs and
considers the notion of PD, EL and LGD of CDO tranches.

Contents
1 Introductory Remark 2

2 Four good reasons for CDO business 2

2.1 Motivation 1: Spread arbitrage opportunities . . . . . . . . . . . . . . . . . . . . . 3
2.2 Motivation 2: Regulatory capital relief . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Motivation 3: Funding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Motivation 4: Economic risk transfer . . . . . . . . . . . . . . . . . . . . . . . . . 8

3 CDOs from a quantitative perspective 9

3.1 Modeling the asset side of a CDO . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Modeling the liability side of a CDO . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 CDO modeling techniques 12

4.1 Calibration of a credit curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.2 The distribution of single-name default times . . . . . . . . . . . . . . . . . . . . 16
4.3 Multivariate default times distributions . . . . . . . . . . . . . . . . . . . . . . . . 18
4.4 Analytic approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.5 Semi-analytic techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

5 Further examples and applications 32

Disclaimer. The content of this paper reflects the personal view of the author and not the opinion of HypoVereinsbank.
The calculation and simulation examples in Sections 4 and 5 are based on realistic but fictitious transactions and data.
Their sole purpose is to illustrate the concepts presented in this paper.
Acknowledgements. The author thanks Christoff Gossl for more than two years of valuable cooperation in CDO
analysis and the ongoing fruitful discussions on all kinds of modeling issues. He also thanks colleagues and friends
for their helpful comments and suggestions on a first version of this paper.

HypoVereinsbank (Munich), Group Credit Portfolio Management, GCP3 - Structured Finance Analytics.
Comments are welcome and can be sent to the author by email: christian.bluhm@hvb.de

1
1 Introductory Remark
A main motivation for writing this survey paper on CDO modeling was an invitation to give the
closing talk at the Autumn School on Risk Management (September 29 - October 02, 2003 in
Herrsching/Ammersee) of the Munich University of Technology. The aim of the talk was to give a
brief and self-contained introduction to CDOs as structured finance instruments from a mathemat-
ical modeling point of view. The slides of the talk (although not all of the topics discussed in this
paper found their way onto them) can be downloaded at the Autumn Schools website
www.mathematik.tu-muenchen.de/gkaam/AutumnSchool/index.html
The paper is divided in two parts. Sections 2 and 3 are written in prose and can be considered as
a very brief and self-contained mini introduction to CDOs as capital market instruments. Sec-
tions 4 and 5 present some ideas regarding the mathematics of CDOs. Readers only interested
in the application of mathematical concepts to structured finance problems can safely omit Sec-
tion 2. However, my hope is that readers without any experience in CDOs but interested in their
functionality and applications will find Sections 2 and 3 useful.
A reference for readers interested in diving deeper into the mathematics of CDO modeling is [14],
where a much more detailled and mathematically more rigorous course in CDO modeling will
hopefully soon be ripe for publication.

2 Four good reasons for CDO business

CDO issuance showed a remarkable growth during recent years. Although this growth has slowed
down in 2003 due to a difficult economic environment, it can be expected that the growth of CDO
issuance will continue over the next years.
What people not involved in CDOs typically ask is the following question: What makes this asset
class so successful in the structured finance market, and why are people still considering new
variations of the scheme with unbroken enthusiasm? It is the purpose of the next four sections
to outline an answer to this question. Besides the four motivations summarized in the sequel, tax
and legal arbitrage play an important role in CDO structuring. Because they are not described by
mathematics, these aspects are not considered in this paper.
In the following we will look at two real life examples and one illustrative example. The first
two examples refer to transactions done by members of HypoVereinsbank (HVB) Group in the
past nine months. I have chosen these two particular structures from a variety of transactions done
by HVB because they nicely illustrate two of the four main motivations for CDO issuance. Both
transactions are public and the reader can find more details regarding the structures in the corre-
sponding presale reports of the rating agencies. Obviously, our purpose here is not to discuss the
performance or to disclose the economics of these transactions but rather to use them as examples
for good reasons why banks issue CDOs. Equivalently, one could have used purely illustrative
examples based on fictitious transactions, as exercised in Sections 4 and 5. However, there is a
good chance that readers will find it interesting to see at least some real examples.
Section 2.1 is intended to be an introductory warm-up and therefore more lengthy than the sub-
sequent parts of this section.

2
2.1 Motivation 1: Spread arbitrage opportunities
We begin this section by quoting some passages from an article which appeared in the web at
www.FinanceAsia.com (written by Rob Davies, March 20, 2003):
HVB Asset Management Asia (HVBAM) has brought to market the first ever hybrid collateralized
debt obligation (CDO) managed by an Asian collateral manager. The deal, on which HVB Asia
(formerly known as HypoVereinsbank Asia) acted as lead manager and underwriter, is backed by
120 million of asset-backed securitization bonds and 880 million of credit default swaps ... Under
the structure of the transaction, Artemus Strategic Asian Credit Fund Limited - a special purpose
vehicle registered in the Cayman Islands - issued 200 million of bonds to purchase the 120 million
of cash bonds and deposit 80 million into the guaranteed investment contract, provided by AIG
Financial Products. In addition, the issuer enters into credit default swap agreements with three
counterparties (BNP Paribas, Deutsche Bank and JPMorgan) with a notional value of 880 million.
On each interest payment date, the issuer, after payments of certain senior fees and expenses and
the super senior swap premium, will use the remaining interest collections from the GIC accounts,
the cash ABS bonds, the hedge agreements, and the CDS premiums from the CDS to pay investors
in the CDO transaction ... The transaction was split into five tranches, including an unrated 20
million junior piece to be retained by HVBAM. The 127 million of A-class notes have triple-A
ratings from Fitch, Moodys and S&P, the 20 million B-notes were rated AA/Aa2/AA, the 20
million C bonds were rated A/A2/A, while the 13 million of D notes have ratings of BBB/Baa2
and BBB.

Figure 1: Artemus Strategic Asian Credit Fund

3
Figure 1 shows a structural diagram of the transaction. As mentioned in the quotation from Fi-
nanceAsia above, the three major rating agencies analyzed the transaction. The reader can find the
results of their analysis in the three presale reports [1], [2] and [3] of the transaction.
We now want to look at the transaction in a more detailled manner. Hereby we focus on some
aspects only and try to simplify things as much as possible. For example, hedge agreements on
interest rates and currencies will be excluded from our discussion.
In the sequel, all amounts of money refer to USD.

Liability side of the structure

The issuer (Artemus Strategic Asian Credit Fund Limited, an SPV at Cayman, from now on shortly
called Artemus) issued 200mm of bonds, split in five tranches reflecting different risk-return
profiles. Artemus (as protection buyer) also entered into a CDS agreement (super senior swap) on a
notional amount of 800mm with a super senior swap counterparty. Such counterparties (protection
sellers on super senior swaps) are typically OECD-banks with excellent credit quality. Because
the liability side has a funded (200mm of notes) and an unfunded (800mm super senior swap) part,
the transaction is called partially funded.

Asset side of the structure

The proceeds of the 200mm issuance have been invested in a guaranteed investment contract (GIC
account; 80mm in eligible collateral assets) and asset backed securities (ABS bonds; 120mm).
Additionally, the issuer sold protection on a pool of names with an aggregated notional amount of
880mm. Because the asset side consists of a mixture of debt securities and synthetic assets (CDS),
the transaction is called hybrid. Note that the GIC is kind of risk-free (AAA-rated securities,
cash-equivalent).

Settlement of credit events

If credit events happen on the 880mm CDS agreement (remember: Artemus is protection seller), a
settlement waterfall takes place as follows.

Proceeds from the GIC account are used by Artemus to make payments on the CDS agree-
ment.

If proceeds from the GIC are not sufficient to cover losses, principal proceeds from the debt
securities are used to pay for losses.

If losses exceed the notional amount of the GIC and principal proceeds, then ABS securities
are liquidated and proceeds from such liquidation are used for payments on the 880mm CDS
agreement.

Only if all of the above mentioned funds are not sufficient for covering losses, the super
senior swap will be drawn (remember: Artemus bought protection from the super senior
swap counterparty).

Note that (at start) the volume of the GIC plus the super senior swap notional amount exactly match
the 880mm CDS agreement, and that the 120mm ABS Securities plus the 880mm CDS volume
asset-back the 1bn total tranche volume on the liability side. However, these coverage equations

4
refer only to principal and swap notionals outstanding. But there is much more credit enhancement
in the structure, because additional to the settlement waterfall, interest proceeds, mainly coming
from the premium payments on the 880mm CDS agreement and from the ABS bonds, mitigate
losses as explained in the following section.

Distribution of proceeds
Principal proceeds (repayment/amortization of debt securities) and interest proceeds (income on
ABS bonds, the GIC, hedge agreements and premium from the 880mm CDS agreement) are gen-
erally distributed sequentially top-down to the note holders in the order of their seniority. On top
of the interest waterfall, fees, hedge costs and other senior expenses and the super senior swap pre-
mium have to be paid. Both, principal and interest payments are subject to change in case certain
coverage test are broken. There are typically two types of coverage tests in such structures:

Overcollateralization tests (O/C) take care that the available (principal) funds in the struc-
ture are sufficient for a certain (over)coverage (encoded by O/C-ratios greater than 100%)
regarding repayments due on the liability side of the transaction.

Interest coverage tests (I/C) make sure that any expenses and interest payments due on the
liability side of the structure and due to other counterparties involved, e.g., hedge counterpar-
ties, are (over)covered (encoded by I/C-ratios greater than 100%) by the remaining (interest)
funds of the transaction.

If a test is broken, cash typically is redirected in a way trying to bring the broken test in line again.
In this way, the interest stream is used to mitigate losses by means of a changed waterfall. It is
beyond the scope of this paper to dive deeper into such cash flow mechanisms.

Excess spread
As already mentioned above, interest proceeds are distributed top-down to the note holders of
classes A, B, C and D. All excess cash left-over after senior payments and payments of coupons
on classes A to D is paid to the subordinated note investors. Here, HVB Asset Management Asia
(HVBAM) retained part of the subordinated note (the so-called equity piece). Such a constel-
lation is typical in arbitrage structures: Most often, the originator/arranger keeps some part of
the most junior piece in order to participate in the excess spread of the interest waterfall. Addi-
tionally, retaining part of the first loss of a CDO to some extent proves to the market that the
originator/arranger itself trusts in the structure and the underlying credits. As indicated above in
our discussion on coverage tests, if tests are broken excess cash typically is redirected in order
to protect senior note holders interests. Here, the timing of defaults (see Section 5) is essential:
If defaults occur at the end of the lifetime of the deal (backloaded), subordinated notes investors
had plenty of time to collect excess spread and typically will achieve an attractive overall return
on their investment even if they loose a substantial part of their invested capital. In contrast, if
defaults occur at an early stage of the transaction (frontloaded), excess cash will be redirected and
no longer distributed to the equity investor. This is a bad scenario for equity investors, because
they bear the first loss (will loose money) but now additionally miss their (spread) upside potential
because excess cash is trapped.

5
Where does the arbitrage come from?
Now where does the arbitrage come from? The key observation is that on the 880mm CDS agree-
ment and on the 120mm ABS securities on the asset side premiums are collected on a single-name
base, whereas premium/interest payments to the super senior swap counterparty and the note hold-
ers refer to a diversified pool of ABS bonds and CDS names. Additionally, the tranching of the
liability side into risk classes contributes to the spread arbitrage in that tranches can be sold for
a comparably low spread if sufficient credit enhancement (e.g., subordinated capital, excess cash
trapping, etc.) is built-up for the protection of senior notes.
Obviously, the total spread collected on single credit risky instruments at the asset side of the
transaction exceeds the total diversified spread to be paid to investors on the tranched liability
side of the structure. Such a mismatch typically creates a significant arbitrage potential which
offers an attractive excess spread to the equity or subordinated notes investor.
There are many such transactions motivated by spread arbitrage opportunities in the CDO mar-
ket. In some cases, structures involve a so-called rating arbitrage which arises whenever spreads
increase quickly and rapidly and the corresponding ratings do not react fast enough to reflect the
increased risk of the instruments. Rating arbitrage as a phenomenon is an important reason why
typically a serious analysis of arbitrage CDOs should not rely on ratings alone but also considers
all kinds of other sources of information.
From a modeling perspective, arbitrage structures constitute an interesting and challenging class
of CDOs because in most cases all kinds of cash flow elements are involved.
Looking at arbitrage structures from an economic point of view one could say that a well-structured
transaction like Artemus Strategic Asian Credit Fund has - due to the arbitrage spread involved -
a potential to offer an interesting risk-return profile to notes investors as well as to the origina-
tor/arranger holding (part of) an unrated junior piece. It is certainly possible that the incorporated
spread arbitrage is sufficiently high to compensate both groups of people adequately for the risk
taken.

A word on super senior swaps

Regarding super senior swaps one should mention that in most transactions the likelihood that
the super senior tranche gets hit by a loss will be close to zero. Scenarios hitting such a tranche
typically are located far out in the tail of the loss distribution of the underlying reference pool.
Looking at super senior swaps from a heuristic (non-mathematical) point of view, one can say that
in order to cause a hit on a super senior tranche the economy has to turn down so heavily that it is
very likely that problems will have reached a level where a super senior swap hit is just the tip of
the iceberg of a heavy global financial crisis.

2.2 Motivation 2: Regulatory capital relief

Regulatory capital relief is another major motivation why banks issue CDOs. Here, most often the
D in CDO becomes an L standing for loan. The CDO is then called a collateralized loan
obligation (CLO). Figure 2 shows a typical CLO issued for the purpose of regulatory capital relief.
Again, the reader has the option to look into presale reports of the rating agencies; see [4] and [5].
In Building Comfort 2002-1, HVB has bought protection on a pool of 5bn Euro of residential
mortgage backed loans. In this particular transaction, not even an SPV has been set-up. Instead,
HVB directly issued notes on the lower 5% of the notional volume (including a 50bps equity piece

6
 Figure 2: Building Comfort 2002-1

in form of a swap, the so-called Class-E Swap) and bought protection from an OECD-bank on the
upper (senior) 95% of the notional volume of the RMBS pool.
Let us briefly outline what such a transaction means for the regulatory capital requirement of the
underlying reference pool. In general, loan pools require regulatory capital in size of 8%RWA
where RWA denotes the risk-weighted assets of the reference pool. Ignoring collateral eligible
for a risk weight reduction, regulatory capital equals 8% of the pools notional amount. After the
(synthetic) securitization of the pool, the only regulatory capital requirement the originator has to
fulfill regarding the securitized loan pool is holding capital for retained pieces. For example, if
HVB would hold the equity piece, the regulatory capital required on the pool would have been
reduced by securitization from 8% down to 50bps, which is the size of the equity tranche. The
50bps come from the fact that retained equity pieces typically require a full capital deduction. As
already mentioned, in the special case of Building Comfort, HVB has implemented a swap (the
Class-E Swap, with a so-called interest subparticipation) into the equity piece.
Altogether one clearly sees that structures like Building Comfort are very useful tools if one is
interested in relief of regulatory capital. Even if a bank keeps an equity piece of, say, 3%, the
securitization implies a 5% relief of regulatory capital, ignoring collateral eligible for RWA re-
duction again. As opportunity costs for capital relief, the originating bank has to pay interest to
notes investors, a super senior swap premium, upfront costs (rating agencies, lawyers, structuring
and underwriting costs) ongoing administration costs and possibly some other expenses. A full
calculation of costs compared to the decline of regulatory capital costs is required to judge about
the economics of such transactions.

7
2.3 Motivation 3: Funding
There is not much to say about this third motivation banks have for CDO issuance. In so-called true
sale transactions, the transfer of assets is not made of derivative constructions but rather involves a
true sale off balance sheet of the underlying reference assets to an SPV which then issues notes
in order to refinance/fund the assets purchased from the originating bank. The advantage for the
originator is the receipt of cash (funding).
Funding can be an issue for banks whose rating has declined to a level where funding from other
sources is expensive. The advantage of refinancing by means of securitizations is that resulting
funding costs are mainly related to the credit quality of the transferred assets and not so much to
the rating of the originator. However, there remains some linkage to the originators rating, if the
SPV also enters into a servicer agreement with the originating bank. In such cases, investors and
rating agencies will evaluate the servicer risk inherent in the transaction.

2.4 Motivation 4: Economic risk transfer

The last of the four major motivations for CDO issuance is economic risk transfer. Figure 3 illus-
trates the impact of a securitization transaction on the loss distribution of the underlying reference
pool. Such a transaction divides the loss distribution in two segments. The left segment, the so-
called first loss, refers to losses carried by the originator (e.g., by retaining the equity piece). The
right segment refers to the excess loss of the first loss piece, taken by the CDO investors. The
upper boundary of the first loss piece is an effective loss cap the originator buys from the CDO
investors. The chart on the right in Figure 3 illustrates the lucky situation that the securitization
costs are much lower than the decline of the expected loss. In such situations, the risk-return profile
of the securitized pool will be improved by the securitization transaction.

Figure 3: Economic risk transfer (illustrative!)

The insurance paradigm

A standard question people ask regarding risk transfer in the context of securitizations is the follow-
ing: How can these deals transfer risk, if the first loss (often significantly higher than the expected

8
loss of the original pool) is kept by the originator? To answer this question, we go back to the
insurance paradigm sometimes used to explain the necessity of taking into account expected and
unexpected losses in credit pricing. In this paradigm the expected loss is calculated as an insurance
premium to be charged and deposited in some loss reserve account in order to cover the historic
mean losses observed w.r.t. the considered class of credit risky instrument. Analogously, a capital
cushion against unexpected losses is calculated by means of an economic capital definition (e.g.,
credit VaR, expected shortfall or some other risk measure). Now, if a bank securitizes a credit
portfolio and retains only the first loss piece (FLP), there is some risk transfer if and only if the
probability P[Loss > FLP] is greater than zero. Applying the insurance paradigm to this case, the
insurance premium for covering losses can be chosen somewhat lower after securitizing the pool
because there is an effective loss cap in place for the benefit of the originating bank.

Costs versus benefit

Summarizing one can say that most securitization transactions actually lead to some risk transfer.
The problem is that the risk cost saving alone does not always justify the securitization costs spent
on the liability side of a CLO (spread payments, hedge costs and other fees and costs). Moreover,
the upfront costs of setting up a CLO can be quite expensive too. There will be expenses for
structuring, underwriting, rating agency fees, lawyer costs, etc. Such costs hit the P&L of an
originating bank right at the start of the transaction.

Impact on the contributory economic capital

Another important point for investigation in the context of risk transfer is the change in contributory
economic capital implied by a securitization. The problem is that securitizing a subportfolio can
cause some negative effect on the economic capital of the residual source portfolio due to the
diversification turn-down caused by taking away a pool of diversifying assets; see [13], pages 256-
258. Obviously, if the volume of the securitized pool is small compared to the volume of the
source portfolio, such negative impact has a chance to be negligible.

Motivation for Monte Carlo simulation tools

Besides others, all of the above mentioned aspects have to be taken into account by originating
banks. Without tailor-made mathematical tools for evaluating the planned CLO in a way consistent
with the banks internal portfolio model it is impossible to draw a complete picture of the impact
of a securitization transaction. Of course, for CDO investors the same principles hold.

3 CDOs from a quantitative perspective

In this section we briefly indicate how CDO evaluation is done in general by means of Monte Carlo
simulations. The general picture we have in mind is given in Figure 4.
On the left side of a CDO there will be always some pool of credit risky instruments, e.g., loans,
bonds, credit derivatives (e.g., CDS), ABS notes, or even a combination of different asset classes.
We refer (and referred in Section 2) to this part of a CDO as the asset side of the structure. On the
right side of a CDO we typically have securities issued in the capital market. For obvious reasons
we call this part of the CDO the liability side of the transaction. The two sides of the CDO are
connected through the structural definition of the transaction. Typically the structure is represented

9
 Figure 4: Modeling CDOs - cause and response

by means of a term sheet, an indenture, an offering circular or whatever documentation the investor
receives from the originator/arranger.
Now, the key to a basic understanding of CDO modeling is the trivial fact that the only 1 reason
for uncertainty regarding the performance of the securities on the liability side of the structure
is the uncertainty regarding the performance of the underlying asset pool. In other words, if we
would be able to predict the economic future of the underlying assets with certainty, we would -
just by applying the structural definitions (cash flows, etc.) - also be able to exactly predict the
performance of securities on the liability side of the transaction.
Mathematically speaking (see Figure 4), the asset side induces a probability space capturing the
randomness of the underlying assets. The cash flow structure of the transaction uniquely defines
a mapping X ~ on this probability space, transforming asset scenarios into liability scenarios. A
liability scenario can be thought of as a vector whose components include all kinds of numbers
relevant for describing the performance of issued securities, e.g., tranche losses, coupons, IRRs,
etc.; see [13], Section 8.3.

3.1 Modeling the asset side of a CDO

The probability space underlying the asset side of a transaction can be constructed by means of a
credit portfolio model. There are various ways to implement a credit risk model suitable for CDO
evaluation. In this paper, we restrict ourselves to a correlated default times model; see Section 4
for an introduction to the mathematics of default times.
The basic idea of how to evaluate a CDO by means of default times is as follows. An asset scenario
in a default times model basically consists of a vector (1 , ..., m ) of default times for a portfolio of,
say, m obligors in the asset pool underlying the CDO. The randomly drawn numbers i represent
1
If the asset pool is managed, the performance of the asset manager is another source of uncertainty.

10
the time until obligor i defaults. Based on an asset scenario, cash flows can be transformed into
default-times-adjusted cash flows. Figure 5 illustrates the cash flow transformation by means of a
bond. Until the default time which depends on the credit quality of the obligor, all cash flows
are in place as scheduled. But at the time of default, cash flows dry-out and the bond investor
receives some recovery amount. Note that in reality most often there will be some delay until the
final settlement/work-out of the defaulted asset. In derivative transactions the time until settlement
most often follows an ISDA master agreement.

Figure 5: Transformation of cash flows by default times

In this way, cash flow CDOs, where the cash collected on the underlying assets is distributed on
the liability side of the structure, can be evaluated. For purely synthetic structures the application
of default times is even more straightforward, because it is primarily loss and default distributions
that matter. More details on default times follow later.

3.2 Modeling the liability side of a CDO

The modeling of the structure of a CDO itself is far from being trivial. First of all it involves a
careful study of the available documentation, starting from presales and ending at a voluminous
offering memorandum. It also involves a careful balance between modeling the full range of all
cash flow elements and simplifying structural elements for the benefit of a better handling. Taking
shortcuts regarding cash flow elements can be dangerous, of course, and has to be done with great
care. To illustrate this, we give two examples.

Example 1: A harmless simplification

Assume that a CDO-SPV issues tranches where one tranche is split and issued in two currencies,
e.g., 40% of the tranche are Euro denominated (Euro LIBOR as reference interest rate) and 60% of
the tranche are Sterling denominated (Sterling LIBOR referenced). Lets say the underlying assets
are all Sterling denominated and (Sterling LIBOR) floating rate notes. Obviously there is some
currency mismatch inherent in the CDO which typically is hedged by means of a basis swap and

11
a currency swap. The good news regarding the CDO model is that as long as the hedges are in
place, there is no need to model the randomness underlying the currency risk. Instead, the hedge
costs can just be implemented as another (senior) deduction from available funds in the interest
waterfall of the CDO.

Example 2: A dangerous simplification

A less harmless shortcut is the following situation. An investor considers buying a mezzanine
tranche of a cash flow CDO. In order to come to a quick decision, the investor only tranches-up
the loss distribution of the underlying pool (see 4.4) in order to get an estimate of the mezzanine
tranches default probability and expected loss. Now, if the CDO is not just a plain vanilla
structure but incorporates some redirection of cash flows based on credit events in the asset pool
(which will be the case in almost all cases of cash flow deals), such a tranched loss approach
(ignoring cash flow triggers and waterfall changes) is very likely to be quite misleading. For more
sophisticated structures at least a semi-analytic approach (see 4.5), or even better, a full Monte
Carlo simulation approach is recommended.

4 CDO modeling techniques

This section will outline the basics of correlated default times. As a reference to research papers
following more or less the same route we mention F INGER [20], L I [26, 27] and S CHMIDT and
WARD [33].
Because default times depend on the credit quality of the considered obligor, well-calibrated credit
curves are a main ingredient for constructing default times. The derivation of such curves is the
topic of the following section.

4.1 Calibration of a credit curve

For reasons of simplicity and to make the applicability clear, we asssume in this section that we
have a set of ratings R =AAA, AA, A, BBB, BB, B, CCC, D, where AAA as always denotes the
best credit quality, CCC refers to the worst non-defaulted credit quality and D denotes the default
state. It is straightforward to find generalizations of the following results for finer rating scales.
It is best practice to assign a unique default probability (PD, in Basel II notation) to every obligor
rating R. Table 1 shows an example of such a PD-calibration of ratings.

Table 1: One-year PDs calibrated to ratings

The aim of this section is to calibrate a credit curve for each of the 7 ratings R, where a credit
curve for a rating R is a mapping
(R)
t 7 pt = P[R * D in time t] (t 0; R {AAA, AA, ..., CCC}),

12
 (R)
where * denotes migration. In other words, pt is the probability that an obligor with a current
rating of R defaults within the next t years. An example for probabilities
(AAA) (AA) (CCC)
pt , pt , ..., pt

for t = 1 is given in Table 1. We will always count time in years.

There are many ways to do this. Here, we follow a well-known Markov chain approach (see
JARROW ET AL . [24], I SRAEL ET AL . [23], K REININ and S IDELNIKOVA [25]), based on a one-
year migration matrix from Standard & Poors ([36], Table 8, Page 13). In practice, the calibration
of credit curves is not as straightforward as exercised here. For example, internal information
on credit migrations and default history will be a main source of data taken into account for the
calibration of credit curves. However, our purpose here is to demonstrate how such a calibration
can be done in principle. For our little exercise, we rely on rating agency data.
(R)
Rating agencies annually publish discrete credit curves (pt )t=1,2,3... based on cohorts of histor-
ically observed default frequencies; [36] is an example. The problem with historically observed
cohortes is that the resulting multi-year PDs have a tendency to look kind of saturated at longer
horizons due to lack-of-data problems. Comparing corporate bond default reports from 5 years ago
with reports as of today, one certainly recognizes a lot of improvement on the data side. Curves
are smoother and do not imply zero forward PDs too early. Nevertheless the data quality is still
not satisfactory enough to rely on historic data without smoothing by some suitable model. The
following Markov chain approach is an elegant way to overcome this problem and to generate
continuous-time credit curves.
Let us now start with the adjusted average one-year migration matrix from S&P (see [36], Table
8, Page 13). We overcome the zero default observation problem for AAA-ratings by replacing the
default column by the values from Table 1. In fact, the PDs in Table 1 have been calibrated based
on a linear regression on a log-scale of the original default column in the S&P-matrix. To assure
that we have a stochastic (migration) matrix, i.e., row sums equal to 1, we renormalize the rows of
the modified S&P-matrix. As a result we obtain the one-year migration matrix M = (m ij )i,j=1,...,8
presented in Table 2. Next, we need the following theorem.

Table 2: Modified S&P average one-year migration matrix

1
Theorem 1 If a migration matrix M = (mij )i,j=1,...,8 is strictly diagonal dominant, i.e., mii > 2
for every i, then the log-expansion
n
X (M I)k
Qn = (1)k+1 (n N)
k=1
k

converges to a matrix Q = (qij )i,j=1,...,8 satisfying

13
 P8
1. j=1 qij = 0 for every i = 1, ..., 8;

2. exp(Q) = M .
The convergence Qn Q is geometrically fast.
Proof. See I SRAEL ET AL . [23]. 2

Remark 1 Recall that the generator of a time-continuous Markov chain is given by a so-called
Q-matrix Q = (qij )1i,j8 satisfying the following properties:
P8
1. j=1 qij = 0 for every i = 1, ..., 8;

2. 0 qii < for every i = 1, ..., 8;

3. qij 0 for all i, j = 1, ..., 8 with i 6= j.

For some background on Markov chains we refer to the book by N ORIS [31].
The following theorem is a standard result from Markov chain theory:

Theorem 2 The following two properties are equivalent for a matrix Q R 88 :

Q satisfies Properties 1 to 3 in Remark 1.

exp(tQ) is a stochastic matrix for every t 0.

Proof. See N ORIS [31], Theorem 2.1.2. 2

Theorem 1, Remark 1 and Theorem 2 open a nice way to construct credit curves which are
compatible with Table 1 at the one-year horizon. We start by calculating the log-expansion
Q = (qij )i,j=1,...,8 of the one-year migration matrix M = (mij )i,j=1,...,8 according to Theorem
1. This can be done with a calculation program like Mathematica or Matlab, but can as easily also
be implemented in Excel/VBA. Table 3 shows the resulting matrix Q.

Table 3: Log-expansion of the modified S&P average one-year migration matrix

Theorem 1 guarantees that Q fulfills Property 1 of generators listed in Remark 1. Condition 2 is

also not hurt by Q, but Condition 3 is not fulfilled. However, there are only three black sheep in
Table 3, namely

qAAA,B = -1bps,

qB,AAA = -1bps,

qCCC,AA = -2bps.

14
Only these three entries disable Q from being a generator matrix. Because these three values are
very small numbers, we feel free to set them equal to zero and decrease the diagonal elements of
rows AAA, B and CCC by an amount compensating for the increased row sums, such that at the
end the row sums are equal to zero again. In [25] this procedure is called a diagonal adjustment.
As a result we obtain a generator matrix Q = (qij )i,j=1,...,8 as shown in Table 4.

Table 4: Approximate generator of the modified S&P average one-year migration matrix

From Theorem 1 we know that we get back the original migration matrix M from Q by exp(Q).
But what about getting M back from Q? Because we manipulated Q in order to arrive at a gen-
erator Q, exp(Q) will not exactly equal M . What we are interested in now is the error coming
from replacing M by exp(Q). Because the manipulation we did is negligible, we already expect
the result of the next proposition.

Proposition 1 M has an approximate Q-matrix representation by Q. The error is

v
u 8
uX
kM exp(Q)k2 = t (mij (exp(Q))ij )2 0.000224
i,j=1

and therefore negligible. We can safely work with Q instead of Q.

Proof. Just calculate the distance. 2

In Markov chain theory the achievement of Q is called an embedding of the time-discrete Markov
chain represented by M into a time-continuous Markov chain represented by its generator or Q-
matrix Q. Of course, our embedding only holds in an approximate manner (see Proposition 1), but
the error is negligibly small. Probabilists know that the existence of such embeddings is far from
being trivial, and to some extent we have been very lucky that it worked so well with the S&P-
based migration matrix M . However, there is more than just one technique to find approximate
generators; see [24], [23] and [25].
We are basically done now with our credit curves. Our efforts have been rewarded by a generator
Q with exp(Q) = M , ignoring the from now on. The credit curves can be read-off from the
collection of matrices (exp(tQ))t0 by looking at the default columns. More precisely, we obtain
(R)
pt = (etQ )i(R),8
where i(R) denotes the transition matrix row corresponding to the given rating R. For the rest of
this paper we will always work with the credit curves just constructed. Figure 6 shows a chart of
our credit curves from t = 0 to t = 50 years (quarterly values).
The curves in Figure 6 are typical. Credit curves assigned to subinvestment grade ratings have a
tendency to slow-down their growth, because conditional on having survived for some time the
chances for further survival improve. For good ratings we see the opposite effect.

15
 (R)
Figure 6: Calibrated curves (pt )t0

4.2 The distribution of single-name default times

(R)
If we believe that our credit curves (pt )t0 are correct and really give us the cumulative default
probabilities for any given rating R over any time interval [0, t], there is one and only one way to
define a default time distribution for an R-rated obligor.
(R)
Proposition 2 Given a credit curve (pt )t0 for a rating R, there exists a unique default times
distribution for R-rated obligors.
(R)
Proof. The proof is obvious: Set FR (t) = pt for t 0. Define a random variable (R) with
values in [0, ) and distribution function F (R) by
(R)
F (R) (t) = P[ (R) t] = FR (t) = pt . (1)
For example, (R) = FR1 (X) with X U [0, 1] will do the job. 2

Proposition 2 shows that as soon as our credit curve is established, there is just one way to come
up with a default time for an obligor admitting such a credit curve. In the sequel, we continue
to use the notation from the proof of Proposition 2. Based on (1), the density of the default time
distribution of R-rated obligors can be obtained by calculating the derivative
d
f (R) (t) = FR (t). (2)
dt
(R)
Figure 7 shows the default time densities w.r.t. our credit curves (p t )t0 for R=AAA, R=AA and
R =A. For (single name) default times we expect in general that

investment grade ratings have a default expectation far in the future, whereas

16
 (R)
Figure 7: Default time densities w.r.t. (pt )t0 for R =AAA,AA,A

subinvestment grade ratings can be expected to default in the near future.

We will illustrate this by calculating the expectation and standard deviation of the default time
distributions calibrated w.r.t. our credit curves. Table 5 reports on our calculation of the mean and
(R)
standard deviation of f ,
Z
E[ (R) ] = tf (R) (t)dt and
0

 Z 21
[ (R) ] = (t E[ (R) ])2 f (R) (t)dt .
0

(R)
Table 5: Mean and standard deviation of default times w.r.t. (pt )t0 (time in years)

The numbers show that if one only considers a single credit risky asset, its expected default time
typically is far away in the future. Even for the worst credit quality, CCC, the expected time
until default is 12 years. However, one should not forget that the chances that a CCC-rated credit
defaults within one year are almost 30%. This is possible because default time distributions are
quite unsymmetric and heavily skewed.

17
 Figure 8: Dependence of default times statistics on ratings

Figure 8 visualizes the dependence of default times means and standard deviations on ratings. It
is quite interesting to observe that in the considered case mean and standard deviation coincide
exactly at the border between investment and subinvestment grade ratings, namely in BBB. The
volatility of default times decreases with decreasing credit quality but increases if considered rela-
tive to the corresponding expected default time.
This concludes our discussion of single-name default times. In the next section we are interested
in default times for a whole portfolio of credits.

4.3 Multivariate default times distributions

Let us now assume that we look at a portfolio of m obligors. We use the index i to refer to the i-th
obligor. The rating assigned to an obligor i will be denoted by R(i). The results from Section 4.2
imply marginal default time distributions for our portfolio. We denote the corresponding densities
and distribution functions by
f1 = f (R(1)) , ..., fm = f (R(m)) and F1 = F (R(1)) , ..., Fm = F (R(m)) .
These are the unique (marginal) default time distributions matching our credit curves, and because
we decided to believe in our credit curves we will fix these distributions for the rest of this section.
What we need now is a method to combine given marginal distributions to a common multivariate
distribution reflecting the dependencies between single default times. And we are lucky: there is
such a concept known in the statistics literature for many years, namely the copula approach. In
fact, the concept of copulas is much older than the history of internal credit risk models. Here in
the context of default times one really finds a good example for a meaningful application of copula
functions. The problem we started with at the beginning of this section is a classical motivation for
the development of copulas. Given m marginal distributions, how can we bind them together to
a reasonable multivariate distribution incorporating the dependencies between the m components?
Copulas are the answer. Therefore, we now briefly mention some key facts.

Definition 1 A copula (function) is a multivariate distribution (function) such that its marginal
distributions are standard uniform. A common notation for copulas is
C : [0, 1]m [0, 1], (u1 , ..., um ) 7 C(u1 , ..., um ) .
In our case m always refers to a number of obligors or credit risky instruments.

18
The following two theorems show that copulas offer a universal tool for constructing and studying
multivariate distributions.

Theorem 3 (S KLAR [34], [35]) For any m-dimensional distribution function F with marginal
distributions F1 , ..., Fm , there exists a copula function C such that
F (x1 , ..., xm ) = C(F1 (x1 ), ..., Fm (xm )) (x1 , ..., xm R).
Moreover, if the marginal distributions F1 , ..., Fm are continuous, then C is unique.
Sketch of Proof. Define the following function on [0, 1]m ,
C(u1 , ..., um ) = F (F11 (u1 ), ..., Fm1 (um )) . (3)
Then one only has to verify that C is a copula representing F , see also N ELSON [29]. 2

But more is true. It is not only the case that any multivariate distribution has a copula representa-
tion, but copulas can be combined with given marginal distributions.

Proposition 3 Given a copula C and distribution functions F1 , ..., Fm on R, the function

F (x1 , ..., xm ) = C(F1 (x1 ), ..., Fm (xm )) (x1 , ..., xm R)
defines a multivariate distribution function with marginal distributions F 1 , ..., Fm .
Proof. The proof is straightforward. 2

It is Proposition 3 we can use to construct multivariate default time distributions. We already found
the true marginal distribution functions F1 , ..., Fm reflecting our credit curves. Choosing a copula
function C will yield a multivariate default times distribution
F (t1 , ..., tm ) = C(F1 (t1 ), ..., Fm (tm )) (4)
for times t1 , ..., tm [0, ). So the concept of copulas is a nice tool for constructing multivariate
default times distributions. Obviously the challenge now is to find a suitable copula best matching
the modeling problem and the underlying data.

Example: Gaussian and Student-t copulas

The most commonly used copula in the context of default times is the Gaussian copula,
C[u1 , ..., um ] = F m, [N 1 (u1 ), ..., N 1 (um )] (5)
where N [] denotes the standard normal distribution function, N 1 () denotes its inverse and
F m, refers to the multivariate Gaussian distribution function on Rm with correlation matrix
= (%ij )1i,jm and zero mean. The parameter m refers to the number of obligors in the con-
sidered portfolio. Gaussian copulas and their application to CDOs have been studied in various
papers, see [20], [26], [33], just to mention a few already quoted research papers. One safely can
say that whenever people are not explicitly addressing the problem of copula selection, they most
often rely on a Gaussian copula.
However, there are many alternatives. The non-Gaussian copula applied most often definitely is
the Student-t copula, defined by
1 1
C[u1 , ..., um ] = F t(n,) [Ft(n) (u1 ), ..., Ft(n) (um )] (6)

19
where F t(n,) denotes the multivariate t-distribution function with n degrees of freedom and (lin-
ear) correlation matrix Rmm , and Ft(n) denotes the t-distribution function with n degrees
of freedom. Note that the t(n, )-distribution can be derived by choosing a Gaussian vector
Y = (Y1 , ..., Ym ) Np (0, ), a 2 (n)-distributed random variable X, independent of Y , and
considering the product n/X Y , which will be t(n, )-distributed.
There is a whole universe of copulas available in the stochastics literature; see for example E M -
BRECHTS ET AL . [17] and B OUY E ET AL . [15]. Making the right choice of copula is far from
being trivial. In the next section, we briefly indicate what difference the choice of copulas makes
to a CDO model; see also Table 7 and Figure 12.

Copula impact on CDO performance: Gaussian versus Student-t

If the degrees of freedom are large enough, the difference between Gaussian und t-copulas will
vanish if they rely on the same linear correlation matrix . This is due to the fact that for large n
t(n) is approximately normal. Moreover, the multivariate t-distribution inherits its linear correla-
tion from the correlation matrix of the involved Gaussian vector,
p p
Corr[ n/X Yi , n/X Yj ] = Corr[Yi , Yj ]

using the same notation as in the previous section. But if one decreases the degrees of freedom,
the t-copula will show more and more tail dependency; see E MBRECHTS ET AL . [17], F REY and
M C N EIL [18], and [13], pages 106-112. This is graphically illustrated in Figure 9. In this example,
the chosen linear correlation equals % = 0.4.

Figure 9: Increasing tail dependency by decreasing the degrees of freedom

For CDO modeling, this insight has important consequences. More tail dependency typically
means a higher potential for joint defaults which in turn implies higher stress for senior tranches
in CDO transactions. In other words, changing the copula used for our default times model means
changing the economics of different tranches in different ways; see Section 5. On one side this
observation shows that the choice of a particular copula function induces a high model risk, but
on the other side it provides a very useful tool for stress testing different CDO tranches. Stress
testing is a must in CDO modeling anyway, because most often there is a significant amount of
uncertainty on the parameter side involved. It is very natural to make copulas a part of the stress
testing program for CDO tranche investments.

20
A remark on asset value models and default times
It is possible to calibrate a default times model compatible in distribution with a one-period asset
value model (e.g., period = one year) according to B LACK and S CHOLES [12] and M ERTON [28];
see [20], [26], [33]. The basic idea is as follows. We start with geometric Brownian motions
(i) (i) (i)
At = A0 exp[(i 21 i2 )t + i Bt ] (i = 1, ..., m; t [0, T ]) (7)
where the Brownian motions are correlated and admit a correlation matrix = (% ij )1i,jm con-
stant over time. The processes (At )t0 are interpreted to give the asset value of the corresponding
obligors at any time. Asset value processes can not be observed in the market, but can be inferred
from equity processes, see C ROSBIE [16], N ICKEL ET AL . [30], and [13], 3.4. The link to default
risk is given by a latent variables approach by means of a Bernoulli mixture model, where joint
default probabilities can be written in the following general form; see J OE [21], F REY and M C
N EIL [18], and [13], 2.1. For 1 , ..., m {0, 1},
Z Y m
P[L1 = 1 , ..., Lm = m ] = pi i (1 pi )1i dF(p1 , ..., pm ).
i=1
[0,1]m

Here, the complexity of the model is completely hidden in the distribution function F. For example,
F in its most simple form is applied in so-called uniform portfolio models, where all obligors are
correlated the same way, admitting a uniform default probability. In this case (see F INGER [19],
VASICEK [37], and [13], 2.1.2), the distribution function F is given by
h N 1 (PD) % y i
1
F = N g , g(y) = N (y R), (8)
1%
where N [] denotes the distribution function of the standard normal distribution, % refers to the uni-
form asset correlation between obligors, and PD denotes the (one-period, e.g., one year) uniform
default probability of the portfolio. The probability for k out of m defaults can then be written as
  Z
m
P[L1 + + Lm = k] = g(y)k (1 g(y))mk dN (y).
k

Note that the function g(y) from Formula (8) represents the heart of the Basel II benchmark risk
weights in the new capital accord.
Now fix some horizon T > 0. Coming back to the general case of geometric Brownian asset value
processes, the Bernoulli variable indicating default of obligor i over time T typically is defined by
(T )
Li = 1{A(i) c(i) } (9)
T T

(i)
where cT denotes the default-critical threshold for obligor i, the so-called default point; see [16].
(R)
The link to our credit curves (pt )t0 from Section 4.1 is given by a calibration of the default
(i)
point cT in a way reflecting the T -year cumulative default probability,
(T ) (i) (i) ! (i)
P[Li = 1] = P[AT cT ] = pT (10)
(i) (R(i))
where pT = pT . Based on a simple reformulation of (7), we obtain the distributional equation
(i) (i)
(T ) (i) (i) ln(cT /A0 ) (i 12 i2 )T
P[Li = 1] = P[B (i) cT ] with cT = . (11)
i T

21
 (i)
where B (i) B1 N (0, 1). Note that conceptually B (i) has nothing but the standard normal
(i)
distribution in common with the value B1 of the driving Brownian motion. The time-dynamics
of the underlying process is not taken into account in this fixed-horizon approach. However, due
to B (i) N (0, 1) and Equations (10) and (11) we can conclude that
(i) (i)
cT = N 1 (pT ) = N 1 (Fi (T )) (12)

for the default point of obligor i, where (Fi (t))t0 denotes the credit curve of that obligor. The
second = in (10) additionally forces that at the horizon T
(T ) (i)
P[Li = 1] = pT = P[i T ] with i Fi (13)

where Fi denotes the default time distribution function of obligor i; see Section 4.2. We have
(T )
P[i T ] = P[Li = 1] = P[B (i) N 1 (Fi (T ))] = P[Fi1 (N [B (i) ]) T ].

This motivates the current standard approach to correlated default times, defining

i = Fi1 (N [B (i) ]) (i = 1, ..., m) where (B (1) , ..., B (m) ) N (0, ) (14)

is multivariate Gaussian with correlation matrix ; see [20], [26], [13], 7.3. By construction we
have i i , so (14) shows a way to define the default time of obligor i as a function of a variable
(i)
in distribution equal to the (standardized) one-year asset value log-return B 1 of obligor i. Natu-
rally, default times defined according to Equation (14) will inherit the dependence structure of the
involved one-period asset value model.
Summarizing, (14) yields correlated default times marginally matching given credit curves and
inheriting the dependence structure of a given one-period asset value model.
Moreover, the cumulative default distribution up to time T arising from the default times approach
and the distribution of the portfolio defaults arising from a one-period (more precisely, the period
is the time interval [0, T ]) asset value model coincide,
m m
(T )
X X
1{i T } Li
i=1 i=1

for any fixed horizon T . Note that this relation refers to equality in distribution only.
However, in order to define correlated default times in line with (14), it is not really necessary
to think in terms of an asset value model, because they simply can be derived by combining our
single default times distributions from Section 4.2 by means of a Gaussian copula.
More precisely, combining (marginal) default time distribution functions F 1 , ..., Fm by means of a
Gaussian copula according to (5), we get a multivariate default times distribution function

F (t1 , ..., tm ) = F m, [N 1 (F1 (t1 )), ..., N 1 (Fm (tm ))].

For the corresponding default times 1 , ..., m we derive at the following equation,

P[1 t1 , ..., m tm ] = F m, [N 1 (F1 (t1 )), ..., N 1 (Fm (tm ))]

= P[X1 N 1 (F1 (t1 )), ..., Xm N 1 (Fm (tm ))]

= P[F1 1
1 (N [X1 ]) t1 , ..., Fm (N [Xm ]) tm ] where (X1 , ..., Xm ) N (0, ).

22
Going back to (14) we find that (1 , ..., m ) (1 , ..., m ).
Altogether the derivation of default times according to (14) explains why marginal default time
distributions are often combined to a multivariate default times distribution by use of a Gaussian
copula. The reason is that a one-period asset value model in the classical setting as introduced
by M ERTON matches a default times model relying on a Gaussian copula. However, in the same
way as other processes than Brownian motions are used in option pricing theory today, other than
Gaussian copulas become more and more popular for default times distributions. The dependence
structure of a non-normal asset value model then can be carried over to the default times copula in
an analogous way as exercised for the Gaussian copula case in (14).

Barrier diffusion models and first passage times

In this section we briefly want to mention other approaches to default times by means of barrier
models in order to give the reader some motivation to investigate other promising ideas applicable
to CDO analysis. A more detailled exposition and discussion of the results indicated in this section
can be found in [14].
For example, OVERBECK and S CHMIDT [32] start with (marginal) default time distributions in a
similar way as we did in 4.2 by writing
P[i < t] = Fi (t) (t 0; i = 1, ..., m). (15)
Hereby, the distribution functions Fi (t) are given from external sources, e.g., credit curves as in
Section 4.1. Additionally they define the pairwise joint default probabilities (JDP) w.r.t. a fixed
horizon T by
JDPij = P[i < T, j < T ]. (16)
The problem they study is the following:
(i)
Problem 1 Given F1 , ..., Fm and (JDPij )1i,jm , find stochastic processes (Xt )t0 and barriers
(i)
(ct )t0 , i = 1, ..., m, such that
(i) (i)
i = inf{t 0 : Xt ct },
where i means the default time of obligor i satisfying (15) and (16).
Problem 1 searches for a first passage or hitting time matching the default time of obligors ad-
mitting a prescribed credit curve and prescribed default event correlations, because the JDPs are
directly related to default correlations and vice versa; see [13], pages 58-61.
The key idea elaborated in [32] for tackling Problem 1 is to define a time scale transformation T i
(i)
for every asset and to transform correlated Brownian motions (Bt )t0 ,
(i) (i)
Xt = B Ti (t) (t 0; i = 1, ..., m),
(i)
then applied for the definition of first passage times w.r.t. barriers (c t )t0 ,
(i) (i)
i = inf{t 0 : Xt ct }. (17)
(i)
The processes (Xt )t0 are called ability-to-pay processes. In [32], Proposition 1, they define
a suitable time scale transformation Ti such that i coincides with the default time i defined by

23
means of the prescribed credit curves (15). The proof is based on the well-known first passage time
distribution for Brownian motion. In a second part of the paper ([32], 4.2) it is shown that the time
changed Brownian motions can be calibrated w.r.t. given JDPs. Based on a paper by Z HOU [38],
the JDPs can be described in closed analytic form as a function of the correlation of the underlying
Brownian motions.
The results in [32] provide an intuitive way (time-transforming Wiener processes) to derive corre-
lated default times by means of a barrier diffusion model. In H ULL and W HITE [22], a comparable
analysis can be found. Regarding default barrier models we refer to A LBANESE ET AL . [6] and
AVELLANEDA and Z HU [8].

4.4 Analytic approximations

For some transactions, analytic or semi-analytic approximations can be applied in order to speed-
up the evaluation of CDOs. Whether analytic shortcuts can be applied or not strongly depends on
the structure of the transaction and the underlying asset pool.

Analytic derivation of expected loss and PD of a tranche

The most typical example where an analytic approach is as good as any Monte-Carlo simulation
approach is the case of a synthetic (balance sheet motivated) transaction for the purpose of regu-
latory capital relief and risk transfer, referenced to a large homogeneous pool of reference assets,
e.g., a large portfolio of retail loans or a highly diversified portfolio of SMEs. More precisely, for
the sequel we consider (as an example) a structure satisfying the following conditions:
1. The underlying reference pool is highly diversified and can be (approximately) represented
by a uniform reference portfolio with infinite granularity.

2. Amortization of notes on the liability side follows sequentially top-down in decreasing order
of seniority (highest seniority tranche first, second highest seniority tranche next, and so on).

3. Losses are written-off sequentially bottom-up in increasing order of seniority (equity tranche
bears the first loss, first junior piece bears the second loss, and so on).

4. CDO notes are referenced to the underlying pool of assets (e.g., by credit linked notes).
Interest payments on notes will be paid by the originator in an amount of

Interest(Tranche) = Volume(Tranche) [LIBOR + Spread(Tranche)]. (18)

Besides loss write-offs (bottom-up; see Condition 3), no other random events trigger repay-
ments or interest payments.
Condition 4 implies that a default on interest payment obligations can only happen if the originator
defaults on its financial/contractual obligation, because the promise made to notes investors is to
pay interest according to (18) unconditionally. However, due to loss write-offs, the nominal amount
of interest payments can decrease if a tranches volume decreases as a consequence of losses.
For the sequel we assume without loss of generality a fixed LGD of 100%. By scaling and a
substitution in the respective formulas, any fixed LGD can be implemented, replacing any gross
loss L by a realized net loss L LGD. As another simplifying condition, we assume that the
asset pool is static (non-managed) and has a bullet exposure profile until maturity. However, by

24
WAL-adjustments (WAL stands for weighted average life) the results of this section also apply
to amortizing asset pools.
For transactions as the one described above, the risk of tranches can be quantified by a closed-
form analytic approach. Condition 1 allows to replace the original reference pool by a uniform
or homogeneous portfolio with infinite granularity admitting a uniform credit curve (p t )t0 as
calibrated in Section 4.1 and a uniform asset correlation %. Let us assume that the considered CDO
matures at time T . Then, the cumulative loss L at the horizon T is given by (cp. (8))
h N 1 (p ) % Y i
T
L = p(T, Y ) = N where Y N (0, 1). (19)
1%
The derivation of this representation is well-known, due to VASICEK [37], and can be found in
a more general setting, e.g., in [13], pages 87-94. The expression on the right side in equation
(19) is the loss variable of a portfolio with infinitely many obligors (limit case) where all obligors
have a PD of pT and are pairwise correlated with an asset correlation %. The variable Y has
an interpretation as a macro-economic factor driving the loss of the portfolio. Because p(T, Y )
corresponds to a portfolio of infinitely many assets, idiosyncratic risk has been completely removed
by diversification, so that the randomness of Y is the sole source of the riskyness of the portfolio
loss p(T, Y ).
In the sequel we will exploit the absolute continuity of p(T, Y ) by relying on its density
r  2 
1% 1 1

2 1  1 p 1
fpT ,% (x) = exp N (x) N (pT ) 1 % N (x) (20)
% 2 2%
see [13], page 91. By construction, we get back the credit curve (p t )t0 by calculating expectations,
Z h 1 Z1
N (pt ) % y i
pt = E[p(t, Y )] = N dN (y) = xfpt ,% (x)dx,
1%
0

assuming that the severity of loss equals 100% in case of defaults.

Now, any tranching consisting of q tranches on the liability side of the CDO can be written as a
partition of [0, 1) in the following way,
i = [i , i+1 ) (i = 1, ..., q; 0 = 1 < 2 < < q+1 = 1).
Hereby we assume losses to be normalized (in percentage) to the unit interval (0 = no loss, 1 = full
loss of the portfolios total notional amount). Given the simple cash flow structure assumed at the
beginning of this section, the loss Li of tranche i is given by
Li = i (L) = min[max(0, L i ), i+1 i ] (i = 1, ..., q) (21)
where L represents the portfolio loss at time T , L = p(T, Y ).

Proposition 4 For a CDO with maturity T satisfying the conditions listed at the beginning of this
section, the expected loss of tranche i , normalized to the tranche size, can be calculated by
Z1
1
E[Li ] = i (x)fpT ,% (x)dx
|i |
0

where i is the function defined in (21) and |i | = i+1 i .

25
 Figure 10: Loss profile function of a CDO tranche

Proof. The assertion of the proposition is obvious. 2

Proposition 4 offers a closed-form expression for the expected loss of a tranche, illustrated by
Figure 10. The hitting probability of tranches also can be expressed in closed form.

Proposition 5 Under the stated conditions, the probability i that tranche i is hit by a loss equals
h 1  p i
i = 1 N N 1 (i ) 1 % N 1 (pT ) .
%
Remember that i denotes the lower boundary of tranche i .
Proof. First of all note that we have

h N 1 (x) 1 % N 1 (pT ) i
P[p(T, Y ) x] = P Y
%
for all x [0, 1] according to (19). Taking Y N (0, 1) into account and considering
i = P[p(T, Y ) > i ] = 1 P[p(T, Y ) i ],
the proof of the proposition follows. 2.

Proposition 5 offers a way to calculate the PD of a CDO tranche and Proposition 4 helps calculating
its expected loss (EL). Then, the loss given default (LGD) can be defined and calculated by
E[Li ]
LGDi = (22)
i
for any tranche i , i = 1, ..., q. In this way, the three main components of basic risk analysis (PD,
EL and LGD) are fully specified. Another interesting application of Proposition 5 is the derivation
of a model-based implied rating of a CDO tranche.

26
An illustrative example
We conclude this section by an example applying the results just derived. Let us consider a CLO
satisfying the assumptions stated at the beginning of Section 4.4. The assumed tranching of the
CLO is reported in Figure 11. The maturity of the CLO is in 10 years, counted from today on.

Figure 11: Tranching of a portfolios loss distribution

Let us assume that the underlying reference pool (static, bullet profile) can be (approximately)
replaced by a uniform portfolio with a BBB-credit curve (pt )t0 , see Figure 6, and a uniform asset
correlation % = 20%. For T = 10 we then obtain p1 = 36bps and pT = 9.8%. The cumulative
loss distribution for pT , % and an assumed fixed LGD of 60% (i.e., an overall recovery of 40%) is
plotted in Figure 11. The cumulative EL of the portfolio equals 5.88%.
We now apply Propositions 4 and 5 to our sample transaction and obtain Table 6 as a result. Here
are some comments:
1. Equity tranche: In the analytic approach, the PD of the equity tranche typically is 100%
because the loss distribution is absolutely continuous, implying that P[L = 0] = 0.

2. Rule-of-thumb for LGDs: In general, the LGD of a tranche is lower for thick tranches and
higher for thin tranches. To illustrate this, assume that a tranche consists of one point only,
say, lower and upper boundaries equal some (0, 1). Then, as soon as a loss hits the
tranche, it will completely be wiped out the same moment. Therefore, its LGD is 100%.
Now set = , keep as the lower boundary of the tranche fixed, but increase as the
upper boundary of the tranche. The higher , the longer it takes for losses to eat into the
tranche before they have consumed all of the tranches capital. The LGDs rule-of-thumb is
reflected by the results in Table 6.

27
 3. Super senior tranche: The cumulative PD of the super senior tranche equals 6.15%. This
is very high for a super senior swap and due to the illustrative character of the example. In
typical real life transactions we considered, the super senior swaps PD never exceeded a
few basispoints. However, even in our illustrative sample tranching one observes that the
cumulative 10-year EL is quite low, also reflected by the small LGD of the tranche. A linear
annualization of the EL would yield an annual EL of about 3bps.

Table 6: Result of applying Propositions 4 and 5 to our example

Obviously, the overall expected loss of the portfolio can be obtained by calculating
5
X
Volume(Tranchei ) EL(Tranchei ),
i=1

Indeed, doing the calculation yields the portfolios cumulative EL of 5.88%. In other words, the
portfolios EL has been allocated to CDO tranches in an additive way as expected.

4.5 Semi-analytic techniques

We now come to a semi-analytic approach applicable to a much broader class of transactions than
the purely analytic approach explained in the previous section. Here, we

only make the assumption that the underlying reference pool is highly diversified and can be
approximately represented by a synthetic homogeneous reference pool,

but allow for all kinds of cash flow elements in the structural definition of the CDO.

In such cases, the semi-analytic technique is a powerful tool to quickly evaluate a CDO.
The approach works as follows. Instead of considering the default times i of single obligors, we
consider the fraction of obligors with a default time within the considered payment period. To make
this precise, denote by i the default time of obligor i and by L(m) the cumulative loss for a portfolio
of m obligors over quarterly payment periods 1, ..., T , where T refers to the maturity of the CDO.
The exposure outstanding on loan i in period/quarter j will be denoted by E i,j . We assume the
following natural conditions, considering an increasing number of obligors in the portfolio:

1. The exposures in the portfolio do not increase over time, i.e., E i,j1 Ei,j for all i =
1, ..., m, j = 2, ..., T , and m N, m .

2. The total exposure at time j,

m
(m)
X
Ej = Ei,j ,
i=1

28
 converges to a limit relative to the portfolios start exposure,

(m)
Ej
lim (m)
= wj
m E1

for every fixed payment period j = 1, ..., T . Due to Condition 1, wj [0, 1].

3. With increasing number of obligors, the total exposure of the portfolio strictly increases to
(m)
infinity for all fixed payment periods j, i.e., Ej for m for every j = 1, ..., T .

4. With increasing number of obligors, exposure weights shrink very rapidly,


X Em,j 2
(m)
<
m=1 Ej

for every payment period j = 1, ..., T .

These conditions are sufficient but obviously not necessary for establishing the following results.
More relaxed conditions are easy to formulate. Here we only illustrate the basic principle.
An example for exposures satisfying Condition 2 is the case of uniform amortization profiles,
Ei,j
1 = w1 w2 w3 wT 0 i, j : = wj . (23)
Ei,1
To give an example, this condition is fulfilled in collateralized swap obligations where on the asset
side protection is sold for m single reference names, typically in form of a 5-year bullet profile of
m (equal amount) CDSs, and on the liability side protection is bought on the (diversified) pool of
CDS in form of tranched securities with a suitable leverage regarding spreads on volumes. In this
particular case, wj = 1 for all j.
An example for exposures fulfilling Conditions 3 and 4 is the case where the exposures are captured
in a band, 0 < a Ei,j b < for all i, j. Then, we get for Condition 3,
m
(m)
X
Ej = Ei,j m a ,
i=1

and for Condition 4,


X Em,j 2 X b2 b2 X 1
(m)
2 a2
= 2 2
< .
m=1 Ej m=1
m a m=1
m

Conditions 1-4 are not really restrictive and will be satisfied in most cases.
We fix Conditions 1-4 for the rest of this section. Assuming for a moment an LGD of 100% and
zero collateral, the percentage cumulative loss for an m-obligor portfolio is
T (m) m
(m)
X Ej (m) (m)
X (m)
L = (m)
Xj , with Xj = wi,j 1{j14i <j} (24)
j=1 E1 i=1

29
 (m)
where wi,j denotes the exposure weight for obligor i in payment period j,

(m) Ei,j
wi,j = (m)
(i = 1, ..., m; j = 1, ..., T ).
Ej
Note that here we write 4 instead of because we consider quarterly payment periods w.r.t. a
time variable t counting in years. A generalization to other payment frequencies is straightforward.
Conditions 3 and 4 in the above list are essentially cloned from [13], Assumption 2.5.2, in order
to establish Proposition 7. But before, we need to state

Proposition 6 For a homogeneous portfolio with a credit curve (pt )t0 and a uniform asset corre-
lation %, the probability that obligor i defaults in the time period [s, t) with s < t conditional on
Y = y is given by
h N 1 (p ) % y i h N 1 (p ) % y i
t s
P[s i < t | Y = y] = N N .
1% 1%
Proof. From Equation (19) we conclude
h N 1 (p ) % y i
t
P[i < t | Y = y] = N .
1%
This immediately implies the assertion of the proposition. 2

Proposition 7 Under the conditions of this section, for a homogeneous portfolio with a credit
curve (pt )t0 and uniform asset correlation % we obtain
h h  h N 1 (p j ) % Y i h N 1 (p j1 ) % Y ii i
(m)
P lim Xj N 4 N 4 = 0 = 1,
m 1% 1%
where Y N (0, 1). Recall that in this section we consider quarterly payment periods.
Proof. The proof is a straightforward modification of the argument provided in [13], pages 88-89,
but for the convenience of the reader we provide the argument. The usual trick to prove such
results is to condition on the factor Y . We write Py = P[ |Y =y] for the conditional probability
measures. Fix y R. Then, the random variables

Zi,j = Ei,j 1{j14i <j} E[Ei,j 1{j14i <j} | Y ] (i = 1, ..., m; j = 1, ..., T )

(k)
are i.i.d. w.r.t. Py and centered. The sequence (Ej )k=1,2,... is strictly increasing to infinity due to
Condition 3 for any j = 1, ..., T . Moreover,

X 1 2
X 1 2
X Ek,j 2
(k)
Ey [Zk,j ] (k)
4 Ek,j = 4 (k)
< .
k=1 (Ej )2 k=1 (Ej )2 k=1 Ej
due to Condition 4. Then a version of the strong law of large numbers based on Kroneckers
Lemma (see, e.g., [9]) implies that
m
1 X
lim (m)
Zi,j = 0 Py -almost surely.
m Ej i=1

30
From this we conlude for every y R
(m) (m)
P[ lim (Xj E[Xj | Y ]) = 0 | Y = y] = 1.
m

Then, to prove almost sure convergence is straightforward by writing

(m) (m)
P[ lim (Xj E[Xj | Y ]) = 0] =
m
Z
(m) (m)
= P[ lim (Xj E[Xj | Y ]) = 0 | Y = y]dN (y) = 1.
m

(m)
Proposition 6 implies that the conditional expectation E[Xj | Y ] for Y = y equals
m
(m) 1 X
E[Xj | Y = y] = (m)
Ei,j E[1{j14i <j} | Y = y] =
Ej i=1

hN 1 h N 1 (p j1 ) % y i
hj 1 j i (p j ) % y i
P i < | Y = y = N 4 N 4 .
4 4 1% 1%
This completes the proof of the proposition. 2

Note that the idea underlying the proof of Proposition 7 does not rely on Y N (0, 1). The
same argument could be used to establish an analogous convergence result for other than normal
distributions. We now come to the final result of this section.

Corollary 1 Under the conditions stated in this section, we have

h  T
X  i
P lim L(m) wj gj (Y ) = 0 = 1,
m
j=1

where the functions gj () are defined on R by

h N 1 (p j ) % y i h N 1 (p j1 ) % y i
gj (y) = N 4 N 4 .
1% 1%
The numbers wj refer to the limit exposure weights from Condition 2.
Proof. The assertion follows from the previous two propositions and Condition 2. 2

The semi-analytic technique can be further developed and refined in practice in several directions,
for example, stochastic recoveries could be implemented into the framework quite easily; see [13],
pages 86-89, for an approach for a single-period model, which can be extended to a multi-period
approach in a straightforward manner.
From a CDO modeling point of view, the tool developed in this section is quite powerful, because
in contrast to the purely analytic approach explained in Section 4.4, the semi-analytic approach
allows for the implementation of all relevant cash flow elements, e.g., redirection of cash flows
due to realized losses or other triggers affecting the performance of CDO notes. This flexibility
is a consequence of considering every single payment period, such that all the specialties of the
considered waterfall can be implemented in an accurate manner.
In the next section we apply the semi-analytic technique to an illustrative sample transaction.

31
5 Further examples and applications
We conclude this paper by two examples. Here, we keep the examples as simple as possible and
do not exercise the modeling of some complicated structure. More sophisticated examples and
illustrations can be found in [14].

An example of an evaluation of a collateralized swap obligation

The first example is a plain-vanilla collateralized swap obligation (CSO) with 5 tranches. The
transaction is assumed to work as follows:

The issuer sells protection by means of 80 single credit default swaps (CDS) with a total
swap volume of 800 million EUR (10 million swap volume on each of the 80 names). This
constitutes the asset side of the transaction.

The issuer also buys protection on the 800 million credit volume he now is exposed to. This
builds-up the liability side of the structure.

For buying protection, the vehicle issues 4 credit-linked notes (CLNs), namely class A, class
B, class C and equity. The total volume of notes issued in the capital market equals 100
million EUR (12.5% on the swap volume of 800 million). The cash received from issuing
the CLNs is invested in an account of risk-free (cash-equivalent) collateral.

For the upper 87.5% (700 million), the issuers enters into a super senior swap agreement
with an OECD bank in order to buy protection against tail events.

In case of a default in the CDS pool, the realized loss is paid on the protection selling
agreement by liquidating collateral and using the proceeds to make the contingent payments.
If losses exceed the 100 million funded volume, the super senior swap counterparty will have
to pay for the residual losses not already covered by the available collateral. On the liability
side, losses eat into the tranches bottom-up in a way that the total funded volume always
matches the amount in the risk-free collateral account. Recovered amounts are reflected on
the liability side by a top-down deleveraging of the outstanding swap volume.

The average rating of the CDS pool is a BBB+, the average PD equals 20bps, the assumed
recovery is 34% and the transaction matures in 5 years.

We now calculate for this transaction the PD, EL and LGD. Table 7 shows the result of the Monte
Carlo simulation. For the calculation, we used a correlated default times approach based on a
normal copula as well as a t-copula with

an average asset correlation of 20% (linear correlation) in both cases and

5 degrees of freedom in the t-copula case;

see Section 4.3 as a reference for copulas and default times. Figure 12 illustrates the impact of a
copula change; the result could already be guessed from the earlier presented Figure 9.
Because the t-copula generates a much stronger tail dependency (see Figure 9) for the joint distri-
bution of marginal default times (calibrated from our credit curves from Section 4.1), joint defaults
occur more often in the t-copula than in the Gaussian copula case.

32
 Table 7: A sample CSO (illustrative!)

Figure 12: Graphical illustration of PDs (left) and ELs (right) from Table 7

Table 7 and Chart 12 illustrate how the change from a Gaussian to a t-copula with suitably low
degrees of freedom stresses senior tranches and - at the same time - implies kind of a risk relief
for the most junior tranche. This is a comparable result to the well-known fact that, e.g., assuming
zero correlation for a first-to-default basket is a conservative approach; see Figure 14.
The higher the correlation and tail dependency, the better for equity investors and first-to-default
takers. The lower the correlation and the more independent the occurrence of joint defaults, the
higher the risk of taking the first loss of a basket or pool.
We conclude our example by modeling the transaction also by means of the (semi-)analytic 2 ap-
proach. Hereby we rely on a Gaussian copula, but other copulas can be implemented easily. Table
8 shows the result.
2
Going back to Corollary 1, we see that if wj = 1 j then the semi-analytic approach and the analytic approach are
essentially the same for a fixed horizon T . The difference between analytic and semi-analytic approaches in this case
is that the telescope sum representation of the cumulative loss according to Corollary 1 allows for all kinds of cash
flow adjustments in every single payment period. Obviously, if the pools exposure profile is not of bullet-type but
follows a certain amortization schedule, then the semi-analytic approach and the analytic approach diverge, although
in some cases a suitable WAL-adjustment of the analytic approach has some chance to yield good approximations.

33
 Table 8: Revisiting the transaction by means of a semi-analytic approach

Here are some comments on the result:

The PD for equity equals 100%, reflecting the absolute continuity of the analytic loss distri-
bution. We already discussed this phenomenon in the illustrative example in Section 4.4.

The differences we see are essentially due to the difference of a simulation of 80 non-
homogeneous single default times compared to a (semi-)analytic approach based on the
assumption of an infinitely granular and homogeneous pool of assets. However, if we would
increase the number of assets and take care that the names underlying the CDSs can be mod-
eled by application of a uniform PD and a uniform correlation, then the two results (Table 7
and 8) would converge towards a common risk profile of tranches.

In [14] we also consider pricing and return aspects of a sample CDO. In transactions where, e.g.,
an excess spread redirection trigger (or, to mention another example, principal deficiency ledgers)
are implemented, the semi-analytic approach really unfolds its strengths.

The importance of a reasonable modeling of the timing of defaults

In this section we want to make the point that an accurate modeling of the timing of defaults can
be quite essential. To give an example, we present Figure 13.

Figure 13: Illustration of the importance of an accurate modeling of default timing

34
The chart shows a scatterplot of loss (x-axis) versus cumulative excess spread (y-axis) paid to
equity investors. The underlying transaction has 5 tranches: 4 funded tranches (equity, classes C,
B, A) and an unfunded super senior tranche. The loss axis in Figure 13 refers to losses cumulating
on classes C, B and A. Based on subordination, any loss eating into class C implies that equity
investors already lost all their invested capital. It is interesting to observe the following:

There are scenarios involving a large loss and nevertheless a substantially high excess spread
cumulation to subordinated note holders. The reason for such scenarios typically is a more
backloaded default timing.

There are also scenarios with a comparably small loss but nevertheless a negligibly small
excess spread cumulation to equity investors. Such scenarios are generated by a more front-
loaded default timing.

The density of points decreases with increasing seniority of tranches. This reflects the shape
of the loss distribution of the underlying asset pool. However, it is worthwhile to mention
that losses located at the upper boundary of the C-tranche admit the whole spectrum of
excess spread scenarios, ranging from zero to almost maximum excess spread cumulation to
equity holders. Mezzanine tranches bear the risk of the second loss but have no participation
in the excess spread of the transaction. Therefore they need an extra careful evaluation and
risk/return assessment.

Altogether one clearly can see that an accurate modeling of default timing is essential for pric-
ing a CDO tranche. To give another example in this direction, Figure 14 shows first-to-default
distributions for the transaction described in Table 7.

Figure 14: First-to-default distributions

What we see here is that a change of copula as well as a change of linear correlation changes
the time dynamics of default times. A typical application of default timing considerations are
the tailor-made definition of coverage ratios in cash flow CDOs. The more precise an ar-
ranger/originator can predict the timing of defaults, the more target-oriented the structuring of
the deal can be done.

35
References
[1] Artemus Strategic Asian Credit Fund Limited; FitchRatings, Structured Finance, Synthet-
ics/Asia New Issue Report, March 5 (2003)

[2] Artemus Strategic Asian Credit Fund Limited; Synthetic Collateralized Debt Obligation;
Moodys Investors Service, Internation Structured Finance, New Issue Report, March 13 (2003)

[3] New Issue: Artemus Strategic Asian Credit Fund Ltd.; Standard & Poors, Structured Finance,
March 12 (2003)

[4] Building Comfort 2002-1; FitchRatings, Structured Finance, RMBS/Germany Presale Report,
November 8 (2002)

[5] Building Comfort 2002-1; Standard & Poors, Structured Finance, RMBS Presale Report,
November 18 (2002)

[6] A LBANESE , C., C AMPOLIETI , G., C HEN , O., Z AVIDONOV, A.; Credit Barrier Models; Risk
16 (6), 109-113 (2003)

[7] A RTZNER , P., D ELBAEN , F., E BER , J., H EATH , D.; Coherent Measures of Risk; Mathemat-
ical Finance 9 (3), 203-228 (1999)

[8] AVELLANEDA , M., Z HU , J.; Distance to Default; Risk 14 (12), 125-129 (2001)

[9] BAUER , H.; Probability Theory; de Gruyter (1996)

[10] B ELKIN , B., S UCHOWER , S., F OREST, L. R. J R .; The effect of systematic credit risk on
loan portfolio value-at-risk and loan pricing; CreditMetrics Monitor, Third Quarter (1998)
[11] B ELKIN , B., S UCHOWER , S., F OREST, L. R. J R .; A one-parameter representation of credit
risk and transition matrices; CreditMetricsT M Monitor, Third Quarter (1998)
[12] B LACK , F., S CHOLES , M.; The Pricing of Options and Corporate Liabilities; Journal of
Political Economy 81, 637-654 (1973)

[13] B LUHM , C., OVERBECK , L., WAGNER , C.; An Introduction to Credit Risk Modeling;
Chapman & Hall/CRC Financial Mathematics Series; 2nd Reprint; CRC Press (2003)

[14] B LUHM , C., OVERBECK , L.; Modeling of Collateralized Debt Obligations; book
manuscript; draft (2003)

[15] B OUY E , E., D URRLEMAN , V., N IKHEGBALI , A., R IBOULET, G., RONCALLI , T HIERRY;
Copulas for Finance. A Reading Guide and Some Applications; July (2000)
[16] C ROSBIE , P.; Modelling Default Risk; KMV Corporation (1999) (www.kmv.com)

[17] E MBRECHTS , P., M C N EIL , A., S TRAUMANN , D.; Correlation and Dependence in Risk
Management: Properties and Pitfalls, Preprint, July 1999.
[18] F REY, R., M C N EIL , A. J.; Modelling Dependent Defaults; Preprint, March (2001)

[19] F INGER , C. C.; Conditional Approaches for CreditMetrics Portfolio Distributions; Credit-
Metrics Monitor, April (1999)

36
[20] F INGER , C. C.; A Comparison of Stochastic Default Rate Models; RiskMetrics Journal,
Volume 1, November (2000)

[21] J OE , H.; Multivariate Models and Dependence Concepts; Chapman & Hall (1997)

[22] H ULL , J., W HITE , A.; Valuing Credit Default Swaps II: Modeling Default Correlations; The
Journal of Derivatives, 12-21, Spring (2001)

[23] I SRAEL , R., ROSENTHAL , J., W EI , J.; Finding Generators for Markov Chains via Empirical
Transition Matrices with Application to Credit Ratings; Mathematical Finance 11 (2), 245-265
(2001)

[24] JARROW, R. A., L ANDO , D., T URNBULL , S. M.; A Markov Model for the Term Structure
of Credit Risk Spreads; Review of Financial Studies 10, 481-523 (1997)

[25] K REININ , A., S IDELNIKOVA , M.; Regularization Algorithms for Transition Matrices; Algo
Research Quarterly 4 (1/2), 25-40 (2001)

[26] L I , D. X.; The Valuation of Basket Credit Derivatives; CreditMetrics Monitor, April (1999)

[27] L I , D. X.; On Default Correlation: A Copula Function Approach; The Journal of Fixed
Income 6, 43-54, March (2000)

[28] M ERTON , R.; On the Pricing of Corporate Debt: The Risk Structure of Interest Rates; The
Journal of Finance 29, 449-470 (1974)

[29] N ELSEN , R.; An Introduction to Copulas; Springer, New York (1999)

[30] N ICKELL , P., P ERRAUDIN , W., VAROTTO , S.; Ratings- versus Equity-based Credit Risk
Modeling: An Empirical Analysis; Working Paper (1999)

[31] N ORIS , J.; Markov Chains; Cambridge Series in Statistical and Probabilistic Mathematics;
Cambridge University Press (1998)

[32] OVERBECK , L., S CHMIDT, W.; Modeling Default Dependence with Threshold Models;
Preprint, submitted for publication (2003)

[33] S CHMIDT, W., WARD , I.; Pricing Default Baskets; RISK 15 (1), 111-114 (2002)

[34] S KLAR , A.; Fonction de Repartition a n Dimension et Leur Marges; Publications de lInsitute
Statistique de lUniversite de Paris 8, 229-231 (1959)

[35] S KLAR , A.; Random Variables, Joint Distribution Functions and Copulas; Kybernetika 9,
449-460 (1973)

[36] S TANDARD & P OOR S; Ratings Performance 2002: Default, Transition, Recovery, and
Spread; February (2003)

[37] VASICEK , O. A.; The Loan Loss Distribution; KMV Corporation

[38] Z HOU , C.; Default Correlation: An Analytical Result; Draft, May (1997)

37