Risk Analysis of Collateralized Debt Obligations

Kay Giesecke∗ and Baeho Kim†

February 13, 2009; this draft March 3, 2010‡

Operations Research, forthcoming

Abstract
Collateralized debt obligations, which are securities with payoffs that are tied to
the cash flows in a portfolio of defaultable assets such as corporate bonds, play a
significant role in the financial crisis that has spread throughout the world. Insuf-
ficient capital provisioning due to flawed and overly optimistic risk assessments is
at the center of the problem. This paper develops stochastic methods to measure
the risk of positions in collateralized debt obligations and related instruments tied
to an underlying portfolio of defaultable assets. It proposes an adaptive point pro-
cess model of portfolio default timing, a maximum likelihood method for estimating
point process models that is based on an acceptance/rejection re-sampling scheme,
and statistical tests for model validation. To illustrate these tools, they are used
to estimate the distribution of the profit or loss generated by positions in multi-
ple tranches of a collateralized debt obligation that references the CDX High Yield
portfolio, and the risk capital required to support these positions.

Keywords: Correlated default risk, collateralized debt obligation, portfolio credit

derivative, actual measure, point process, intensity, re-sampling, thinning, accep-
tance/rejection sampling, exact simulation.

∗
Department of Management Science & Engineering, Stanford University, Stanford, CA 94305-
4026, USA, Phone (650) 723 9265, Fax (650) 723 1614, email: [email protected], web:
www.stanford.edu/∼giesecke.
†
Korea University Business School, Anam-dong, Seongbuk-gu, Seoul 136-701 Korea, Phone +82 2
3290 2626, Fax +82 2 922 7220, email: [email protected].
‡
We are grateful for default data from Moody’s and market data from Morgan Stanley. We thank
Eymen Errais, Igor Halperin and Jack Kim for helpful discussions. We are also grateful to three anonymous
referees, an associate editor, Shahriar Azizpour, Lisa Goldberg, Steve Kou, and especially Jeremy Staum
for insightful comments.

1
1 Introduction
The financial crisis highlights the need for a holistic, objective and transparent approach
to accurately measuring the risk of investment positions in portfolio credit derivatives
such as collateralized debt obligations (CDOs). Portfolio credit derivatives are securities
whose payoffs are tied, often through complex schemes, to the cash flows in a portfolio
of credit instruments such as corporate bonds, loans, or mortgages. They facilitate the
trading of insurance against the default losses in the portfolio. An investor providing the
insurance is exposed to the default risk in the portfolio.
There is an extensive literature devoted to the valuation and hedging of portfolio
derivatives.1 The basic valuation problem is to estimate the price of default insurance,
i.e., the arbitrage-free value of the portfolio derivative at contract inception. This value
is given by the expected discounted derivative cash flows relative to a risk-neutral pricing
measure. After inception, the derivative position must be marked to market; that is, the
value of the derivative under current market conditions must be determined. The basic
hedging problem is to estimate the sensitivities of the derivative value to changes of
the default risk of the portfolio constituents. These sensitivities determine the amount
of constituent default insurance to be bought or sold to neutralize the derivative price
fluctuations due to small changes of the constituent risks.
The valuation and hedging problems are distinct from the risk analysis problem,
which is to measure the exposure of the derivative investor, who provides default in-
surance, to potential payments due to defaults in the portfolio. More precisely, the goal
is to estimate the distribution of the investor’s cumulative cash flows over the life of the
contract. The distribution is taken under the actual measure describing the empirical like-
lihood of events, rather than a risk-neutral pricing measure. The distribution describes the
risk/reward profile of a portfolio derivative position, and is the key to risk management
applications. For example, it allows the investor or regulator to determine the amount of
risk capital required to support a position. The financial crisis indicates the significance of
these applications and the problems associated with the traditional rating based analysis
of portfolio credit derivative positions; see SEC (2008).
The risk analysis problem has been largely ignored in the academic literature. This
paper provides stochastic methods to address this problem. It makes several contributions.
First, it develops a maximum likelihood approach to estimating point process models of
portfolio default timing from historical default experience. Second, it devises statistical
tests to validate a fitted model. Third, it formulates, fits and tests an adaptive point
process model of portfolio default timing, and demonstrates the utility of the estimation
and validation methods on this model. Fourth, it addresses important risk management
1
See Arnsdorf & Halperin (2008), Brigo, Pallavicini & Torresetti (2007), Chen & Glasserman (2008),
Cont & Minca (2008), Ding, Giesecke & Tomecek (2009), Duffie & Garleanu (2001), Eckner (2009), Errais,
Giesecke & Goldberg (2009), Kou & Peng (2009), Longstaff & Rajan (2008), Lopatin & Misirpashaev
(2008), Mortensen (2006), Papageorgiou & Sircar (2007) and many others.

2
applications, including the estimation of profit and loss distributions for positions in mul-
tiple tranches of a CDO. These distributions quantify and differentiate the risk exposure
of alternative investment positions, and the impact of complex contract features. They
are preferable to agency ratings, which are often based on the first moment only.
Estimating a stochastic point process model of portfolio default timing under the
actual probability measure presents unique challenges. Most importantly, inference must
be based on historical default timing data, rather than market derivative pricing data.
However, the default history of the reference portfolio underlying the credit derivative is
often unavailable, so direct inference is usually not feasible. We confront this difficulty
by developing an acceptance/rejection re-sampling scheme that allows us to generate
alternative portfolio default histories from the available economy-wide default timing data.
These histories are then used to construct maximum likelihood estimators for a portfolio
point process model that is specified in terms of an intensity process. A time-scaling
argument leads to testable hypotheses for the fit of a model.
The re-sampling approach is predicated on a top-down formulation of the portfolio
point process model. This formulation has become popular in the credit derivatives pricing
literature. Here, the point process intensity is specified without reference to the portfolio
constituents. In our risk analysis setting, the combination of top-down formulation and re-
sampling based inference leads to low-dimensional estimation, validation and prediction
problems that are highly tractable and fast to address even for the large portfolios that
are common in practice. Alternative bottom-up formulations in Das, Duffie, Kapadia
& Saita (2007), Delloye, Fermanian & Sbai (2006) and Duffie, Eckner, Horel & Saita
(2009) require the specification and estimation of default timing models for the individual
portfolio constituent securities. They allow one to incorporate firm-specific data into the
estimation, but lead to high-dimensional computational problems.
To demonstrate the effectiveness of the re-sampling approach and the appropriateness
of the top-down formulation, we develop and fit an adaptive intensity model. This model
extends the classical Hawkes (1971) model by including a state-dependent drift coefficient
in the intensity dynamics. The state-dependent drift involves a reversion level and speed
that are proportional to the intensity at the previous event. While this specification is
as tractable as the Hawkes model, it avoids the constraints imposed by the constant
Hawkes reversion level and speed. This helps to better fit the regime-dependent behavior
of empirical default rates. In- and out-of-sample tests show that our adaptive model
indeed captures the clustering in the default arrival data. Because it can capture the
default clustering better than other models that have been used in practice, our model
leads to more realistic estimates of the probabilities of losses to senior tranches.
This rest of this paper is organized as follows. Section 2 develops the re-sampling
approach to point process model estimation and validation. Section 3 formulates and
analyzes the adaptive point process model, and uses the re-sampling approach to fit it.
Section 4 applies the fitted model to the risk analysis of synthetic CDOs, while Section 5
analyzes the risk of cash CDOs. Section 6 concludes. There are several appendices.

3
2 Re-sampling based inference
This section develops a re-sampling approach to estimating a stochastic point process
model of default timing. It also provides a method to validate the estimators.

2.1 Preliminaries and problem

The uncertainty in the economy is modeled by a complete probability space (Ω, F, P ),
where P is the actual (statistical) probability measure. The information flow of investors
is described by a right-continuous and complete filtration F = (Ft )t≥0 .
Consider a non-explosive counting process N with event stopping times 0 < T1 <
T2 < . . . The Tn represent the ordered default times in a reference portfolio of firms,
with the convention that a defaulted name is replaced with name that has the same
characteristics as the defaulter. A portfolio credit derivative is a security with cash flows
that depend on the financial loss due to default in the reference portfolio.
Rt
Suppose N has a strictly positive intensity λ such that 0 λs ds < ∞ almost surely.
The intensity represents the conditional mean default rate in the sense that E(Nt+∆ −
R·
Nt | Ft ) ≈ λt ∆ for small ∆ > 0. This means that N − 0 λs ds is a local martingale relative
to P and F. The process followed by λ determines the distribution of N . It is the modeling
primitive, and is specified without reference to the constituent firms.
Our goal is to estimate a given model of λ. If the data consist of a path of N , then
this is a classical statistical problem; see Ogata (1978), for example. However, a path of
N is rarely available, so direct inference is typically not feasible. Instead, the data consist
of a path of the economy-wide default process N ∗ generated by the default stopping times
0 < T1∗ < T2∗ < . . . in the universe of names. We develop a re-sampling approach to
estimating λ from the realization of N ∗ . The basic idea is to generate alternative portfolio
default histories from N ∗ , and to estimate λ from these histories.

2.2 Acceptance/rejection re-sampling

We propose to generate paths of N from the realization of N ∗ by acceptance/rejection
sampling. Here, we randomly select an event time of N ∗ as an event time of N with
a certain conditional probability. The following basic observation, whose proof is found
in Appendix A, provides the foundation of this mechanism, and the justification of our
estimation approach. It also facilitates the design of tests to evaluate the approach.

Proposition 2.1. Let Z ∗ be a predictable process with values in [0, 1]. Select an economy-
wide event time Tn∗ with probability ZT∗n∗ . If the economy-wide default process N ∗ has
intensity λ∗ , then the counting process of the selected times has intensity Z ∗ λ∗ .

Proposition 2.1 states that the counting process obtained by thinning N ∗ according
to Z ∗ has intensity given by the product of the thinning process Z ∗ and the intensity of N ∗ .

4
Algorithm 1 (Acceptance/Rejection Re-Sampling) Generating a sample path of
the portfolio default process N from the realization of the economy-wide default process
N ∗ over the sample period [0, τ ].
1: Initialize m ← 0.
2: for n = 1 to Nτ∗ do
3: Draw u ∼ U (0, 1).
4: if u ≤ ZT∗n∗ then
5: Assign Tm+1 ← Tn∗ and update m ← m + 1.
6: end if
7: end for

The specification of Z ∗ is subject to a mild predictability condition: the selection prob-

ability at Tn∗ can only depend on information accumulated up to but not including time
Tn∗ . Proposition 2.1 is related to the construction of a marked point process from a Pois-
son random measure through a state-dependent thinning mechanism in Proposition 3.1
of Glasserman & Merener (2003). It is a generalization of the classical, state-independent
thinning scheme for the Monte Carlo simulation of time-inhomogeneous Poisson processes
proposed by Lewis & Shedler (1979).
We generate paths of the portfolio default process N by thinning N ∗ with a fixed
process Z ∗ , see Algorithm 1. Each of these paths represents an alternative event sequence,
or portfolio default history, assuming a defaulter is replaced with a name that has the
same characteristics as the defaulter. Proposition 2.1 implies that the events in each
alternative sequence arrive with intensity λ = Z ∗ λ∗ . This justifies the estimation of λ
from the alternative paths of the portfolio default process N .
It is important to note that this estimation approach does not require the specification
of a model for the economy-wide intensity λ∗ . We only need to formulate a model for the
thinning process Z ∗ , to which we turn next.

2.3 Thinning process specification

The random measure argument behind Proposition 3.1 in Giesecke, Goldberg & Ding
(2009) can be adapted to show that Z ∗ takes the form

E(Nt+ − Nt | Ft )
Zt∗ (ω) = lim ∗
(ω) (1)
→0 E(Nt+ − Nt∗ | Ft )

in all those points (ω, t) ∈ Ω × (0, ∞] where the limit exists.2 The quotient on the right
side of equation (1), which is taken to be zero when the denominator vanishes, represents
the conditional probability at time t that the next defaulter is a reference name, given
that a default occurs in the economy by time t + .
2
The limit in (1) exists and is equal to Zt∗ (ω) almost surely with respect to a certain measure on the
product space Ω × (0, ∞]. See Giesecke et al. (2009) for more details.

5
 We specify Z ∗ nonparametrically, guided by formula (1). Intuitively, Z ∗ must reflect
the relation between the issuer composition of the economy and the issuer composition
of the reference portfolio. We propose to describe the issuer composition in terms of the
credit ratings of the respective constituent issuers. In this case, N ∗ is identified with the
default process in the universe of rated issuers.
The use of ratings does not limit the applicability of our specification, because the
reference portfolios of most derivatives consist of rated names only. Moreover, the rating
agencies maintain extensive and accessible data bases that record credit events including
defaults in the universe of rated names, and further attributes associated with these
events, such as recovery rates. The agencies have maintained a high firm coverage ratio
throughout the sectors of the economy, and therefore the universe of rated names is a
reasonable representation of the firms in the economy.
Let [0, τ ] be the sample period, with τ denoting the (current) analysis time. Let R
be the set of rating categories, Xτ (ρ) be the number at time τ of reference firms with
rating ρ ∈ R, and Xt∗ (ρ) be the number at time t ∈ [0, τ ] of ρ-rated firms in the universe
of rated names. The number of firms X ∗ (ρ) is an adapted process. It varies through time
because issuers enter and exit the universe of rated names. Exits can be due to mergers
or privatizations, for example. We assume that the thinning process Z ∗ is equal to the
predictable projection of the process defined for times 0 < t ≤ τ by3
Xτ (ρ∗Nt−
∗ +1 )

∗
1{Xt−
∗ (ρ∗ )>0} (2)
Xt− (ρ∗Nt−
∗ +1 )
N∗
t− +1

where ρ∗n ∈ FTn∗ is the rating at the time of default of the nth defaulter in the economy.4
We require that Xτ (ρ) ≤ Xt∗ (ρ) for all t ≤ τ and ρ ∈ R. Since Xτ (ρ) is a fixed integer
∗
prescribed by the portfolio composition and Xt− (ρ) is predictable for fixed ρ,
X Xτ (ρ)
Zt∗ = ∗
P (ρ∗Nt−
∗ +1 = ρ | Ft− ) (3)
ρ∈R∗
X t− (ρ)
t−

almost surely, where Rt∗ is the set of rating categories ρ ∈ R for which Xt∗ (ρ) > 0. Formula
(3) suggests to interpret the value Zt∗ as the conditional “empirical” probability that the
next defaulter is a reference name. This conditional probability respects the ratings of
the reference names, as P (ρ∗Nt−∗ +1 = ρ | Ft− ) is the conditional probability that the next

defaulter has rating ρ. Our estimator νt∗ (ρ) of this latter conditional probability is based
on the ratings of the defaulters in [0, t). For ρ ∈ R, it is given by
∗
PNt−
1{ρ∗ =ρ} + α
νt∗ (ρ) = PN ∗ n=1 n 1{ρ∈Rt−
∗ } (4)
t− ∗
n=1 1{ρ ∗ ∈R∗ } + α|R
n t− t− |
3
Here and below, if Y is a right-continuous process with left limits, Yt− = lims↑t Ys .
4
Taking Z ∗ to be the predictable projection of the process given by formula (2) guarantees that Z ∗ is
defined up to indistinguishability; see Dellacherie & Meyer (1982).

6
where α ∈ (0, 1] is an additive smoothing parameter guaranteeing that νt∗ (ρ) is well-defined
for t < T1∗ . For α = 0, equation (4) defines the empirical rating distribution, which treats
the observations ρ∗1 , . . . , ρ∗Nt−
∗ as independent samples from a common distribution and
ignores all other information contained in Ft− . Our implementation assumes α = 0.5, a
value that can be justified on Bayesian grounds, see Box & Tiao (1992, pages 34-36).5

2.4 Likelihood estimators and fitness tests

Based on a collection of re-sampling paths {N (ωi ) : i ≤ I} of the portfolio default process
N generated by Algorithm 1, we estimate a model λ = λθ , where θ ∈ Θ is a parameter
vector and Θ is the set of admissible parameters. The paths of N induce intensity paths
{λθ (ωi ) : i ≤ I, θ ∈ Θ}. We fit θ by solving the log-likelihood problem6
Z τ I
X
sup (log λθs− (ωi )dNs (ωi ) − λθs (ωi )ds). (5)
θ∈Θ 0 i=1

The adequacy of the fitted intensity as a model of portfolio defaults depends on

the effectiveness of the re-sampling procedure and the appropriateness of the parametric
intensity specification. More precisely, it depends on how well the alternative re-sampling
scenarios generated by Z ∗ capture the actual default clustering in the reference portfolio,
and how well the fitted model for λ replicates these clusters. We require statistical tests
to assess this. The following result allows us to design such tests.

Proposition 2.2. Suppose that Zt∗ > 0, almost surely. Then the economy-wide default
process N ∗ is a standard Poisson process under a change of time defined by
Z t
∗ 1
At = λ ds.
∗ s
(6)
0 Zs

Proposition 2.2 expresses the compensator A∗ to N ∗ in terms of Z ∗ and λ, and states

that this compensator can be used to time-scale N ∗ into a standard Poisson process. We
evaluate the joint specification of Z ∗ and λ by testing whether the fitted (mean) paths
of these processes generate a realization of A∗ that time-scales the observed Tn∗ into a
standard Poisson sequence. The Poisson property can be tested with a battery of tests.

2.5 Portfolios without replacement

The intensity model λ estimated from the re-sampling scenarios generated by Algorithm
1 is based on a portfolio with replacement of defaulters. This is without loss of generality
because we can extend the reach of the fitted model to portfolios without replacement.
5
This choice makes (4) the so-called expected likelihood estimator. A sensitivity analysis indicates
that the model parameter estimates reported in Table 1 are robust with respect to variations of α.
6
The asymptotic properties of the maximum likelihood estimator of θ are developed by Ogata (1978).
The estimator is shown to be consistent, asymptotically normal, and efficient.

7
Algorithm 2 (Replacement Thinning) Generating a sample path of the portfolio
default process N 0 without replacement from a path of the portfolio default process N
with replacement over [τ, H], for a horizon H > τ .
1: Initialize m ← 0 as we define T00 = τ , and set Y (ρ) ← Xτ (ρ) for all ρ ∈ R.
2: for n = Nτ + 1 to NH do
3: Draw u ∼ U (0, 1).
4: if u ≤ ZTn then
5: Assign Tm0 ← Tn and update m ← m + 1.
6: Draw ρm ∼ ν 0 , where

Y (ρ)
ν 0 (ρ) = ν(ρ)/ZTn , ρ ∈ R. (9)
Xτ (ρ)

7: Update Y (ρm ) ← Y (ρm ) − 1.

8: end if
9: end for

Consider event times Tn of N over some interval [τ, H], where H > τ . The event
times Tn0 of the portfolio default process without replacement, N 0 , can be obtained from
the Tn by removing event times due to replaced defaulters. This is done by thinning. To
formalize this, let Xt (ρ) be the number of ρ-rated reference names at time t ≥ τ , assuming
defaulters are not replaced. Thus, for fixed ρ, Xt (ρ) ≤ Xτ (ρ) almost surely for every t ≥ τ .
For fixed ρ, the process X(ρ) decreases and vanishes when all ρ-rated reference names are
in default. It suffices to specify Z, the thinning process for N , at the event times Tn ≥ τ
of N . Motivated by formula (3), we suppose that
X XT − (ρ)
ZTn = n
P (ρn = ρ | FTn− ) (7)
ρ∈R
X τ (ρ)
τ

where Rτ is the set of rating categories ρ ∈ R for which Xτ (ρ) > 0, and ρn is the rating of
the firm defaulting at Tn . Note that the thinning probability (7) vanishes when all firms
in the portfolio are in default. We estimate the conditional distribution P (ρn = · | FTn− ) by
the smoothed empirical distribution ν of the ratings of the defaulters in the re-sampling
scenarios {N (ωi ) : i ≤ I}, where
PI PNτ (ωi )
i=1 n=1 1{ρn (ωi )=ρ} + α
ν(ρ) = PI PNτ (ωi ) 1{ρ∈Rτ } (8)
i=1 n=1 1{ρn (ωi )∈Rτ } + α|Rτ |

for an additive smoothing parameter α ∈ [0, 1]; see formula (4). This estimator treats the
observations ρn (ωi ) of all paths ωi as independent samples from a common distribution.
Algorithm 2 summarizes the steps required to generate N 0 from N .

8
3 An adaptive intensity model
This section demonstrates the effectiveness of the re-sampling approach. It formulates,
fits and evaluates an adaptive intensity model for the CDX High Yield portfolio, which is
a standard portfolio that is referenced by a range of credit derivatives. The fitted intensity
model will be used in Sections 4 and 5 to analyze the risk of such derivatives.

3.1 Re-sampling scenarios

We begin by examining the re-sampling scenarios of the default process of the CDX High
Yield Series 6 portfolio, or CDX.HY6. We adopt the rating system of Moody’s, a leading
rating agency.7 The reference portfolio consists of 2 Baa, 45 Ba, 39 B and 14 C rated firms.
The realization of N ∗ comes from Moody’s Default Risk Service, and covers all Moody’s
rated issuers between 1/1/1970 and 11/7/2008. We observe a total of 1447 defaults.
Figure 1 shows the annual economy-wide default rate between 1970 and 2008, along
with the mean default rate for the CDX.HY6 portfolio, obtained from Algorithm 1.8
Portfolio defaults cluster heavily, indicating the positive dependence between the defaults
of reference names. The excessive default rate in 1970 is due to an exceptional cluster
of 24 railway defaults on June 21, 1970. Other major event clusters are related to the
1987 crash and the burst of the internet bubble around 2001. These clusters translate into
substantial fluctuations of the empirical portfolio default rate that the model portfolio
intensity λ must replicate. This calls for special model features.

3.2 Intensity specification

Our specification of λ is informed by the results of the empirical analysis in Azizpour &
Giesecke (2008b), which is based on roughly the same historical default data used here.
7
We distinguish issuers in terms of their “senior rating,” which is an issuer-level rating generated by
Moody’s from ratings of particular debt obligations using its Senior Rating Algorithm, Hamilton (2005).
We follow a common convention and subsume the categories Caa and Ca into the category C. We also
subsume the numerical sub-categories Aa1, Aa2, Aa3 into the category Aa, and similarly for the other
numerical sub-categories. Then the set of rating categories is R = {Aaa, Aa, A, Baa, Ba, B, C, WR}.
The category WR indicates a withdrawn rating.
8
Each default event in the data base has a time stamp; the resolution is one day. There are days with
multiple events whose exact timing during the day cannot be established. Thus, the sequence of raw
event dates in the data set is not strictly increasing. Algorithm 1 requires distinct event times, however.
To address this issue, we assume that an event date is measured with uniformly distributed noise. The
noise is sampled when the paths of N are generated. We convert a raw economy-wide event date to a
real-valued calendar time equal to 12am on the day of the event, and draw the noise from a uniform
distribution on [0, 1/365]. With this choice, the randomization does not alter the original time stamp of
an observed event. We have experimented with several alternative randomization schemes but have found
that the model estimation results are insensitive to the chosen randomization scheme. Further, since the
data set allows us to measure the number of ρ-rated firms in the economy Xt∗ (ρ) only daily, in formula
(3) for Zt∗ we take Xt−
∗
(ρ) as the number of ρ-rated firms on the day prior to the day that contains t.

9
 4 12

3.5
10

8
2.5
Percent

Percent
2 6

1.5
4

2
0.5

0 0
1970 1975 1980 1985 1990 1995 2000 2005 1970 1975 1980 1985 1990 1995 2000 2005

Figure 1: Left panel : Annual default rate, relative to the number of rated names at the
beginning of a year, in the universe of Moody’s rated corporate issuers in any year between
1970 and 2008, as of 11/7/2008. Source: Moody’s Default Risk Service. Right panel : Mean
annual default rate for the CDX.HY6 portfolio for I = 10K re-sampling scenarios.

Using in- and out-of-sample tests, Azizpour & Giesecke (2008b) found that prediction of
economy-wide default activity based on past default timing outperforms prediction based
on exogenous economic covariates. Intuitively, the timing of past defaults provides infor-
mation about the timing of future defaults that is statistically superior to the information
contained in exogenous covariates. Further, if the past default history is the conditioning
information set, then the inclusion of additional economic covariates does not improve
economy-wide default forecast performance. These empirical findings motivate the formu-
lation of a parsimonious portfolio intensity model whose conditioning information set is
given by the past default history.
We assume that λt is a function of the path of N over [0, t]. More specifically, we
propose that λ evolves through time according to the equation

dλt = κt (ct − λt )dt + dJt (10)

where λ0 > 0 is the initial intensity value, κt = κλTNt is the decay rate, ct = cλTNt is the
reversion level, and J is a response jump process given by
X
Jt = max(γ, δλTn− )1{Tn ≤t} . (11)
n≥1

The quantities κ > 0, c ∈ (0, 1), δ > 0 and γ ≥ 0 are parameters. We denote by θ
the vector (κ, c, δ, γ, λ0 ). We give a sufficient condition guaranteeing that Nt < ∞ almost
surely, for all t. This condition relates the reversion level parameter c to the parameter δ,
which controls the magnitude of a jump of the intensity at an event.
Proposition 3.1. If c(1 + δ) < 1, then the counting process N is non-explosive.

10
 The intensity (10) follows a piece-wise deterministic process with right-continuous
sample paths. It jumps at an event time Tn . The jump magnitude is random. It is equal
to max(γ, δλTn− ), and depends on the intensity just before the event, which itself is a
function of the event times T1 , . . . , Tn−1 . The minimum jump size is γ. From Tn onwards
the intensity reverts exponentially to the level cλTn , at rate κλTn . Since the reversion rate
and level are proportional to the value of the intensity at the previous event, they depend
on the times T1 , . . . , Tn and change adaptively at each default. For Tn ≤ t < Tn+1 , the
behavior of the intensity is described by the FTn -measurable function

λt = cλTn + (1 − c)λTn exp(−κλTn (t − Tn )). (12)

The dependence of the reversion level ct , reversion speed κt and jump magnitude
max(γ, δλTn− ) on the path of the counting process N distinguishes our specification (10)
from the classical Hawkes (1971) model. The Hawkes intensity follows a piece-wise de-
terministic process dλt = κ(c − λt )dt + δdUt , where U = u1 + · · · + uN and the jump
magnitudes un are drawn from a fixed distribution on R+ . This model is more rigid than
our adaptive model (10): it imposes a global, state-independent reversion level c and speed
κ, and the magnitude of a jump in the Hawkes intensity is drawn independently of past
event times. Figure 2 contrasts the sample paths of (λ, N ) for the Hawkes model with
those for our model (10). The paths exhibit different clustering behavior, with the Hawkes
model generating a more regular clustering pattern. While Azizpour & Giesecke (2008a)
found that a variant of the Hawkes model performs well on the economy-wide default
data, we had difficulty fitting this and several other variants of the Hawkes model to the
portfolio default times generated by the re-sampling mechanism. We found the constant
reversion level and speed to be too restrictive. Our adaptive specification (10) relaxes
these constraints while preserving parsimony and computational tractability.
The jumps of the intensity process (10) are statistically important. They generate
event correlation: an event increases the likelihood of further events in the near future.
This feature facilitates the replication of the event clusters seen in Figure 1, a fact that
we establish more formally below. The intensity jumps can also be motivated in economic
terms. They represent the impact of a default on the other firms, which is channeled
through the complex web of contractual relationships in the economy. The existence of
these feedback phenomena is indicated by the ripple effects associated with the default
of Lehman Brothers on September 15, 2008, and is further empirically documented in
Azizpour & Giesecke (2008a), Jorion & Zhang (2007) and others. Our jump size specifi-
cation guarantees that the impact of an event increases with the default rate prevailing at
the event: the weaker the firms the stronger the impact. An alternative motivation of the
intensity jumps is Bayesian learning: an event reveals information about the values of un-
observed event covariates, and this leads to an update of the intensity, see Collin-Dufresne,
Goldstein & Helwege (2009), Duffie et al. (2009) and others.

11
 35 40 10 40

9
30 35 35
8
30 30
25 7

Number of Defaults

Number of Defaults
25 25
6
20
Intensity

Intensity
20 5 20
15
4
15 15

10 3
10 10
2
5 5 5
1

0 0 0 0
0 2 4 6 8 10 0 2 4 6 8 10

Figure 2: Left panel : Sample path of the intensity (left scale, solid) and default process
(right scale, dotted) for the adaptive model (10) with θ = (0.25, 0.05, 0.4, 0.8, 2.5), values
that are motivated by our estimation results in Section 3.4. The reversion level and speed
change at each event. Algorithm 3 is used to generate the paths. Right panel : Sample
path of the intensity and default process for the Hawkes model dλt = κ(c − λt )dt + δdUt
where the jump magnitudes un = 1 so U = N , λ0 = 2.5, and κ = 2.5 and c = δ = 1.5 are
chosen so that the expected number of events over 10 years matches that of the model
(10), roughly 37. Note that the Hawkes intensity cannot fall below the global reversion
level c. The algorithm in Ogata (1981) is used to generate the Hawkes model paths.

3.3 Event and loss simulation

The piece-wise deterministic dynamics of the intensity (10) generate computational ad-
vantages for model calculation. One benefit of this feature is that it allows us to estimate
the distribution of N and related quantities by exact Monte Carlo simulation of the jump
times of N . The inter-arrival times of N can be generated sequentially by acceptance-
rejection sampling from a dominating counting process with intensity λTNt− ≥ λt− .
The steps are summarized in Algorithm 3. In Step 1, we draw the state at the
analysis time τ from the re-sampling scenarios of N , and the corresponding fitted intensity
scenarios. The rate λS of the dominating Poisson process, from which a candidate default
time is generated in Step 4, is not only redefined at each acceptance, but also at each
rejection of a candidate time. This improves the efficiency of the algorithm. The acceptance
probability is increased and fewer candidate times are wasted.
Algorithm 3 leads to a trajectory of the default process N for a portfolio with replace-
ment of defaulters. A trajectory of the corresponding portfolio loss process L = `1 +· · ·+`N
is obtained by drawing the loss `n at each event time Tn of N . The distribution µ` of `n
is estimated by the empirical distribution of the losses associated with all defaults in the
re-sampling scenarios {N (ωi ) : i ≤ I}.

12
Algorithm 3 (Default Time Simulation) Generating a sample path of the portfolio
default process N over [τ, H] for a horizon H > τ and the model (10).
1: Draw (Nτ , TNτ , λTNτ ) uniformly from {(Nτ (ωi ), TNτ (ωi ), λTNτ (ωi )) : i ≤ I}.
2: Initialize (n, S) ← (Nτ , τ ) and λS ← cλTn + (1 − c)λTn exp(−κλTn (S − Tn )).
3: loop
4: Draw E ∼ Exp(λS ) and set T ← S + E.
5: if T > H then
6: Exit loop.
7: end if
8: λT ← cλTn + (λS − cλTn ) exp(−κλTn (T − S))
9: Draw u ∼ U (0, 1).
10: if u ≤ λT /λS then
11: λT ← λT + max(γ, δλT )
12: Assign Tn+1 ← T and update n ← n + 1.
13: end if
14: Set S ← T and λS ← λT .
15: end loop

Once a path of N is generated, Algorithm 2 can be used to obtain a path of the

default process N 0 for a portfolio without replacement of defaulters. A trajectory of the
corresponding portfolio loss process L0 = `01 + · · · + `0N 0 is obtained by drawing the loss `0n
at each event time from µ` .
Algorithm 3 is easy to implement and runs fast. For example, generating 100K paths
of N over 5 years for the fitted model parameters in Table 1 takes 1.67 seconds. Generating
100K paths of N 0 takes, in the same configuration, 2.14 seconds.9

3.4 Likelihood estimators

Another advantage of the piece-wise deterministic intensity dynamics (10) is that the
log-likelihood function in problem (5) takes a closed form that can be computed exactly.
Based on I = 10K re-sampling scenarios, we address the problem (5) with the Nelder-
Mead algorithm, which uses only function values. The algorithm is initialized at a set of
random parameter values, which are drawn from a uniform distribution on the parameter
space Θ = (0, 2)×(0, 1)×(0, 2)2 ×(0, 20). For each of 100 randomly chosen initial parameter
sets, the algorithm converges to the optimal parameter value θb reported in Table 1.10 We
9
The numerical experiments in this paper were performed on a desktop PC with an AMD Athlon 1.00
GHz processor and 960 MB of RAM, running Windows XP Professional. The codes were written in C++
and the compiler used was Microsoft Visual C++ .NET version 7.1.3088. For random number generation,
numerical optimization, and numerical root-finding, we used the GNU Scientific Library, Version 1.11.
10
In the computing environment described in Footnote 9, it takes 0.15 seconds to evaluate the log-
likelihood function for I = 10K and a given parameter set. The full optimization takes 146.88 seconds

13
 Parameter κ c δ γ λ0
MLE 0.254 0.004 0.419 0.810 8.709
95% CI (A) (0.252,0.255) (0.004,0.004) (0.417,0.422) (0.806,0.813) (8.566,8.851)
95% CI (B) (0.250,0.258) (0.003,0.004) (0.406,0.433) (0.791,0.829) (8.538,8.882)
Median 0.263 0.004 0.430 0.779 9.405

Table 1: Maximum likelihood estimates of the parameters of the intensity λ for the
CDX.HY6 portfolio, along with estimates of asymptotic (A) and bootstrapping (B) 95%
confidence intervals (10K bootstrap samples were used). The “Median” row indicates the
median of the empirical distribution of the per-path MLEs over all re-sampling paths. The
estimates are based on I = 10K re-sampling scenarios, generated by Algorithm 1 from the
observed defaults in the universe of Moody’s rated names from 1/1/1970 to 11/7/2008.

also provide asymptotic and bootstrapping confidence intervals. The left panel of Figure
3 shows the path of the fitted mean intensity λ b = I −1 PI λθb(ωi ).
i=1
To provide some perspective on the parameter estimates, we employ an alternative
inference procedure. Instead of maximizing the total likelihood (5) associated with all
paths of N , we maximize the path log-likelihood
Z τ
θ(ω )
sup (log λs− i (ωi )dNs (ωi ) − λθ(ω
s
i)
(ωi )ds)
θ(ωi )∈Θ 0

for each i = 1, 2, . . . , I. The last row in Table 1 shows the median of the empirical distri-
bution of the per-path MLE θ(ωi ) over all paths ωi . These values are in good agreement
with the MLEs, supporting our total likelihood estimation strategy.

3.5 Testing in-sample fit

Above we have stressed the computational advantages of our intensity model (10). In this
section and the next, we evaluate the model statistically. We show that the model fits the
default data, and that the fitted model leads to accurate event forecasts.
We evaluate the joint specification of Z ∗ and λ based on Proposition 2.2.11 The right
panel of Figure 3 shows the fitted path of the economy-wide intensity λ∗ = λ/Z ∗ , which
defines the time change. We test the Poisson property of the time-scaled inter-arrival times
R T∗
Wn∗ = T ∗n Z1∗ λs ds using a Kolmogorov-Smirnov (KS) test, a QQ plot, and Prahl’s (1999)
n−1 s
test. The KS test addresses the deviation of the empirical distribution of the (Wn∗ ) from
their theoretical standard exponential distribution. Prahl’s test is particularly sensitive to
when the initial parameters are set to the mid-points of the parameter space.
11
Proposition 2.2 requires the process Z ∗ to be strictly positive. The fitted values Zt∗ are indeed strictly
positive during our sample period. We approximate the paths of the fitted process Z ∗ and the fitted mean
intensity λ
b on a discrete-time grid with daily spacing. When multiple economy-wide defaults occur on
the same day, then the arrival times are spaced equally.

14
 20 110
Fitted Mean Portfolio Intensity Semi−annual Defaults
18 100 Fitted Economy−wide Intensity

16 90

80
14
70
12
60
10
50
8
40
6
30
4 20

2 10

0 0
1970 1975 1980 1985 1990 1995 2000 2005 1970 1975 1980 1985 1990 1995 2000 2005

b = I −1 PI λθb(ωi ) vs. [5%, 95%]

Figure 3: Left panel : Fitted mean portfolio intensity λ i=1
percentiles (boxes), [1%, 99%] percentiles (whiskers) and the mean number of portfolio
defaults in any given year between 1970 and 2008, based on I = 10K re-sampling scenarios
for the CDX.HY6. Right panel : Fitted economy-wide intensity λ/Z b ∗ vs. economy-wide
defaults between 1970 and 2008, semi-annually. The fitted intensity matches the time-
series fluctuation of observed economy-wide default rates.

large deviations of the (Wn∗ ) from their theoretical mean 1. Prahl shows that if the (Wn∗ )
are independent samples from a standard exponential distribution, then
1 X  W∗
M= 1− n (13)
m n:W ∗ <µ µ
n

is asymptotically normally distributed with mean µM = e−1 − 0.189/m and standard

√
deviation σM = 0.2427/ m, where m = Nτ∗ is the number of arrivals and µ is the sample
mean of (Wn∗ )n≤m . We reject the null hypothesis of a correct joint specification of Z ∗ and
λ if the test statistic ∆M = (M − µM )/σM lies outside of the interval (−1, 1).
Figure 4 contrasts the empirical distribution of the (Wn∗ ) with the theoretical standard
exponential distribution. The KS test has a p-value of 0.094, indicating that the deviation
of the empirical distribution from the theoretical distribution is not statistically significant
at standard confidence levels. The value of Prahl’s test statistic ∆M = 0.87 leads to the
same conclusion. The results of these and other12 tests suggest that the economy-wide
intensity λ∗ generated by (3) and (10) replicates the substantial time-series variation
of default rates during 1970–2008. This means that our adaptive intensity model (10)
captures the clustering of defaults observed during the sample period.
12
To detect a possible serial dependence of the Wn∗ , we also consider the test statistics SC1 and SC2
described by Lando & Nielsen (2009). These tests check the dependence of the number of re-scaled events
in non-overlapping time bins. Under the null of Poisson arrivals, the event counts are independent. We
cannot reject the null for bin sizes 2, 4, 6, 8 and 10 at standard confidence levels.

15
 9

1
8

Theoretical Exponential Quantiles 7

0.8
6

5 0.6

0.4
3

2
0.2
1

0 0
0 1 2 3 4 5 6 7 8 9 0 0.5 1 1.5 2 2.5 3 3.5 4
Empirical Quantiles

Figure 4: In-sample fitness tests: empirical distribution of time-scaled, economy-wide inter-

arrival times generated by the fitted (mean) paths of Z ∗ and λ, based on I = 10K
re-sampling scenarios for the CDX.HY6. Left panel : Empirical quantiles of time-scaled
inter-arrival times vs. theoretical standard exponential quantiles. Right panel : Empirical
distribution function (solid) of time-scaled inter-arrival times vs. theoretical standard
exponential distribution function (dotted) along with 1% and 5% bands.

A benefit of the time-scaling tests is that they can be applied to alternative model for-
mulations, facilitating a direct comparison of fitting performance. Consider the bottom-up
specification of Das et al. (2007), which appears to be much richer than ours: they specify
firm-level intensity models with conditioning information given by a set of firm-specific
and macro-economic covariates. Das et al. (2007) find that the time-scaled, economy-wide
times generated by their model deviate significantly from those of a Poisson process. In
particular, they find that their specification does not completely capture the event clusters
in the data. Thus, they reject this model formulation at standard confidence levels. This
may indicate that, relative to our portfolio-level formulation with conditioning informa-
tion given by past default timing, the additional modeling and estimation effort involved
in a firm-level intensity model formulation with a large set of exogenous covariates may
not translate into better fits and forecast performance.
Finally we note that the time-scaling tests can also be used to evaluate the specifica-
tion of λ directly on the re-sampling scenarios {N (ωi ) : i ≤ I}. For each path ωi , we time-
R·
scale the portfolio default process N (ωi ) with its fitted compensator A(ωi ) = 0 λs (ωi )ds,
and calculate the statistics of the KS test and Prahl’s test for the time-scaled inter-arrival
R Tn (ωi ) θb
times Wn = Tn−1 λ (ωi )ds, which can be calculated exactly thanks to formula (12).
(ωi ) s
The 95% confidence interval for the p-value of the KS test is 0.24 ± 0.01. The 95% confi-
dence interval for Prahl’s test statistic, defined for (Wn ) in analogy to (13), is 0.81 ± 0.07.
Also these tests indicate that the adaptive model (10) is well-specified.

16
 10 8
Realized Loss Realized Loss
9
7
8
6
7
5
6
1 Year Loss

1 Year Loss
5 4

4
3
3
2
2
1
1

0 0
1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007

Figure 5: Out-of-sample test of loss forecasts for a test portfolio with the same rating
composition as the CDX.HY6. We show the [25%, 75%] percentiles (box), the [15%, 85%]
percentiles (whiskers) and the median (horizontal line) of the fitted conditional distribu-
tion of the incremental portfolio loss L0τ +1 − L0τ given Fτ for τ varying annually between
1/1/1996 and 1/1/2007, estimated from 100K paths generated by Algorithms 3 and 2,
and the realized portfolio loss during [τ, τ + 1]. Left panel : The test portfolio is selected
at the beginning of each period. Right panel : The test portfolio is selected in 1996.

3.6 Testing out-of-sample loss forecasts

Next we assess the out-of-sample event forecast performance of the fitted model (10).
We contrast the loss forecast for a test portfolio without replacement with the actual
loss in the test portfolio. The firms in the test portfolio are randomly selected from the
universe of rated issuers, such that the test portfolio has the same rating composition
as the CDX.HY6.13 We apply Algorithms 3 and 2 to obtain an unbiased estimate of the
conditional distribution of the incremental test portfolio loss L0τ +1 − L0τ given Fτ . We then
contrast this distribution with the actual loss in the test portfolio during (τ, τ + 1], for
yearly analysis times τ between 1/1/1996 and 1/1/2007.
Figure 5 shows the results for two settings that represent common situations in prac-
tice. In one setting (left panel), we roll over the test portfolio in each 1 year test period.
That is, for each test period, we select a new test portfolio of 100 issuers at the beginning
of the period, estimate the portfolio loss distribution based on data available at the be-
ginning of the period, and compare it with the realized portfolio loss during the forecast
period. In the other setting (right panel), we select the test portfolio only once, in 1996.
For each period, we estimate the loss distribution for the portfolio of names that have
13
We do not use the CDX.HY6 itself, since it was formed only in 2006, and we want to test the loss
process over a longer time period, starting in 1996. The analysis would suffer from survivorship bias if
we were to take the CDX portfolio itself.

17
survived to the beginning of the period, and compare it with the realized portfolio loss
during the forecast period. This setting is motivated by the situation of a buy and hold
investor. The graphs indicate that the portfolio loss forecasts are very accurate.

4 Synthetic collateralized debt obligations

Having established the fitting and forecast performance of the adaptive model (10) of
portfolio default timing, we use it to estimate the exposure of an investor selling default
protection on a tranche of a synthetic CDO. A synthetic CDO is based on a portfolio of
C single-name credit swaps with notional 1 and maturity date H, the maturity date of
the CDO. The constituent credit swaps are referenced on bonds issued by rated firms.
When a bond issuer defaults, the corresponding credit swap pays the loss associated with
the event. A defaulter is not replaced.
A tranche of a synthetic CDO is a swap contract specified by a lower attachment point
K ∈ [0, 1), and an upper attachment point K ∈ (K, 1]. An index swap has attachment
points K = 0 and K = 1. The protection seller agrees to cover all losses due to default
in the reference portfolio, provided these losses exceed a fraction K of the total notional
C of the reference contracts but are not larger than a fraction K of C. In exchange, the
protection buyer pays to the protection seller an upfront fee at inception, and a quarterly
fee, both of which are negotiated at contract inception.
We estimate the risk exposure of the tranche protection seller, which is measured by
the conditional distribution at contract inception of the cumulative tranche loss, i.e., the
portfolio loss allocated to the tranche over the life of the contract. With the convention
that the portfolio loss at contract inception is equal to zero, the cumulative tranche loss
at a post-inception time t ≤ H is given by the “call spread”

Ut (K, K) = (L0t − CK)+ − (L0t − CK)+ . (14)

The left panel of Figure 6 shows the conditional distribution of the normalized 5 year cu-
mulative tranche loss UH (K, K)/C(K − K) for the CDX.HY6 on 3/27/2006, the Series 6
contract inception date,14 for each of several standard attachment point pairs. The matu-
rity date H for Series 6 contracts is 6/27/2011.15 To estimate the tranche loss distribution,
we generate default scenarios during 1/1/1970 – 3/27/2006 for the CDX portfolio from
Moody’s default history by the re-sampling Algorithm 1. Next we estimate the intensity
model (10) from these scenarios, as described in Section 3.4, with parameter estimates re-
ported in Table 2. Then we generate event times during the prediction interval 3/27/2006
14
Every 6 months, the CDX High Yield index portfolio is “rolled.” That is, a new portfolio with a new
serial number is formed by replacing names that have defaulted since the last roll, and possibly other
names. The index and tranche swaps we consider are tied to a fixed series.
15
Although the actual time to maturity is 5 years and 3 months at contract inception, here and below,
we follow the market convention of referring to the contract as a “5 year contract.” At an index roll, new
“5 year contracts” are issued, and these mature in 5 years and 3 months from the roll date.

18
 Index 0−10% 10−15% 15−25% Index 0−10% 10−15% 15−25%
1 1

0.9 0.9

0.8 0.8
Cumulative Loss Distribution

Cumulative Loss Distribution

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Index/Tranche Loss (%) Index/Tranche Loss (%)

Figure 6: Kernel smoothed conditional distribution of the normalized 5 year cumulative

tranche loss UH (K, K)/C(K − K) for the CDX.HY6 on 3/27/2006, the Series 6 contract
inception date, for each of several standard attachment point pairs (K, K). Here, H
represents the maturity date 6/27/2011. Left panel : Actual distribution, estimated using
the model and fitting methodology developed in Sections 2 and 3, based on I = 10K re-
sampling scenarios for N , and 100K replications. Right panel : Risk-neutral distribution,
estimated from the market prices paid for 5 year tranche protection on 3/27/2006 using
the method explained in Appendix B, based on 100K replications.

– 6/27/2011 from the fitted intensity using Algorithm 3, which we thin using Algorithm 2
to obtain paths of portfolio default times and losses without replacement. The trajectories
for U (K, K) thus obtained lead to an unbiased estimate of the desired distribution.
The loss distributions indicate the distinctions between the risk profiles of the different
tranches. The equity tranche, which has attachment points 0 and 10%, carries the highest
exposure among all tranches. For the equity protection seller, the probability of trivial
losses is roughly 15%, and the probability of losing the full tranche notional is 20%.
For the (10%, 15%)-mezzanine protection seller, the probabilities are roughly 80% and
10%, respectively. The mezzanine tranche is less risky, since the equity tranche absorbs
the first 10% of the total portfolio loss. For the (15%, 25%)-senior protection seller, the
probabilities are roughly 90% and 5%, respectively.
To provide some perspective on these numbers, we also estimate the risk-neutral
tranche loss distributions implied by the market prices paid for 5 year tranche protection
on 3/27/2006. The estimation procedure, explained in Appendix B, is based on the in-
tensity model (10) relative to a risk-neutral pricing measure. The risk-neutral intensity
parameters are chosen such that the model prices match the market index and tranche
prices as closely as possible. The fitting errors are small, which indicates that our adaptive
intensity model (10) performs also well under a risk-neutral pricing measure.
The right panel of Figure 6 shows the risk-neutral distribution of UH (K, K)/C(K−K)

19
 Parameter κ c δ γ λ0
MLE 0.260 0.003 0.439 0.856 8.494
95% CI (A) (0.259,0.261) (0.003,0.003) (0.437,0.441) (0.853,0.859) (8.368,8.621)
95% CI (B) (0.256,0.265) (0.003,0.004) (0.424,0.454) (0.836,0.876) (8.327,8.650)

Table 2: Maximum likelihood estimates of the parameters of the portfolio intensity λ

for the CDX.HY6, along with estimates of asymptotic (A) and bootstrapping (B) 95%
confidence intervals (10K bootstrap samples were used). The estimates are based on I =
10K re-sampling scenarios for N , generated by Algorithm 1 from the observed economy-
wide defaults in Moody’s universe of rated names from 1/1/1970 to 3/27/2006.

for the CDX.HY6, for each of several standard attachment point pairs. For the equity
protection seller, the risk-neutral probability of trivial losses is zero, and the risk-neutral
probability of losing the full tranche notional is 70%. Compare with the actual probabilities
indicated in the left panel of Figure 6. For any tranche, the risk-neutral probability of
losing more than any given fraction of the tranche notional is much higher than the
corresponding actual probability. The distinction between the probabilities reflects the risk
premium the protection seller requires for bearing exposure to the correlated corporate
default risk in the reference portfolio. Figure 7 shows the distributions of the normalized
cumulative portfolio loss L0H /C for multiple horizons H. For all horizons, the risk-neutral
distribution has a much fatter tail than the corresponding actual distribution. In other
words, when pricing index and tranche contracts, the market overestimates the probability
of extreme default scenarios relative to historical default experience.
Our tools can also be used to analyze more complex investment positions. Investors
often trade the CDO capital structure, i.e. they simultaneously sell and buy protection in
different tranches. A popular trade used to be the “equity-mezzanine” trade, in which the
investor sells equity protection and hedges the position by buying mezzanine protection
with matching notional and maturity. Figure 8 contrasts the conditional distribution on
3/27/2006 of the normalized cumulative loss (UH (0, 0.1) − UH (0.1, 0.15))/0.1C generated
by this trade with the conditional distribution of the normalized cumulative equity loss
UH (0, 0.1)/0.1C, both for the CDX.HY6. While the mezzanine hedge does not alter the
probability of trivial losses in the equity-only position, it does reduce the probability of a
total loss of notional from 20% to virtually zero. This is because the mezzanine protection
position generates cash flows when equity is wiped out. The magnitude of these cash flows
is however capped at 50% of the total equity notional, and this property generates the
point mass at 5% in the loss distribution of the equity-mezzanine position.
We can also measure the risk of positions in tranches referenced on different portfolios.
For example, an investor may sell and buy protection on several of the reference portfolios
in the CDX family, including the High Yield, Investment Grade and Crossover portfolios.
In this case, we estimate default and loss processes for each of the portfolios based on the

20
 0.1 0.1

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
10 10
9 9
8 8
7 7
6 40 6 40
5 30 Time to 5 30
Time to 20 20
Maturity (Years) 4 Maturity (Years) 4
10 10
3 0 3 0
Loss (%) Loss (%)

Figure 7: Kernel smoothed conditional distribution of the normalized cumulative portfolio

loss L0H /C for the CDX.HY6 on 3/27/2006 for horizons H varying between 3/27/2009
and 3/27/2016. Left panel : Actual distribution, estimated using the model and fitting
methodology developed in Sections 2 and 3, based on I = 10K re-sampling scenarios
for N , and 100K replications. Right panel : Risk-neutral distribution, estimated from the
market prices paid for 5 year tranche protection on 3/27/2006 using the method explained
in Appendix B, based on 100K replications.

realization of the economy-wide default process. We generate events for each portfolio,
and then aggregate the corresponding position losses as in the equity-mezzanine case.

5 Cash collateralized debt obligations

Next we use the adaptive model to estimate the exposure of an investor in a tranche
of a cash CDO. A cash CDO is based on a portfolio of corporate bonds, mortgages or
other credit obligations, which is bought by a special purpose vehicle (SPV) that finances
the purchase by issuing a collection of tranches that may pay coupons. The interest and
recovery cash flows generated by the reference bonds are collected by the SPV in a reserve
account, and are then allocated to the tranches according to a specified prioritization
scheme. Losses due to defaults are absorbed by the tranche investors, usually in reverse
priority order. Figure 9 illustrates a typical structure.
The reference portfolio of a cash CDO may be actively managed. In this case, the
asset manager can buy and sell collateral bonds, and replace defaulted obligations. We
analyze a basic unmanaged cash CDO, in which the reference bonds are held to maturity
and defaulted bonds are not replaced. We assume that the C reference bonds are straight
coupon bonds with notional 1, maturity date H equal to the CDO maturity date, issuance
date t0 equal to the CDO inception date, coupon dates (tm )1≤m≤M with tM = H, and

21
 1

0.9

0.8

Cumulative Loss Distribution

0.7

0.6

0.5

0.4

0.3

0.2

0.1 Equity−Mezzanine
Equity Only
0
0 2 4 6 8 10
Loss (%)

Figure 8: Kernel smoothed conditional distribution on 3/27/2006 of the normalized cu-

mulative loss (UH (0, 0.1) − UH (0.1, 0.15))/0.1C associated with selling equity protection
and buying mezzanine protection with matching notional and maturity, along with the
distribution of the normalized cumulative equity loss UH (0, 0.1)/0.1C, for the CDX.HY6.

per-period coupon rate v. The total interest income to the SPV in period m is therefore

W (m) = v(C − Nt0m ). (15)

When a reference bond defaults, the SPV collects the recovery at the coupon date following
the event. The total recovery cash flow in period m is

K(m) = Nt0m − Nt0m−1 − (L0tm − L0tm−1 ).

The total cash flows from coupons and recoveries are invested in a reserve account that
earns interest at the risk-free rate r. For period m, they are given by

W (m) + K(m) if 1 ≤ m < M
B(m) = 0
W (m) + K(m) + C − Ntm if m = M.

The SPV issues three tranches, of which two are debt tranches that promise to pay a
specified coupon at times (tm ). There is one senior debt tranche, represented by a sinking-
fund bond with initial principal p1 and per-period coupon rate c1 , and a junior debt tranche
that is represented by a sinking-fund bond with initial principal p2 and per-period coupon
rate c2 . The initial principal of the residual equity tranche is p3 = C −p1 −p2 . Each tranche
has maturity date H.
For the debt tranches j = 1, 2 and coupon period m, we denote by Fj (m) the re-
maining principal. The scheduled interest payment is Fj (m)cj , and the accrued unpaid

22
 Figure 9: Cash flow “waterfall” of a sample cash CDO.

interest is Aj (m − 1). If Qj (m), the actual interest paid in period m, is less than Fj (m)cj ,
then the difference is accrued at cj to generate an accrued unpaid interest of

Aj (m) = (1 + cj )Aj (m − 1) + cj Fj (m) − Qj (m).

The actual payments to the debt tranches are prioritized. We consider two prioritization
schemes, the uniform and fast schemes, which were introduced by Duffie & Garleanu
(2001) and which are reviewed in Appendix C for completeness. A prioritization scheme
specifies the actual interest payment Qj (m), the pre-payment of principal Pj (m), and the
contractual unpaid reduction in principal Jj (m). The total cash payment in period m is
Qj (m) + Pj (m). The remaining principal after interest payments is

Fj (m) = Fj (m − 1) − Pj (m) − Jj (m) ≥ 0.

The par coupon rate on tranche j is the scheduled coupon rate cj with the property that
the initial market value of the bond is equal to its initial face value Fj (0) = pj .
The residual equity tranche does not make scheduled coupon or principal payments
so c3 = 0. Instead, at maturity H, after the debt tranches have been paid as scheduled,
any remaining funds in the reserve account are allocated to the equity tranche.
We analyze the exposure of a tranche investor for a 5 year cash CDO referenced on
the CDX.HY6 of C = 100 names. This choice of reference portfolio allows us to compare
the results with those for the synthetic CDO. Before we can start this analysis, we need
to price the reference bonds and the debt tranches.16 The first step is to estimate the
par coupon rate of the reference bonds. This is the scheduled coupon rate v ∗ with the
16
In practice, this is done by the SPV, which then offers the tranche terms to potential investors. We
have to perform this task here since we do not have access to actual data for the structure we analyze.

23
 Reference Uniform Scheme Fast Scheme
Bond Senior Junior Senior Junior
Annualized Par Coupon Rate (bp) 811.23 476.58 1299.42 483.38 749.75
Par Coupon Spread (bp) 339.20 23.68 846.52 11.29 277.77

Table 3: Fitted annualized par coupon rates and spreads on 3/27/2006 for the constituent
bonds and debt tranches of a 5 year cash CDO referenced on the CDX.HY6, for each
of two standard prioritization schemes. The principal values are (p1 , p2 , p3 ) = (85, 10, 5).
The fitting procedure is described in Appendix D.

property that the initial market value of a reference bond is equal to its initial face value
1. Given v ∗ , the second step is to estimate the par coupon rates c∗j of the debt tranches.
Appendix D gives details on these steps, and Table 3 reports the annualized par coupon
rates and spreads. The par coupon spread is the difference between the par coupon rate
and the (hypothetical) par coupon rate that would be obtained if the reference bonds
were not subject to default risk.
Next we estimate the conditional distribution of the discounted cumulative tranche
loss, which is the loss the tranche investor faces during the life of the contract, discounted
at the risk-free rate r. At a post-inception time t ≤ H, the discounted cumulative tranche
loss is given by

Ujt (H, pj , v ∗ , c∗1 , c∗2 ) = V jt (H, pj , v ∗ , c∗1 , c∗2 ) − Vjt (H, pj , v ∗ , c∗1 , c∗2 ) (16)

where Vjt (H, pj , v ∗ , c∗1 , c∗2 ) is the present value at time t of the coupon and principal cash
flows actually paid to tranche j over the life of the tranche, and V jt (H, pj , v ∗ , c∗1 , c∗2 ) is the
present value of these cash flows assuming no defaults occur during the remaining life. For
a debt tranche, V jt (H, pj , v ∗ , c∗1 , c∗2 ) represents the present value of all scheduled coupon
and principal payments. For the equity tranche, this quantity represents the present value
of the reserve account value after the debt tranches have been paid as scheduled.
Note that due to the complex structure of the cash flow prioritization, the tranche loss
Ujt (H, pj , v ∗ , c∗1 , c∗2 ) does not admit a simple analytic expression in terms of the portfolio
default count and loss of the reference portfolio. This leaves simulation as the only feasible
tool for analyzing the cash CDO tranche loss, regardless of whether or not the models of
the portfolio default count and loss processes are analytically tractable.
Figure 10 shows the fitted conditional distribution on 3/27/2006 of the normalized
loss Ujτ (H, pj , v ∗ , c∗1 , c∗2 )/V jτ (H, pj , v ∗ , c∗1 , c∗2 ) for a 5 year structure whose maturity date
H is 6/27/2006, for each of several tranches. To estimate that distribution, we proceed as
described in Section 4 for the synthetic CDO, and generate paths of the fitted portfolio
default and loss processes. These are then fed into the cash flow calculator, which computes
Vjt (H, pj , v ∗ , c∗1 , c∗2 ) for each path. We assume that coupons are paid quarterly. The risk-
free interest rate r is deterministic, and is estimated from Treasury yields for multiple
maturities on 3/27/2006, obtained from the website of the Department of Treasury.

24
 1 1

0.9 0.9

0.8 0.8
Cumulative Loss Distribution

Cumulative Loss Distribution

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
Senior Senior
0.1 Junior 0.1 Junior
Residual Residual
0 0
0 20 40 60 80 100 0 20 40 60 80 100
Tranche Loss (%) Tranche Loss (%)

Figure 10: Kernel smoothed conditional distribution on 3/27/2006 of the normalized dis-
counted loss Ujτ (H, pj , v ∗ , c∗1 , c∗2 )/V jτ (H, pj , v ∗ , c∗1 , c∗2 ) for a 5 year cash CDO referenced on
the CDX.HY6, whose maturity date H is 6/27/2011. The initial tranche principals are
p1 = 85 for the senior tranche, p2 = 10 for the junior tranche, and p3 = 5 for the resid-
ual equity tranche. We apply the model and fitting methodology developed in Sections 2
and 3, based on I = 10K re-sampling scenarios for N , and 100K replications. Left panel :
Uniform prioritization scheme. Right panel : Fast prioritization scheme.

As in the synthetic CDO case, the risk profile of a cash CDO tranche depends on the
degree of over-collateralization, i.e. the location of the attachment points. It also depends
on the prioritization scheme specified in the CDO terms. With the uniform scheme, the
cash flows generated by the reference bonds are allocated sequentially to the debt tranches
according to priority order, and default losses are applied to all tranches in reverse priority
order. With the fast scheme, the cash flows are used to retire the senior tranche as soon as
possible. After the senior tranche is retired, the cash flows are used to service the junior
tranche. After the junior tranche is retired, any residual cash flows are distributed to
the equity tranche. Therefore, the senior tranche investor is less exposed under the fast
scheme: the probability of zero losses is increased from roughly 95% for the uniform scheme
to roughly 98%. This reduction in risk is reflected by the par coupon spreads reported in
Table 3. For the junior investor, the probability of zero losses increases from roughly 82%
for the uniform scheme to roughly 93% for the fast scheme, but the probability of a loss
equal to the present value of all scheduled coupon and principal payments increases from
zero to about 3%. Nevertheless, the junior tranche commands a much smaller spread
under the fast scheme. The risk reduction for the debt tranche investors comes at the
expense of the equity investor: while the probability of zero losses is about 15% for both
schemes, the probability of a loss equal to the present value of the reserve account value
after the debt tranches have been paid as scheduled increases from zero to roughly 8% for
the fast scheme. On the other hand, the equity investor has a much higher upside under

25
 1
Uniform Scheme
0.9 Fast Scheme

0.8

0.7

Cumulative Distribution
0.6

0.5

0.4

0.3

0.2

0.1

0
−100 −50 0 50 100 150 200 250 300
Tranche P&L (%)

Figure 11: Kernel smoothed conditional distribution on 3/27/2006 of the normalized dis-
counted profit and loss (V3τ (H, 5, v ∗ , c∗1 , c∗2 ) − 5)/5 for the residual equity tranche of a 5
year cash CDO referenced on the CDX.HY6, whose maturity date H is 6/27/2011. We
apply the model and fitting methodology developed in Sections 2 and 3, based on I = 10K
re-sampling scenarios for N , and 100K replications.

the fast scheme. Figure 11 shows the distribution of the normalized discounted profit and
loss (V3τ (H, 5, v ∗ , c∗1 , c∗2 ) − 5)/5 for an equity tranche position. For the equity investor, the
probability of more than doubling the principal p3 after discounting is roughly 75% under
the fast scheme, while it is only 60% under the uniform scheme. Figure 12 shows that
for the debt tranches, the upside is much more limited, for any scheme. For example, the
normalized discounted profit on the senior tranche can be at most 0.5% under the fast
scheme and about 1% under the uniform scheme.
A standard measure to quantify the risk of a position is value at risk, a quantile of
the position’s loss distribution. We estimate the value at risk at time t ≤ H of a position
in a cash CDO tranche maturing at H, given by

Ujt (H, pj , v ∗ , c∗1 , c∗2 )

    
∗ ∗ ∗
VaRjt (α, H, pj , v , c1 , c2 ) = inf x ∈ R : P ≤ x  Ft > α

V jt (H, pj , v ∗ , c∗1 , c∗2 )

for some level of confidence α ∈ (0, 1). Figure 13 shows the 99.5% value at risk for the 5
year senior tranche with maturity date H given by 6/27/2011, for analysis times varying
weekly between 3/27/2006 and 11/7/2008, for each of the two prioritization schemes.
Assuming risk capital is allocated according to the value at risk, a value in the time series
represents the amount of risk capital that is needed at that time to support the tranche
position over its remaining life. For both prioritization schemes, the time series behavior
reflects the rising trend in corporate defaults that started in 2008 and that is evidenced
in Figure 1. The fast scheme requires less risk capital than the uniform scheme but leads

26
 1 1
Uniform Scheme Uniform Scheme
0.9 Fast Scheme 0.9 Fast Scheme

0.8 0.8

0.7 0.7
Cumulative Distribution

Cumulative Distribution
0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
−12 −10 −8 −6 −4 −2 0 2 −100 −50 0 50
Tranche P&L (%) Tranche P&L (%)

Figure 12: Kernel smoothed conditional distribution on 3/27/2006 of the normalized dis-
counted profit and loss (Vjτ (H, pj , v ∗ , c∗1 , c∗2 ) − pj )/pj for the debt tranches of a 5 year
cash CDO referenced on the CDX.HY6, whose maturity date H is 6/27/2011. We apply
the model and fitting methodology developed in Sections 2 and 3, based on I = 10K re-
sampling scenarios for N , and 100K replications. Left panel : Senior tranche with principal
p1 = 85. Right panel : Junior tranche with principal p2 = 10.

to higher volatility of risk capital.

6 Conclusion
This paper develops, implements and validates stochastic methods to measure the risk of
investment positions in collateralized debt obligations and related credit derivatives tied
to an underlying portfolio of defaultable assets. The ongoing financial crisis highlights the
need for sophisticated yet practical tools allowing potential and existing investors as well
as regulators to quantify the exposure associated with such positions, and to accurately
estimate the amount of risk capital required to support a position.
The key to address the risk analysis problem is a model of default timing that cap-
tures the default clustering in the underlying portfolio, and a method to estimate the
model parameters based on historical default experience. This paper contributes to each
of these two sub-problems. It formulates an adaptive, intensity-based point process model
of default timing that performs well according to in- and out-of-sample tests. Moreover,
it develops a maximum likelihood approach to estimating point process models. This ap-
proach is based on an acceptance/rejection re-sampling scheme that generates alternative
portfolio default histories from the available economy-wide default process.
The point process model and inference method have potential applications in other
areas dealing with correlated event arrivals, within and beyond financial engineering.

27
 Uniform (Senior) Fast (Senior)
23 16

22 15

21 14
20
99.5% Value−at−Risk (%)

99.5% Value−at−Risk (%)

13
19
12
18
11
17
10
16
9
15

14 8

13 7

12 6
Mar−06 Aug−06 Jan−07 Jun−07 Nov−07 Apr−08 Sep−08 Mar−06 Aug−06 Jan−07 Jun−07 Nov−07 Apr−08 Sep−08

Figure 13: 99.5% value at risk VaR1τ (0.995, H, 85, v ∗ , c∗1 , c∗2 ) for the 5 year senior tranche
of a cash CDO referenced on the CDX.HY6 with maturity date 6/27/2011, for τ varying
weekly between 3/27/2006 and 11/7/2008, based on I = 10K re-sampling scenarios for
N , and 100K replications. Left panel : Uniform prioritization scheme. Right panel : Fast
prioritization scheme.

Financial engineering examples include the pricing and hedging of securities exposed to
correlated default risk, and order book modeling. Other example areas are insurance,
healthcare, queuing, and reliability.

A Proofs
Proof of Proposition 2.1. Let (Un∗ ) be a sequence of standard uniform random variables
that are independent of one another and independent of the (Tn∗ ). Let π(du, dt) be the
random counting measure with mark space [0, 1] associated with the marked point process
RtR1
(Tn∗ , Un∗ ). Note that Nt∗ = 0 0 π(du, ds), and that π(du, dt) has intensity λπ (u, t) = λ∗t−
for u ∈ [0, 1], which we choose to be predictable. This means that the process defined by
RtR1
0 0
(π(du, ds) − λπ (u, s)duds) is a local martingale. Now for u ∈ [0, 1] define the process
I(u, ·) by I(u, t) = 1{u≤Zt∗ } . Then the counting process N generated by the selected event
times can be written as
X Z tZ 1
∗ ∗
Nt = I(Un , Tn )1{Tn∗ ≤t} = I(u, s)π(du, ds).
n≥1 0 0

Since Z ∗ is predictable, I(u, ·) is predictable for u ∈ [0, 1]. The process I(u, ·) is also
bounded for u ∈ [0, 1]. Therefore, a local martingale M is defined by
Z tZ 1
Mt = Nt − I(u, s)λπ (u, s)du ds.
0 0

28
But the definitions of I(u, s) and λπ (u, s) yield
Z t Z Zs∗ Z t
∗
Mt = Nt − λs− du ds = Nt − Zs∗ λ∗s ds,
0 0 0

which means that N has intensity Z ∗ λ∗ .

R·
Proof of Proposition 2.2. Proposition 2.1 implies that the compensator A = 0 λs ds to
N is absolutely continuous with respect to the compensator A∗ to N ∗ , with density Z ∗ .
Because Zt∗ > 0 almost surely, A∗ is also absolutely continuous with respect to A, with
density 1/Z ∗ . It follows that A∗ can be written as
Z t Z t
∗ 1 1
At = ∗
dAs = λ ds.
∗ s
0 Zs 0 Zs

The time change theorem of Meyer (1971) implies the result, since A∗ is continuous and
increases to ∞, almost surely.
Rt
Proof of Proposition 3.1. It suffices to show that 0 λs ds < ∞ almost surely for each
t > 0. We assume that γ = 0, without loss of generality. For s ≥ 0 and h(0) > 0, let

h(s) = ch(0) + (1 − c)h(0) exp(−κh(0)s).

The function h describes the behavior of the intensity (10) between events. For Tn ≤ t <
Tn+1 and h(0) = λTn , we have that λt = h(t − Tn ). The inverse h−1 to h is given by
 
−1 1 h(0)(1 − c)
h (u) = log
κh(0) u − ch(0)

for ch(0) < u ≤ h(0). Consider the first hitting time

ϕ(h(0)) = inf{s ≥ 0 : h(s) < h(0)/(1 + δ)}.

For c(1 + δ) < 1, we have

 
−1 1 (1 + δ)(1 − c)
ϕ(h(0)) = h (h(0)/(1 + δ)) = log .
κh(0) 1 − c(1 + δ)
Let T0 = 0. The conditional probability given FTn of the intensity jumping at an
event to a value that exceeds the value taken at the previous event is

P (λTn+1 > λTn | FTn ) = P (Tn+1 − Tn < ϕ(λTn ) | FTn )

 Z ϕ(λT ) 
n
= 1 − exp − h(s)ds
0
   
c (1 + δ)(1 − c) δ
= 1 − exp − log − =: M,
κ 1 − c(1 + δ) κ(1 + δ)

29
where M is strictly less than 1 and independent of n = 0, 1, 2, . . . It follows that the
unconditional probability P (λTn+1 > λTn ) = M , for any n. Further, an induction argument
can be used to show that for any n and k = 1, 2, . . .
P (λTn+k > λTn+k−1 > · · · > λTn ) = M k .
Now letting
Ctk = {ω : Nt (ω) ≥ k and λTn+1 (ω) > λTn (ω) for n = Nt (ω) − 1, . . . , Nt (ω) − k}
for t > 0 and k = 1, 2, . . ., we have that
P (Ctk ) = P (Nt ≥ k and λTNt > λTNt −1 > · · · > λTNt −k )

≤ P (λTNt +k > λTNt +k−1 > · · · > λTNt )

= Mk
independently of t. We then conclude that
Z t   
P λs (ω)ds < ∞ ≥ P sup λs (ω) < ∞
0 s∈[0,t]

∞
 \ c 
≥P Ctk (17)
k=1
∞
\ 
=1−P Ctk
k=1

= 1 − lim P (Ctk ) (18)

k→∞

≥ 1 − lim M k
k→∞

= 1.
Equation (18) is due to the fact that Ctk ∈ Ft for all k = 1, 2, . . . and Ct1 ⊇ Ct2 ⊇ · · · . To
justify the inequality (17), we argue as follows. For given t > 0 and ω ∈ Ω, define

At (ω) = k ∈ {0, 1, . . . , Nt (ω) − 1} : λTk+1 (ω) ≥ λTk (ω) .
Let ω̂ ∈ { ∞ k c k
T
k=1 Ct } . Then, by definition of Ct , either (i) Nt (ω̂) < ∞, or (ii) there exists
some n < ∞ such that λTn+1 (ω̂) < λTn (ω̂) and |At (ω̂) \ ATn (ω̂)| < ∞.
In case (i), thanks to the condition |At (ω̂)| ≤ Nt (ω̂) < ∞, each λTk+1 (ω̂)−λTk (ω̂) < ∞
for all k ∈ At (ω̂). Thus, λs (ω̂) < ∞ for all s ∈ [0, t]. Hence, we conclude that ω̂ ∈ {ω :
sups∈[0,t] λs (ω) < ∞} in this case.
Now consider case (ii). Note that for any s ∈ [0, t], the condition cλTNs (ω̂) < λs (ω̂)
implies that λs (ω̂) remains at infinity once λTNs (ω̂) achieves infinity. Hence, λTn+1 (ω̂) <
λTn (ω̂) implies that λTn (ω̂) < ∞, and we conclude that sups∈[0,Tn ] λs (ω̂) < ∞. Moreover,
due to the condition |At (ω̂) \ ATn (ω̂)| < ∞, we conclude that sups∈[Tn ,t] λs (ω̂) < ∞.
Therefore, { ∞ k c
T
k=1 Ct } ⊆ {ω : sups∈[0,t] λs (ω) < ∞} and (17) is justified.

30
B Risk-neutral tranche loss distributions
To describe the estimation of the risk-neutral tranche loss distributions discussed in Sec-
tion 4, it is necessary to explain the arbitrage-free valuation of index and tranche swaps.
These contracts are based on a portfolio of C credit swaps with common notional 1,
common maturity date H and common quarterly premium payment dates (tm ).
In an index swap, the protection seller covers portfolio losses as they occur (default
leg), and the protection buyer pays Sam (C − Nt0m ) at each premium date tm , where S is
the swap rate and am is the day count fraction for coupon period m (premium leg). The
value at time t ≤ H of the default leg is given by Dt (H, 0, 1), where
Z H  Z s   
Dt (H, K, K) = EQ exp − ru du dUs (K, K)  Ft . (19)

t t

Here, EQ (· | Ft ) denotes the conditional expectation operator relative to a risk-neutral

pricing measure Q. The value at time t ≤ H of the premium leg is given by
X   Z tm   
0
Pt (S) = S am EQ exp − ru du (C − Ntm )  Ft . (20)

tm ≥t t

The index rate at t is the solution S = St (H) to the equation Dt (H, 0, 1) = Pt (S).
In a tranche swap with upfront rate G and running rate S, the protection seller covers
tranche losses (14) as they occur (default leg) and, for K < 1, the protection buyer pays
GKC at inception and Sam (KC − Utm ) at each premium date tm , where K = K − K is
the tranche width (premium leg). The value at τ ≤ H of the default leg is given by (19).
The value at time t ≤ H of the premium leg is given by
X   Z tm   
Pt (K, K, G, S) = GKC + S am EQ exp − ru du (KC − Utm )  Ft . (21)

tm ≥t t

For a fixed upfront rate G, the running rate S is the solution S = St (H, K, K, G) to the
equation Dt (H, K, K) = Pt (K, K, G, S). For a fixed rate S, the upfront rate G is the
solution G = Gt (H, K, K, S) to the equation Dt (H, K, K) = Pt (H, K, K, G, S).
The valuation relations are used to estimate the risk-neutral portfolio and tranche
loss distributions from market rates of index and tranche swaps. First we formulate a
parametric model for the risk-neutral dynamics of N and L, the portfolio default and loss
processes with replacement. Our risk-neutral model parallels the model under P . Suppose
that N has risk-neutral intensity λQ with Q-dynamics

dλQ Q Q Q
t = κt (ct − λt )dt + dJt (22)

where λQ Q Q Q Q Q Q
0 > 0, κt = κ λTNt is the decay rate, ct = c λTNt is the reversion level, and J is
a response jump process given by
X
Jt = max(γ Q , δ Q λQ )1
T − {Tn ≤t}
. (23)
n
n≥1

31
The quantities κQ > 0, cQ ∈ (0, 1), δ Q > 0 and γ Q ≥ 0 are parameters such that
cQ (1 + δ Q ) < 1. We let θQ = (κQ , cQ , δ Q , γ Q , λQ
0 ).
Q
Since the Q-dynamics of λ mirror the P -dynamics of the P -intensity λ, we can apply
the algorithms developed above to estimate the expectations (19)–(21) and to calculate
the model index and tranche rates for the model (22). We first generate event times Tn of
N by Algorithm 3, with λ and its parameters replaced by their risk-neutral counterparts,
and with initial condition (Nτ , TNτ , λQ Q
TNτ ) = (0, 0, λτ ). Then we apply the replacement
Algorithm 2 as stated to generate event times Tn0 without replacement, and the corre-
sponding paths of N 0 and L0 required to estimate the expectations (19)–(21). In this last
step, we implicitly assume that the thinning probability (7), the rating distributions (8)
and (9), and the distribution µ` of the loss at an event are not adjusted when the measure
is changed from P to Q. The risk-free interest rate r is assumed to be deterministic, and
is estimated from Treasury yields for multiple maturities on 3/27/2006, obtained from
the website of the Department of Treasury.
The risk-neutral parameter vector θQ is estimated from a set of market index and
tranche rates by solving the nonlinear optimization problem
X  MarketMid(i) − Model(i, θQ ) 2
min (24)
θQ ∈ΘQ
i
MarketAsk(i) − MarketBid(i)

where ΘQ = (0, 2)×(0, 1)×(0, 2)2 ×(0, 20) and the sum ranges over the data points. Here,
MarketMid is the arithmetic average of the observed MarketAsk and MarketBid quotes.
We address the problem (24) by adapted simulated annealing. The algorithm is initialized
at a set of random parameter values, which are drawn from a uniform distribution on the
parameter space ΘQ . For each of 100 randomly chosen initial parameter sets, the algorithm
converges to the optimal parameter values given in Table 5. The market data and fitting
results for 5 year index and tranche swaps referenced on the CDX.HY6 on 3/27/2006 are
reported in Table 4. The model fits the data, with an average absolute percentage error
of 2.9%.

C Cash CDO prioritization schemes

This appendix describes the prioritization schemes for the cash CDOs analyzed in this
paper. These schemes were introduced by Duffie & Garleanu (2001).

C.1 Uniform prioritization

The total interest income W (m) from the reference bonds is sequentially distributed as
follows. In period m, the debt tranches receive the coupon payments
Q1 (m) = min{(1 + c1 )A1 (m − 1) + c1 F1 (m), W (m)}
Q2 (m) = min{(1 + c2 )A2 (m − 1) + c2 F2 (m), W (m) − Q1 (m)}.

32
 Contract MarketBid MarketAsk MarketMid Model
Index 332.88 333.13 333.01 333.08
0-10% 84.50% 85.00% 84.75% 86.75%
10-15% 52.75% 53.75% 53.25% 48.19%
15-25% 393.00 403.00 398.00 398.88
25-35% 70.00 85.00 77.50 79.43
MinObj 33.77
AAPE 2.90%

Table 4: Market data from Morgan Stanley and fitting results for 5 year index and tranche
swaps referenced on the CDX.HY6 on 3/27/2006. The index, (15 − 25%) and (25 − 35%)
contracts are quoted in terms of a running rate S stated in basis points (10−4 ). For these
contracts the upfront rate G is zero. The (0, 10%) and (10, 15%) tranches are quoted in
terms of an upfront rate G. For these contracts the running rate S is zero. The values in
the column Model are fitted rates based on model (22) and 100K replications. We report
the minimum value of the objective function MinObj and the average absolute percentage
error AAPE relative to market mid quotes.

Unpaid reductions in principal from default losses, Jj (m), occur in reverse priority order,
so that the residual equity tranche suffers the reduction

J3 (m) = min{F3 (m − 1), ξ(m)},

where
ξ(m) = max{0, L0tm − L0tm−1 − [W (m) − Q1 (m) − Q2 (m)]}
is the cumulative loss since the previous coupon date minus the collected and undistributed
interest income. Then, the debt tranches are reduced in principal by

J2 (m) = min{F2 (m − 1), ξ(m) − J3 (m)}

J1 (m) = min{F1 (m − 1), ξ(m) − J3 (m) − J2 (m)}.

With uniform prioritization there are no early payments of principal, so P1 (m) = P2 (m) =
0 for m < M . At maturity, principal and accrued interest are treated identically, while
the remaining reserve is paid in priority order. The payments of principal at maturity are

P1 (M ) = min{F1 (M ) + A1 (M ), R(M ) − Q1 (M )},

P2 (M ) = min{F2 (M ) + A2 (M ), R(M ) − Q1 (M ) − P1 (M ) − Q2 (M )}

where R(M ) is the value of the reserve account at M . The residual equity tranche receives

D3 (M ) = R(M ) − Q1 (M ) − P1 (M ) − Q2 (M ) − P2 (M ).

33
 Parameter κQ cQ δQ γQ λQτ
Estimate 0.522 0.333 0.366 1.646 8.586

Table 5: Estimates of the parameters of the risk-neutral portfolio intensity λQ , obtained

from market rates of 5 year index and tranche swaps referenced on the CDX.HY6 on
3/27/2006.

C.2 Fast prioritization

The senior tranche collects interest and principal payments as quickly as possible until
maturity or until its remaining principal becomes zero, whichever is first. In period m,
the senior tranche receives interest and principal payments

Q1 (m) = min{B(m), (1 + c1 )A1 (m − 1) + c1 F1 (m)}

P1 (m) = min{F1 (m − 1), B(m) − Q1 (m)}

As long as the senior tranche receives payments the junior tranche accrues coupons. After
the senior tranche has been retired, the junior tranche is allocated interest and principal
until maturity or until its principal is written down, whichever is first:

Q2 (k) = min{(1 + c2 )A2 (m − 1) + c2 F1 (m), B(m) − Q1 (m) − P1 (m)}

P2 (k) = min{F2 (m − 1), B(m) − Q1 (m) − P1 (m) − Q2 (m)}.

The equity tranche receives any residual cash flows. There are no contractual reductions
in principal.

D Cash CDO valuation

This appendix explains the valuation of the cash CDO reference bonds and debt tranches.
At time t ≤ H, the value of the reference portfolio is
X   Z tm   
Ot (v) = EQ exp − ru du B(m, v)  Ft (25)

tm ≥t t

where the notation B(m) = B(m, v) indicates the dependence of the cash flow B(m) on
the coupon rate v of a reference bond. The par coupon rate of a reference bond is the
number v = v ∗ such that Ot0 (v) = C at the CDO inception date t0 .
The par coupon rates for the debt tranches are determined similarly. Let Ojt (v ∗ , c1 , c2 )
be the value at time t ≤ H of the bond representing tranche j = 1, 2 when the reference
bonds accrue interest at their par coupon rate v ∗ , and when the coupon rates on the debt
tranches are c1 and c2 , respectively. Because of the complexity of the cash flow “waterfall,”

34
this quantity does not admit a simple analytic expression. The par coupon rates are given
by the pair (c1 , c2 ) = (c∗1 , c∗2 ) such that Ojt0 (v ∗ , c1 , c2 ) = pj for j = 1, 2.
For the CDX.HY6 of reference bonds, assuming that the cash CDO is established on
3/27/06 and has a 5 year maturity, the par coupon rates are estimated as follows. We first
generate a collection of 100K paths of N 0 and L0 under the risk neutral measure, based on
the risk-neutral intensity model (22) with calibrated parameters in Table 5, as described
in Appendix B. Based on these paths, we can estimate Ot0 (v) for fixed v, and then solve
numerically for the par coupon rate v ∗ . The risk-free interest rate r is deterministic, and
is estimated from Treasury yields for multiple maturities on 3/27/2006. Given v ∗ , c1 , c2 ,
we can calculate the tranche cash flows for a given path of (N 0 , L0 ) according to the spec-
ified prioritization scheme, and then estimate Ojt0 (v ∗ , c1 , c2 ). We then solve numerically
a system of two equations for (c∗1 , c∗2 ).

References
Arnsdorf, Matthias & Igor Halperin (2008), ‘BSLP: markovian bivariate spread-loss model
for portfolio credit derivatives’, Journal of Computational Finance 12, 77–100.

Azizpour, Shahriar & Kay Giesecke (2008a), Premia for correlated default risk. Working
Paper, Stanford University.

Azizpour, Shahriar & Kay Giesecke (2008b), Self-exciting corporate defaults: Contagion
vs. frailty. Working Paper, Stanford University.

Box, George & George Tiao (1992), Bayesian Inference in Statistical Analysis, Wiley, New
York.

Brigo, Damiano, Andrea Pallavicini & Roberto Torresetti (2007), ‘Calibration of CDO
tranches with the dynamical generalized-Poisson loss model’, Risk 20(5), 70–75.

Chen, Zhiyong & Paul Glasserman (2008), ‘Fast pricing of basket default swaps’, Opera-
tions Research 56(2), 286–303.

Collin-Dufresne, Pierre, Robert Goldstein & Jean Helwege (2009), How large can jump-
to-default risk premia be? Modeling contagion via the updating of beliefs. Working
Paper, Carnegie Mellon University.

Cont, Rama & Andreea Minca (2008), Recovering portfolio default intensities implied by
CDO quotes. Working Paper, Columbia University.

Das, Sanjiv, Darrell Duffie, Nikunj Kapadia & Leandro Saita (2007), ‘Common failings:
How corporate defaults are correlated’, Journal of Finance 62, 93–117.

35
Dellacherie, Claude & Paul-André Meyer (1982), Probabilities and Potential, North Hol-
land, Amsterdam.

Delloye, Martin, Jean-David Fermanian & Mohammed Sbai (2006), ‘Dynamic frailties and
credit portfolio modeling’, Risk 19(1), 101–109.

Ding, Xiaowei, Kay Giesecke & Pascal Tomecek (2009), ‘Time-changed birth processes
and multi-name credit derivatives’, Operations Research 57(4), 990–1005.

Duffie, Darrell, Andreas Eckner, Guillaume Horel & Leandro Saita (2009), ‘Frailty corre-
lated default’, Journal of Finance 64, 2089–2123.

Duffie, Darrell & Nicolae Garleanu (2001), ‘Risk and valuation of collateralized debt
obligations’, Financial Analysts Journal 57(1), 41–59.

Eckner, Andreas (2009), ‘Computational techniques for basic affine models of portfolio
credit risk’, Journal of Computational Finance 15, 63–97.

Errais, Eymen, Kay Giesecke & Lisa Goldberg (2009), Affine point processes and portfolio
credit risk. Working Paper, Stanford University.

Giesecke, Kay, Lisa Goldberg & Xiaowei Ding (2009), A top-down approach to multi-name
credit. Operations Research, forthcoming.

Glasserman, Paul & Nicolas Merener (2003), ‘Numerical solution of jump-diffusion LIBOR
market models’, Finance and Stochastics 7, 1–27.

Hamilton, David (2005), Moodys senior ratings algorithm and estimated senior ratings.
Moodys Investors Service.

Hawkes, Alan G. (1971), ‘Spectra of some self-exciting and mutually exciting point pro-
cesses’, Biometrika 58(1), 83–90.

Jorion, Philippe & Gaiyan Zhang (2007), ‘Good and bad credit contagion: Evidence from
credit default swaps’, Journal of Financial Economics 84(3), 860–883.

Kou, Steven & Xianhua Peng (2009), Default clustering and valuation of collateralized
debt obligations. Working Paper, Columbia University.

Lando, David & Mads Stenbo Nielsen (2009), Correlation in corporate defaults: Contagion
or conditional independence? Working Paper, Copenhagen Business School.

Lewis, P. & G. Shedler (1979), ‘Simulation of nonhomogeneous poisson processes by thin-

ning’, Naval Logistics Quarterly 26, 403–413.

Longstaff, Francis & Arvind Rajan (2008), ‘An empirical analysis of collateralized debt
obligations’, Journal of Finance 63(2), 529–563.

36
Lopatin, Andrei & Timur Misirpashaev (2008), ‘Two-dimensional Markovian model for
dynamics of aggregate credit loss’, Advances in Econometrics 22, 243–274.

Meyer, Paul-André (1971), Démonstration simplifée d’un théorème de Knight, in

‘Séminaire de Probabilités V, Lecture Notes in Mathematics 191’, Springer-Verlag
Berlin, pp. 191–195.

Mortensen, Allan (2006), ‘Semi-analytical valuation of basket credit derivatives in

intensity-based models’, Journal of Derivatives 13, 8–26.

Ogata, Yosihiko (1978), ‘The asymptotic behavior of maximum likelihood estimators

of stationary point processes’, Annals of the Institute of Statistical Mathematics
30(A), 243–261.

Ogata, Yosihiko (1981), ‘On Lewis’ simulation method for point processes’, IEEE Trans-
actions on Information Theory 27, 23–31.

Papageorgiou, Evan & Ronnie Sircar (2007), ‘Multiscale intensity models and name group-
ing for valuation of multi-name credit derivatives’, Applied Mathematical Finance
15(1), 73–105.

Prahl, Jürgen (1999), A fast unbinned test of event clustering in poisson processes. Work-
ing Paper, Universität Hamburg.

SEC (2008), Summary report of issues identified in the commission staff’s examinations
of select credit rating agencies. Report, US Securities and Exchange Commission.

37