Credit Risk Models:

An Overview

Paul Embrechts, Rüdiger Frey, Alexander McNeil

ETH Zürich

c 2003 (Embrechts, Frey, McNeil)
 A. Multivariate Models for Portfolio Credit Risk

1. Modelling Dependent Defaults: Introduction

2. Latent Variable Models for Default

3. Bernoulli Mixture Models for Default

4. Mapping Between Latent Variable and Mixture Models

5. Statistical Issues in Default Modelling

6. Implications for Pricing Basket Credit Derivatives

c 2003 (Embrechts, Frey, McNeil) 1
 A1. Motivation

• Focus in credit risk research has mainly been on modelling of

default of individual firm.

• Modelling of joint defaults in standard models (KMV,

CreditMetrics) is relatively simplistic (based on multivariate
normality).

• In large balanced loan portfolios main risk is occurrence of many

joint defaults – this might be termed extreme credit risk.

• For determining tail of loss distribution, the specification of

dependence between defaults is at least as important as the
specification of individual default probabilities.

c 2003 (Embrechts, Frey, McNeil) 2
 Modelling of Default – Overview

Consider portfolio of m firms and a time horizon T (typically

1 year). For i ∈ {1, . . . , m} define Yi to be default indicator of
company i, i.e. Yi = 1 if company defaults by time T , Yi = 0
otherwise. (Reduction to two states for simplicity.)

Model Types

• Latent variable models. Default occurs, if a latent variable Xi (often

interpreted as asset value at horizon T ) lies below some threshold Di
(liabilities). Examples: Merton model (1974), CreditMetrics, KMV.
• Mixture Models. Bernoulli default probabilities are made stochastic.
Yi | Qi ∼ Be (Qi) where Qi is a random variable taking values in
[0, 1] and Q1, . . . , Qm are dependent. Example: CreditRisk+.

c 2003 (Embrechts, Frey, McNeil) 3
 Simplifications

• We consider only a two-state model (default/no-default). All of the

ideas generalise to more-state models with different credit-quality
classes. Probabilistic ideas are more easily understood in two-state
setting.

• We neglect the modelling of exposures. The basic messages do

not change when different exposures and losses-given-defaults are
introduced.

c 2003 (Embrechts, Frey, McNeil) 4
 A2. Latent Variable Models
0
Given random vector X = (X1, . . . , Xm) with continuous marginal
distributions Fi and thresholds D1, . . . , Dm, define Yi := 1{Xi≤Di}.
Default probability of counterparty i given by

pi := P (Yi = 1) = P (Xi ≤ Di) = Fi (Di) .

Notation: (Xi, Di)1≤i≤m denotes a latent variable model.

Examples

• Classical Merton-model.
Xi is interpreted as asset value of company i at T . Di is value of
liabilities. Assume X ∼ N (µ, Σ).

c 2003 (Embrechts, Frey, McNeil) 5
 Industry Examples of Latent Variable Models

• KMV-model. As Merton but Di is now chosen so that default

probability pi equals average default probability of companies with
same “distance-to-default” as company i.
• CreditMetrics. We assume X ∼ N (0, Σ). Threshold Di is chosen
so that pi equals average default probability of companies with
same rating class as company i.
• Model of Li. (CreditMetrics Monitor 1999) Xi interpreted as
survival time of company i. Assume Xi exponentially distributed
with parameter λi chosen so that P (Xi ≤ T ) = pi, with pi
determined as in CreditMetrics. Multivariate distribution of X
specified using Gaussian copula.

c 2003 (Embrechts, Frey, McNeil) 6
 Model Calibration

In both KMV and CreditMetrics, µi, Σii and Di are chosen so that
pi equals average historical default frequency for companies with
a similar credit quality.
To determine further structure of Σ (i.e. correlations) both models
assume a classical linear factor model for p < m.
p
X
Xi = µi + ai,j Θj + σiεi
j=1

for Θ ∼ Np(0, Ω), independent standard normally distributed rv’s

ε1, . . . , εm, which are also independent of Θ.
Θ global, country and industry effects impacting all companies.
ai,j weights for company i, factor j;  idiosyncratic effects.

c 2003 (Embrechts, Frey, McNeil) 7
 Equivalent Latent Variable Models and Copulas

Definition: Two latent variable models (Xi, Di)1≤i≤m and

(Xei , D
e i)1≤i≤m generating multivariate Bernoulli vectors Y and Y
e
d e
are said to be equivalent if Y = Y.

Proposition: (Xi, Di)1≤i≤m and (X ei , D

e i)1≤i≤m are equivalent if:
 
1. P (Xi ≤ Di) = P X ei ≤ D
e i , ∀i.
2. X and Xe have the same copula.

CreditMetrics and KMV are equivalent, as are all latent variable

models that use the Gaussian dependence structure for latent
variables, such as the model of Li, regardless of how marginals are
modelled.

c 2003 (Embrechts, Frey, McNeil) 8
 Special Case: Homogeneous Groups

It is common to group obligors together to form homogeneous

groups. This corresponds to the mathematical concept of
exchangeability.
A random vector X is exchangeable if

d

(X1, . . . , Xm) = Xp(1), . . . , Xp(m) ,

for any permutation (p(1), . . . , p(m)) of (1, . . . , m).

c 2003 (Embrechts, Frey, McNeil) 9
 Exchangeable Default Model

We talk of an exchangeable default model if the default indicator

vector Y is exchangeable.
If a latent variable vector X is exchangeable (or has an exchangeable
copula) and all individual default probabilities P (Xi ≤ Di) are
equal, then Y is exchangeable.

Exchangeability allows a simplified notation for default probabilities:

πk := P Yi1 = 1, . . . , Yik = 1 ,
{i1, . . . , ik } ⊂ {1, . . . , m}, 1 ≤ k ≤ m,
π := π1 = P (Yi = 1) , i ∈ {1, . . . , m} .

c 2003 (Embrechts, Frey, McNeil) 10
 The Copula is Critical

To see this consider special case of exchangeable default model.

Consider any subgroup of k companies {i1, . . . , ik } ⊂ {1, . . . , m}.

πk = P Yi1 = 1, . . . , Yik = 1 = P Xi1 ≤ Di1 , . . . , Xik ≤ Dik
= C1,...,k (π, . . . , π) ,

where C1,...,k is the k–dimensional margin of C.

The copula C crucially determines higher order joint default

probabilities and thus extreme risk that many companies default.
For π small, copulas with lower tail dependence will lead to higher
πk ’s and more joint defaults.

c 2003 (Embrechts, Frey, McNeil) 11
Comparison of Exchangeable Gaussian and t Copulas
If X is given an asset value interpretation large (downward)
movements of the Xi might be expected to occur together; therefore
tail dependence may be realistic.
Two cases: (extensions such as generalized hyperbolic distributions
can be considered analogously).

1. X ∼ Nm(0, R)
2. X ∼ tm,ν (0, R) .

R is an equicorrelation matrix with off–diagonal element ρ > 0, so

that X is exchangeable with correlation matrix R in both cases. We
also fix thresholds so that Y is exchangeable in both cases and
P (Yi = 1) = π, ∀i, in both models. We vary the value for ν.

c 2003 (Embrechts, Frey, McNeil) 12
 Simulation Study

We consider m = 10000 companies. All losses given default are one

unit; total loss is number of defaulting companies. Set π = 0.005
and ρ = 0.038, these being values corresponding to a homogeneous
group of “medium” credit quality in the KMV/CreditMetrics
Gaussian approach. We set ν = 10 in t–model and perform
100000 simulations to determine loss distribution.
The risk is compared by comparing high quantiles of the loss
distributions (the so–called Value–at–Risk approach to measuring
risk).
Results Min 25% Med Mean 75% 90% 95% Max
Gauss 1 28 43 49.8 64 90 109 331
t 0 1 9 49.9 42 132 235 3238

c 2003 (Embrechts, Frey, McNeil) 13
 Ratio of quantiles of loss distributions (t:Gaussian)
Ratio of Quantiles of Loss Distributions
••
•••
•
••
••
••
•
•
•••
6

••
•
•
••
•
Student t : Gauss

•
••
•
••
•••
•
•••
•
4

•••••
•
••
••
••••••
•
•••
••
••
••
••••••••
••••••
••
••
•
••••••••••••
•
•••••••••
••
•
••
••
•
•••
•
•••••••••••••••
2

•
••••••••••••••••••
••
••
•
••
••
•
••
••
•
•••
••
••••••••••••••••••••••
•
••••••••••••••••••••••
•••
•
••
••
•
••
••
•
••
••
•
••
••
•
••
••
•
••
••
•
••
••••••••••••••••••••••••••••••••••••••••
•••••••••••••••••••••••••••••••••••••••
••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••

0.80 0.85 0.90 0.95 1.00

Quantile

m = 10000, π = 0.005, ρ = 0.038 and ν = 10.

c 2003 (Embrechts, Frey, McNeil) 14
 A3. Exchangeable Bernoulli Mixture Models
The default indicator vector (Y1, . . . , Ym) follows an exchangeable
Bernoulli mixture model if there exists a rv Q taking values in (0, 1)
such that, given Q, Y1, . . . , Ym are iid Be(Q) rvs.

π = P (Yi = 1) = E (Yi) = E (E (Yi | Q)) = E(Q)

Z 1
k
q k dG(q) ,



πk = P Yi1 = 1, . . . , Yik = 1 = E Q =
0

where G(q) is the mixture distribution function of Q. Unconditional

default probabilities and higher order joint default probabilities are
moments of the mixing distribution.
It follows that, for i 6= j, cov (Yi, Yj ) = π2 − π 2 = var Q ≥ 0.
π2−π 2
Default correlation is given by ρY := corr (Yi, Yj ) = π−π2 .

c 2003 (Embrechts, Frey, McNeil) 15
 Examples of Mixing Distributions

• Beta Q ∼ Beta(a, b), g(q) = β(a, b)−1q a−1(1 − q)b−1, a, b > 0

−1 2


• Probit–Normal Φ (Q) ∼ N µ, σ (CreditMetrics/KMV)
 
Q 2


• Logit–Normal log 1−Q ∼ N µ, σ (CreditPortfolioView)

Parameterizing Mixing Distributions

These examples all have two parameters. If we fix the default
probability π and default correlation ρY (or equivalently the first two
moments of the mixing distribution π and π2) then we fix these two
parameters and fully specify the model.
Example: Exchangeable Beta–Bernoulli Mixture Model
π = a/(a + b), π2 = π(a + 1)/(a + b + 1).

c 2003 (Embrechts, Frey, McNeil) 16
 Beta Mixing Distribution

150
100
g(q)
50
0

0.0 0.02 0.04 0.06 0.08 0.10

Beta Density g(q) of mixing variable Q in exchangeable Bernoulli

mixture model with π = 0.005 and ρY = 0.0018.

c 2003 (Embrechts, Frey, McNeil) 17
 Extreme Risk in Large Balanced Portfolios

In exchangeable models for large homogeneous groups with similar

exposures the tail of the loss distribution is proportional to the tail of
the mixing distribution (Frey & McNeil 2001).

For portfolio size m large

VaRα(Loss) ≈ m e VaRα(Q).

where e is mean exposure.

This result underlies loss distribution approximation in KMV and

scaling rule in Basel II.

c 2003 (Embrechts, Frey, McNeil) 18
 Tail of mixing distribution with first two moments
fixed

10^0
10^-16 10^-14 10^-12 10^-10 10^-8 10^-6 10^-4 10^-2
P(Q>q)

Probit-normal
Beta
Logit-normal

0.0 0.2 0.4 0.6 0.8

q

Tail of the mixing distribution G in three exchangeable Bernoulli

mixture models: probit–normal; logit–normal; beta.

c 2003 (Embrechts, Frey, McNeil) 19
 More General Bernoulli Mixture Models
Definition: (Mixture Model with Factor Structure)
(Y1, . . . , Ym) follow a Bernoulli mixture model with p–factor
structure if there is a random vector Ψ = (Ψ1, . . . , Ψp) with p < m
and continuous functions fi : Rp → (0, 1), such that
1. Yi | Ψ ∼ Be(Qi), i = 1, . . . , m, where

Qi = fi (Ψ1, . . . , Ψp) for all 1 ≤ i ≤ m .

2. (Y1, . . . , Ym) are conditionally independent given Ψ.

Remark: Poisson mixture models with factor structure can be
defined analogously, by making the Poisson rate parameters
dependent on Ψ.
Example: CreditRisk+ has this kind of structure.

c 2003 (Embrechts, Frey, McNeil) 20
 A4. Mapping Latent Variable to Mixture Models

It is often possible to transform a latent variable model to obtain an

equivalent Bernoulli mixture model with factor structure. This is
useful in Monte Carlo simulation, since Bernoulli mixture models are
generally easier to simulate than latent variable models.
Example: KMV/Creditmetrics
X is Gaussian and follows a classical linear p–factor model.
p
X
Xi = ai,j Θj + σiεi = a0iΘ + σiεi
j=1

for a p–dimensional random vector Θ ∼ Np(0, Ω), independent

standard normally distributed rv’s ε1, . . . , εm, which are also
independent of Θ.

c 2003 (Embrechts, Frey, McNeil) 21
 CreditMetrics/KMV as a Bernoulli Mixture Model

For the mixing factors take Ψ = Θ.

P (Yi = 1 | Ψ) = P (Xi ≤ Di | Ψ) = P (εi ≤ (Di − a0iΨ) /σi | Ψ)

= Φ ((Di − a0iΨ) /σi) .

Clearly Yi | Ψ ∼ Be (Qi) where Qi = Φ ((Di − a0iΨ) /σi).

Thus Qi has a probit–normal distribution.

Moreover, conditional on Ψ, the Yi are independent.

c 2003 (Embrechts, Frey, McNeil) 22
 Mapping Other L.V. Models to Mixture Models

A similar mapping is possible when the latent variables follow

a multivariate normal mixture model, as in the case of t or
generalised hyperbolic latent variables.

X has a normal mixture distribution if Xi = gi(W ) + W Zi where

W ≥ 0 is independent of Z, gi : (0, ∞) → R, and Z is Gaussian
vector with E(Z) = 0.

If Gaussian vector Z follows a linear factor model as before then it is

possible to derive explicitly an equivalent Bernoulli mixture model.

Examples: p
1. Student t model: W = ν/V , V ∼ χ2ν and gi(W ) = µi.
2. Generalized hyperbolic: W ∼ NIG and gi(W ) = µi + βiW .

c 2003 (Embrechts, Frey, McNeil) 23
 Normal and t: Equivalent Mixture Approach

The profound differences between the Gaussian and t copulas with

similar asset correlation can be understood in terms of the
differences between the corresponding mixture distributions.
Consider two cases (again in exchangeable special case).

Case 1: Asset Correlation held fixed.

Here we observe clear differences between the densities of the
equivalent mixing distributions as we vary degrees of freedom. These
account for differences in distribution of number of defaults.

Case 2: Default Correlation held fixed.

Here differences between densities are much less obvious.
Distributions of the number of defaults very similar 95th and
99th percentiles; differences visible only very far in the tail.

c 2003 (Embrechts, Frey, McNeil) 24
 Mixing densities – similar asset correlation
Densities of mixing distribution
30

t3
t5
t10
normal
20
probability

10
0

0.0 0.1 0.2 0.3 0.4 0.5

Distribution of (Q) for exchangeable Gaussian and t copulas;

π = 0.04 and ρ = 0.3.

c 2003 (Embrechts, Frey, McNeil) 25
 Mixing densities – similar default correlation
Densities of mixing distribution - fitted pi2
20

t5, rho adjusted

normal
15
probability

10
5
0

0.0 0.1 0.2 0.3 0.4 0.5

Distribution of Q for exchangeable Gaussian and t copulas; π = 0.04

and in the normal model ρ = 0.3.

c 2003 (Embrechts, Frey, McNeil) 26
 A5. Statistical Issues – Model Calibration

Methods of model calibration used in practice seem ad hoc. Very

little actual statistical fitting of credit models to historical data takes
place. Parameters, particularly those governing dependence, often
chosen using rational economic arguments, rather than estimated.

Reasons: lack of quality historical data on historical default; feeling

that the existing data (S&P or Moodys) not relevant for own
portfolio, or not relevant for the future.

c 2003 (Embrechts, Frey, McNeil) 27
 Historical Default Data
Typical Data Format:
Year Rating Companies Defaults
2000 A 317 2
B 500 25
.. .. ..
1999 A 280 1
B 560 37
For illustration consider single homogeneous group (say B–rated).
Heterogeneity can be modelled using covariates in various ways.
Suppose our time horizon of interest is one year and we have n years
of historical data {(mj , Mj ) , j = 1, . . . , n}, where mj denotes the
number of obligors observed in year j and Mj is the number of these
that default.

c 2003 (Embrechts, Frey, McNeil) 28
 Statistical Approaches

Assume an exchangeable Bernoulli mixture model in each year

period with Q1, . . . , Qn identically distributed.

Method 1: Maximum Likelihood (Assume independence of Qi)

Parameters of mixing distribution (e.g. beta, logit–, or

probit–normal) can be estimated by maximum likelihood.
Particularly easy for beta: M1, . . . , Mn have a beta–binomial
distribution with probability function:
 
m
P (M = k) = β(a + k, b + m − k)/β(a, b) .
k

c 2003 (Embrechts, Frey, McNeil) 29
 Method 2: Moment Estimation

We have seen the importance of π = E(Q) and π2 = E(Q2) (or ρY )

in homogeneous groups. How do we estimate these moments?
M


Lemma. Let k be (random) number of subgroups of k companies
in those that default. Then E M m



k = k πk .
M

P
Proof. k = (i1,...,ik )⊂(1,...,m) Yi1 · · · Yik .

An unbiased and consistent estimator of πk is

n
1 X Mj (Mj − 1) · · · (Mj − k + 1)
π
bk = , k = 1, 2, 3, . . . .
n j=1 mj (mj − 1) · · · (mj − k + 1)

c 2003 (Embrechts, Frey, McNeil) 30
A6. Implications for pricing basket credit derivatives
Insights on dependence–modelling for loan portfolios have also
implications for pricing of basket credit derivatives. Consider
portfolio with m obligors (the basket) held by bank A. We are
interested in pricing of following stylized default swap:
Second to default swap: Fix horizon {T }. Bank A receives from
counterparty B a fixed payment K at time T if at least two obligors
in the basket have defaulted (i.e. had a credit event) until time T ;
otherwise it receives nothing. At t = 0 A pays to B a fixed premium.
Intuition: pricing sensitive to occurrence of joint defaults.
Remark: Real second–to–default swaps are more complicated. The
payments depend on identities of defaulted counterparties; moreover,
payment due at time of credit event.

c 2003 (Embrechts, Frey, McNeil) 31
 A pricing model

Stylized version of reduced-form model à la Duffie–Singleton or

Jarrow–Lando–Turnbull. Our simplifications:
– interest–rate r is deterministic
– default-intensities are rv’s instead of processes.
Denote by τi the default–time of obligor i in the basket.
Assumption 1: The default–times τi, 1 ≤ i ≤ m follow a mixed
exponential distribution, i.e. there is some p–dimensional random
vector Ψ (p < m) such that conditional on Ψ the τi are independent
exponentially distributed rv’s with parameter λi(Ψ). In particular,

P (τi < T | Ψ) = 1 − exp (−λi(Ψ) T ) ≈ λi(Ψ)T . (1)

Defaults then follow a Bernoulli–mixture model with π as in (1).

c 2003 (Embrechts, Frey, McNeil) 32
 Pricing of credit–derivatives

Following standard–practice we assume that Assumption 1 holds

under a pricing–measure Q. Hence for every claim H depending on
τ1, . . . , τm the price at t = 0 equals

P0 = e−rT E (H (τ1, . . . , τm)) .

In particular we get for our second–to–default swap

m
!
X
−rT
P0 = e Q Yi ≥ 2 .
i=1

c 2003 (Embrechts, Frey, McNeil) 33
 Specific model:

We choose λ and Ψ so that the one–year default probability

corresponds to the default–probability in the one–factor latent
variable model with t copula, i.e.
p √ !!
t−1
ν (π) W/ν − ρΘ
λi = − ln 1 − Φ √ ,
1−ρ
Θ ∼ N (0, 1), W ∼ χ2(ν) .

c 2003 (Embrechts, Frey, McNeil) 34
 Simulations:

Homogeneous portfolio with m = 14, T = 1, and varying values for

default probability π and asset correlation ρ.
Portfolio A: π = 0.15% ρ = 0.38%
Portfolio B: π = 0.50% ρ = 3.80%
t
 normal
In the following table we give the ratio P0 P0 of the price of
stylized second–to–default swap in in t–model and normal model.
Portfolio ν = 5 ν = 10 ν = 20
A 11.0 7.3 4.4
B 3.3 2.6 2.0
Choice of the copula has again drastic effect!

c 2003 (Embrechts, Frey, McNeil) 35
 Conclusions

• Extreme risk in latent variable models is driven by copula of X.

• The assumption of a multivariate normal distribution and
a calibration based on asset correlations alone may seriously
underestimate the extreme risk in latent variable models.
• Extreme risk in Bernoulli mixture models with factor structure is
driven by the mixing distribution of the factors.
• The two model types may often be mapped into one another. It is
particularly useful (Monte Carlo simulation and also for fitting) to
represent latent variable models as Bernoulli mixture models.
• Model calibration should use historical default data and not be
based solely on assumptions about asset value correlations.

c 2003 (Embrechts, Frey, McNeil) 36
 References

On credit risk modelling:

• [Gordy, 2000] A comparative evaluation of credit risk models

• [Crouhy et al., 2000] A comparison of current credit risk models.

• [Frey and McNeil, 2001] paper underlying the presentation

c 2003 (Embrechts, Frey, McNeil) 37
Bibliography

[Crouhy et al., 2000] Crouhy, M., Galai, D., and Mark, R. (2000).
A comparative analysis of current credit risk models. Journal of
Banking and Finance, 24:59–117.

[Frey and McNeil, 2001] Frey, R. and McNeil, A. (2001). Modelling

dependent defaults. Preprint, ETH Zürich. available from
http://www.math.ethz.ch/~frey.

[Gordy, 2000] Gordy, M. (2000). A comparative anatomy of credit

risk models. Journal of Banking and Finance, 24:119–149.

c 2003 (Embrechts, Frey, McNeil) 38