2.

Default correlation

Correlation of defaults of a pair of risky assets

Consider two obligors A and B and a fixed time horizon T .

pA = probability of default of A before T
pB = probability of default of B before T
pAB = joint default probability that A and B default before T
pA|B = probability that A defaults before T , given that B has
defaulted before T
pAB pAB
pA|B = , pB|A =
pB pA
ρAB = linear correlation coefficient between default events
h i h i h i
E 1{A}1{B} −E 1{A} E 1{B}
=
σ1{A} σ1{B}
pAB − pApB
= q ,
pA(1 − pA)pB (1 − pB )
(
1 A defaults before T
where 1{A} = 0 otherwise
Since default probabilities are very small, the correlation ρAB can
have a much larger effect on the joint risk of a position
q
pAB = pApB + ρAB pA(1 − pA)pB (1 − pB )
s
pA
pA|B = pA + ρAB (1 − pA)(1 − pB ) and
pB
s
pB
pB|A = pB + ρAB (1 − pA)(1 − pB ).
pA

Assume ρAB = ρ which is not small but pA = pB = p ¿ 1, then

pAB ≈ p2 + ρp ≈ ρp
pA|B ≈ ρ.

The joint default probability and the conditional default probability

are dominated by the correlation coefficient ρ.

• When there are 2 obligors, we can compute the probabilities of

all elementary events by using the linear correlation coefficient.
Limitations of linear correlation coefficient

• With 3 obligors, we have 8 elementary events but only 7 re-

strictions (3 individual probabilities, 3 correlations and sum of
probabilities). The probability of the joint default of all 3 oblig-
ors is not determined by the 3 pairs of correlation.

• For N obligors, we have N (N − 1)/2 correlations, N individ-

ual default probabilities. Yet we have 2N possible joint default
events. The correlation matrix only gives the bivariate marginal
distributions, while the full distribution remains undetermined.

• For any pair of random variables X and Y

cov(1 − X, 1 − Y ) = cov(X, Y )
so that the linear correlation coefficient between survival events
is the same as that between default events. Hence, it is unable
to capture the fact that risky assets exhibit greater tendency to
crash together than to boom together.
Price bounds for first-to-default (FtD) swaps under low default cor-
relation

fee on CDS on ≤ fee on FtD ≤ portfolio of

worst credit swap CDSs on all
credits
sC ≤ sFtD ≤ sA + sB + sC

With low default probabilities and low default correlation

sFtD ≈ sA + sB + sC .
To see this, the probability of at least one default is
p = 1 − (1 − pA)(1 − pB )(1 − pC )
= pA + pB + pC − (pApB + pApC + pB pC ) + pApB pC
so that
p / pA + pB + pC for small pA, pB and pC .
 Counterparty risk in CDS
Before the Fall 1997 crisis, several Korean banks were willing to offer
credit default protection on other Korean firms.

40 bp
US commercial Korea exchange
bank bank

LIBOR + 70bp

Hyundai
(not rated)

* Higher geographical risks lead to higher default correlations.

Advice: Go for a European bank to buy the protection.

Nature of counterparty risk

1. Replacement cost
• If the Protection Seller defaults prior to the Reference En-
try, then the Protection Buyer renews the CDS with a new
counterparty.
• Supposing that the default risks of the Protection Seller and
Reference Entity are positively correlated, there will be an
increase in the swap rate in the new CDS.

2. Settlement risk
• The Protection Seller defaults during the settlement period
after the default of the Reference Entity.
 Funding cost arbitrage – Credit default swap

A-rated institution 50bps AAA-rated institution LIBOR-15bps Lender to the

as Protection Seller annual as funding AAA-rated
as Protection Buyer
premium cost Institution
funding cost of coupon
LIBOR + 50bps = LIBOR + 90bps

Lender to the BBB risky

A-rated Institution reference asset

In order that the funding cost arbitrage works, the funding cost of
the default protection seller must be higher than that of the default
protection buyer.
The combined risk faced by the Protection Buyer:

• default of the BBB-rated bond

• default of the Protection Seller on the contingent payment

The AAA-rated Protection Buyer creates a synthetic AA-asset with

a coupon rate of LIBOR + 90bps − 50bps = LIBOR + 40bps.

This is better than LIBOR + 30bps, which is the coupon rate of a

AA-asset (net gains of 10bps).

For the A-rated Protection Seller, it gains synthetic access to a

BBB-rated asset with earning of net spread of

50bps − [(LIBOR
|
+ 90bps) −
{z
(LIBOR + 50bps)}] = 10bps
the A-rated Protection Seller earns 40bps
if it owns the BB asset directly
Need for theoretical models of default correlations

Models for default correlation must be able to explain and predict

default correlations from few and more fundamental variables than
a simple full description of the joint default distribution.

Several possible sources of data, none of which is perfect.

• Historically observed joint rating and default events

• Credit spreads

• Equity correlations

1. We simply do not have direct data on default dependencies.

2. Specification of full joint default probabilities is simply too com-

plex. There are 2N joint default events for N obligors.
Poisson model of default

Suppose default of obligor i is modeled by a Poisson-distributes

random variable: Li ∼ Pois(λi).

Probability of default = pi = Pr [Li ≥ 1]

= 1 − e−λi ≈ λi for small λi.

Interacting intensities

Two-firm contagion model

t = a0 + a1 1{τ B ≤t}
λA

t = b0 + b11{τ A ≤t}
λB

The default intensity λA B

t (λt ) jumps by the amount a1 (b1) when
Firm B(A) defaults.
Stochastic intensities

Default intensities for different companies follow correlated stochas-

tic processes:
q
dλj (t) = K[θ − λ(t)] dt + σj λ(t) dZj dZj dZk = ρjk dt.

λj (t) is said to follow the CIR process and ρjk is the correlation
coefficient between the Brownian motions.
Contagion model

Contagion means that once a firm defaults, it may bring down other
firms with it.

Davis-Lo model (2001)

Yij is an “infection” variable which, when equal to one, implies that

default of Firm i immediately triggers default of Firm j.

Assume Xi, Yij are independent Bernuolli variables with

P [Xi = 1] = p and P [Yij = 1] = q.
Define Zi be the default indicator of Firm i
 
Y
Zi = Xi + (1 − Xi) 1 − (1 − Xj Yji) .
j6=i
Note that Zi equals one either when there Y is a direct default of
firm i or if there is no direct default and (1 − Xj Yji) = 0. The
j6=i
latter case occurs when at least one of the factor Xj Yji is 1, which
happens when firm j defaults and infects firm i.

Define Dn = Z1 + · · · + Zn, Davis and Lo (2001) find that

E[Dn] = n[1 − (1 − p)(1 − pq)n−1]
var(Dn) = n(n − 1)βnpq − (E[Dn])2
where
βnpq = p2 + 2p(1 − p)[1 − (1 − q)(1 − pq)n−2]
+(1 − p)2[1 − 2(1 − pq)n−2 + (1 − 2pq + pq 2)n−2].

cov(Zi, Zj ) = βnpq − var(Dn/n)2.

Moody’s binomial expansion technique

• Diversity score, weighted average rating factor and binomial ex-

pansion technique.

• Generate the loss distribution.

To build a hypothetical pool of uncorrelated and homogeneous as-

sets that mimic the default behaviors of the original pool of corre-
lated and inhomogeneous assets.

Additional assumptions

• Every instrument in the comparison portfolio can be uniquely

assigned to one industry group.

• Two instruments in the comparison portfolio have positive cor-

relation if and only if they belong to the same industry group.
Moody’s diversity score

The diversity score of a given pool of participations is the number n

of bonds in a idealized comparison portfolio that meets the following
criteria:

• Comparison portfolio and collateral pool have the same face

value.

• Bonds in the comparison portfolio have equal face values.

• Comparison bonds are equally likely to default, and their default

is independent.

• Comparison bonds are of the same average default probability

as the participations of the collateral pool.

• Comparison portfolio has, according to some measure of risk,

the same total risk as does the collateral pool.
Binomial probability formula

Once the “average” default probability p is known, then the prob-

ability of k defaults out of n bonds will be given by the probability
formula of k successes out of n independent trials.
n!
P (k defaults) = pk (1 − p)n−k
(n − k)!k!
where default probability
n
X
piFi
p = n=1
n .
X
Fi
i=1
 Binomial mixture model

Mixture distribution randomizes the default probability of the bino-

mial model to induce dependence, thus mimicking a situation where
a common background variable affects a collection of firms. The
default events of the firms are then conditionally independent given
the mixture variable.

Binomial distribution

Suppose X is binomially distributed (n, p), then

E[X] = np and var(X) = np(1 − p).
We randomize the default parameter p. Recall the following relation-
ships for random variables X and Y defined on the same probability
space
E[X] = E[E[X|Y ]] and var(X) = var(E[X|Y ]) + E[var(X|Y )].
Suppose we have a collection of n firms, Xi = Di(T ) is the default
indicator of firm i. Assume that pe is a random variable which is
independent of all the Xi. Assume that pe takes on values in [0, 1].
Conditional on p,e X1, · · · , Xn are independent and each has default
probability p.
e Let p denote the mean of p, e where
Z 1
p = E[p]
e = pf (p) dp.
0
We have
E[Xi] = p and var(Xi) = p(1 − p)
and
cov(Xi, Xj ) = E[pe2] − p2, i 6= j.

(i) When pe is a constant, we have zero covariance.

(ii) By Jensen’s inequality, cov(Xi, Xj ) ≥ 0.
(iii) Default event correlation
E[pe2] − p2
ρ(Xi, Xj ) = .
p(1 − p)
 n
X
Define Dn = Xi, which is the total number of defaults; then
i=1

E[Dn] = np and var(Dn) = np(1 − p) + n(n − 1)(E[pe2] − E[p]

e 2)

(i) When pe = p, corresponding no randomness, var(Dn) = np(1−p),

like usual binomial distribution.
(ii) When pe = 1 with prob p and zero otherwise, then var(Dn) =
n2p(1 − p), corresponding to perfect correlation between all de-
fault events.
(iii) One can obtain any default correlation in [0, 1]; correlation of
default events depends only on the first and second moments of
f . However, the distribution of Dn can be quite different.
µ ¶
Dn p(1 − p) n(n − 1)
(iv) var = + 2
var(p)
e −→ var(p)
e as n → ∞, that
n n n
is, when considering the fractional loss for n large, the only
remaining variance is that of the distribution of p.
e
Copula approach for modeling default dependency

Two aspects of modeling the default times of several obligors

1. Default dynamics of a single obligor.

2. Model the dependence structure of defaults between the oblig-
ors.

Question How to specify a joint distribution of survival times, with

given marginal distributions?

• Knowing the joint distribution of random variables allows us to

derive the marginal distributions and the correlation structure
among the random variables but not vice versa.

• A copula function links univariate marginals to their full multi-

variate distribution.
Proposition

If an one-dimensional continuous random variable X has distribution

function F , that is, F (x) = P [X ≤ x], then the distribution of the
random variable U = F (X) is a uniform distribution on [0, 1].

Remark

To simulate an outcome of X, one may simulate an outcome u from

a uniform distribution then let the outcome of X be x = F −1(u).
Multi-variate distribution function
FX1,X2,··· ,Xn (x1, x2, · · · , xn) = P [X1 ≤ x1, X2 ≤ x2, · · · , Xn ≤ xn]

• It is a non-decreasing, right continuous function which maps a

subset of the real numbers into the unit interval [0, 1].
• Monotonicity property for vector a and b
a < b ⇒ F (b) − F (a) ≥ 0
a < b means b − a is a vector with non-negative entries and at
least one strictly positive entry.
 Definition of a copula function

A function C : [0, 1]n → [0, 1] is a copula if

(a) There are random variables U1, U2, · · · , Un taking values in [0, 1]
such that C is their distribution function.
(b) C has uniform marginal distributions; for all i ≤ n, ui ∈ [0, 1]
C(1, · · · , 1, ui, 1, · · · , 1) = ui.

In the analysis of dependency with copula function, the joint dis-

tribution can be separated into two parts, namely, the marginal
distribution functions of the random variables (marginals) and the
dependence structure between the random variables which is de-
scribed by the copula function.
 Construction of multi-variate distribution function

Given univariate marginal distribution functions F1(x1), F2(x2), · · · , Fn(xn),

the function
C(F1(x1), F2(x2), · · · , Fn(xn)) = F (x1, x2, · · · , xn)
which is defined using a copula function C, results in a multivariate
distribution function with univariate marginal distributions specified
as F1(x1), F2(x2), · · · , Fn(xn).
 • Any multi-variate distribution function F can be written in the
form of a copula function.

Theorem

If F (x1, x2, · · · , xn) is a joint multi-variate distribution function with

univariate marginal distribution functions F1(x1), · · · Fn(xn), then
there exists a copula function C(u1, u2, · · · , un) such that
F (x1, x2, · · · , xn) = C(F1(x1), F2(x2), · · · , Fn(xn)).
If each Fi is continuous, then C is unique.

Going through all copula functions gives us all the possible types
of dependence structures that are compatible with the given one-
dimensional marginal distributions.
 Bivariate normal copula function

C(u, v) = N2(N −1(u), N −1(v); γ), −1 ≤ γ ≤ 1.

Suppose we use a bivariate normal copula function with a correlation
parameter γ, and denote the default times for A and B as TA and
TB . The joint default probability is given by
P [TA < 1, TB < 1] = N2(N −1(FA(1)), N −1(FB (1)), γ) (B)
where FA and FB are the distribution functions for the default times
TA and TB .
We observe that
qi = P [Ti < 1] = Fi(1) and Zi = N −1(qi) for i = A, B,
Eqs. (A) and (B) are equivalent if we have ρ = γ.

Note that this correlation parameter is not the correlation coefficient

between the two default times.
 Simulation of default times of a basket of obligors

Assume that for each credit i in the portfolio, we have constructed

a credit curve or a hazard rate function for its default time Ti. Let
Fi(t) denote the distribution function of Ti.

Using a copula function C, we obtain the joint distribution of the

default times
F (t1, t2, · · · , tn) = C(F (t1), F2(t2), · · · , Fn(tn)).
For example, suppose we use the normal copula function, we have
F (t1, t2, · · · , tn) = Nn(N −1(F1(t1)), N −1(F2(t2)), · · · , N −1(Fn(tn))).
To simulate correlated default times, we introduce
Y1 = N −1(F1(T1)), Y2 = N −1(F2(T2)), · · · , Yn = N −1(Fn(Tn)).
There is a one-to-one mapping between Y and T . This provides the
linkage between the random default time T and the credit index Y
(modeled as standard normal random variable).
Simulation scheme

• Simulate Y1, Y2, · · · , Yn from an n-dimensional normal distribu-

tion with correlation coefficient matrix Σ.

• Obtain T1, T2, · · · , Tn using Ti = Fi−1(N (Yi)), i = 1, 2, · · · , n.

With each simulation run, we generate the default times for all the
credits in the portfolio. With this information we can value any
credit derivative structure written on the portfolio.
 CreditMetrics

• CreditMetrics uses the normal copula function in its default cor-

relation formula even though it does not use the concept of
copula function explicitly.

• CreditMetrics calculates joint default probability of two credits

A and B using the following steps:
(i) Let qA and qB denote the one-year default probabilities for A
and B, respectively. Let ZA and ZB denote the credit index of A
and B, respectively, both are standard normal random variables.
qA = P [ZA < zA] and qB = P [ZB < zB ]

(ii) Let ρ denote the asset correlation, the joint default probability
for credit A and B is given by
Z z Z z
A B
P [ZA < zA, ZB < zB ] = n2(x, y; ρ) dxdy = N2(zA, zB ; ρ).
−∞ −∞
(A)
 Factor copula model

Consider a portfolio of N companies and assume that the marginal

risk neutral probabilities of default are known for each company.
τi :
time of default of company i
Qi(t) :
cumulative risk neutral probability that company i will default
before time t
Si(t) = 1 − Qi(t) = risk neutral survival probability.

To generate a one-factor model for τi, we define

q
xi = aiM + 1 − a2 i Zi (1)
where M is the common factor and Zi represents the idiosyncratic
risk.
We take M and Zi to be independent zero-mean unit-variance dis-
tributions and −1 ≤ ai ≤ 1.

Eq. (1) defines a correlation structure between xi dependent on a

common factor M . The correlation between xi and xj is aiaj .

Let Fi denote the cumulative distribution of xi. Under the copula

model, xi is mapped to τi using percentile-to-percentile transforma-
tion. The point xi = x is transformed to τi = t so that
t = Q−1
i [Fi (x)].
Assuming Zi to be identically distributed, we let H denote the cu-
mulative distribution of Zi.
 
 x − aiM 
Pr [xi < x|M ] = H  q .
2
1 − ai

When x = Fi−1[Qi(t)], Pr [τi < t] = Pr [xi < x] so that

 
−1
 Fi (Qi(t)) − aiM 
Pr [τi < t|M ] = H  q .
1 − ai2

Conditional survival probability that the ith obligor will survive be-
yond T
 
−1
 i (Qi(T )) − aiM 
F
Si(T |M ) = 1 − H  q .
1 − a2
i
Extension to multi-factor

q
xi = ai1M1 + · · · + aimMm + Zi 1 − a2
i1 − a2 − · · · − a2
i2 im
 
−1
 Fi (Qi(T )) − ai1M1 − · · · − aimMm 
Si(T |M1, M2, · · · , Mm) = 1 − H  q .
2 2
1 − ai1 − · · · − aim

Define πT (k) = probability that exactly k defaults before T

N
Y
πT (0|M1, M2, · · · , Mm) = Si(T |M1, M2, · · · , Mm).
i=1
πT (1|M1, M2, · · · , Mm) = πT (0|M1, M2, · · · , Mm)
N
X 1 − Si(T |M1, M2, · · · , Mm)
.
i=1 Si(T |M1, M2, · · · , Mm)

Calculations of πT (k|M1, M2, · · · , Mm) for k ≥ 2 can be relegated to

multinomial formula.
Spearman’s rho correlation

It is simply the linear correlation of the probability-transformed ran-

dom variables. It can be expressed in terms of copula K τ as follows:
ρS (τ1, τ2) = ρ(F1(τ1), F2(τ2))
Z 1Z 1
= 12 [K τ (u, v) − uv] dudv.
0 0

Kendall’s tau correlation

ρτ (τ1, τ2) = E[sign(τ1 − τe1)(τ2 − τe2)]
where (τe1, τe2) is an independent copy of (τ1, τ2).
Z 1Z 1
ρτ (τ1, τ2) = 4 K τ (u, v) dK τ (u, v) − 1.
0 0
 CreditRisk+

• A credit risk model developed by Credit Suisse Financial Prod-

ucts (CSFP).

• A typical representative of the group of Poisson mixture models.

• Gamma distribution
1 −x/β xα−1,
γα,β (x) = e x≥0
β αΓ(α)

E[Λ] = αβ and var(Λ) = αβ 2.

• Sector model
Every sector could be thought of as generated by a single under-
lying factor. Sectors can be identified with industries, countries
or regions, or any other systematic influence on the economic
performance of counterparties with a positive weight in this sec-
tor.
1. Each sector s ∈ {1, · · · , ms} has its own gamma-distributed ran-
dom intensity
Λ(s) ∼ Γ(αs, βs)
and Λ(1), · · · , Λ(mS ) are assumed to be independent.

2. Every obligor i admits a breakdown into sector weights wis ≥ 0

with
m
X S
wis = 1.
s=1

3. The risk of sector s is captured by two parameters

(i) mean default intensity of the sector

λ(s) = E[Λ(s)] = αsβs

(ii) default intensity’s volatility

2 = var(Λ(s) ) = α β 2.
σ(s) s s
Every obligor i admits a random default intensity Λi with mean value
E[Λi] = λi
which could be calibrated to the obligor’s one-year default probabil-
ity.

The sector parametrization of Λi

m
X S Λ(s)
Λi = wisλi , i = 1, 2, · · · , m.
s=1 λ(s)

Two obligors are correlated iff there is at least one sector such
that both obligors have a positive sector weight with respect to this
sector – admit a common source of systematic default risk.
Mixed Poisson random variable L0s

The default risk of obligor i is modeled by a mixed Poisson random

variable L0s with random intensity Λi.

Any conditional default intensity of obligor i arising from realiza-

tions θ1, · · · , θmS of the sector’s default intensities Λ(1), · · · , Λ(mS )
generates a conditional one-year default probability pi(θ1, · · · , θmS )
of obligor i by
pi(θ1, · · · , θmS ) = P [L0i ≥ 1|Λ(1) = θ1, · · · , Λ(mS ) = θmS ]
PmS
−λi s=1 wis θs /λ(s) .
= 1−e
Summary

A good portfolio credit risk model should have the following prop-
erties

• Default dependence – produce default correlations of a realistic

magnitude.

• Estimation – number of parameters should be limited.

• Timing risk – producing “clusters” of defaults in time, several

defaults that occur close to each other

• Calibration (i) Individual term structures of default probabilities

(ii) Joint defaults and correlation information

• Implementation - existence of a variable implementation mech-

anism, say Monte Carlo simulation method