A new approach to the modelling and pricing of

correlation credit derivatives

Giuseppe Di Graziano∗
Statistical Laboratory
University of Cambridge

L.C.G Rogers†
Statistical Laboratory
University of Cambridge

29th June 2005

Abstract

The modelling of credit events is in effect the modelling of the times to default
of various names. The distribution of individual times to default can be calibrated
from CDS quotes, but for more complicated instruments, such as CDOs, the joint
law is needed. Industry practice is to model this correlation through a copula,
an approach with significant deficiencies. We present a new approach to default
correlation modelling, where defaults of different names are driven by a common
continuous-time Markov process. Individual default probabilities and default
correlations can be calculated in closed form. As illustrations, the prices of K th -
to-default baskets and CDO tranches with name-dependent random losses are
computed using Laplace transform techniques.

∗
Wilberforce Road, Cambridge CB3 0WB, UK (phone = +44 1223 9798, e-mail =
[email protected]).
†
Wilberforce Road, Cambridge CB3 0WB, UK (phone = +44 1223 766806, fax = +44 1223 337956,
e-mail = [email protected]).

1
1 Introduction

The current industry approach to the pricing of multi-name credit derivatives makes
use of copula functions to model the dependence between issuers in a given portfolio of
defaultable securities. This approach is problematic for two main reasons: there is no
dynamic consistency, and there is no theoretical basis for the choice of any particular
dependence structure. The root cause of the problems is bad modelling - the depen-
dence is forced into the model at the very last stage, rather than growing organically
from the modelling assumptions.
What we propose here is a new approach to the problem based on the use of a Markov
process within the reduced-form framework. This completely deals with the main
problems of the current copula-based approach. Default correlation is determined from
market data by fitting the model to CDS and portfolio derivative data.
We start by assuming there exists a process (ξt )t≥0 which drives the common dynamics
of the credits in the portfolio. We then model the survival probabilities up to time t
of a given obligor, say i, conditional on the filtration generated by the process Ftξ as
 
ξ
P τ ≥ t|Ft = exp −Cti ,
i


(1)

where τ i indicates the default time of the ith reference entity, and C i is an additive
functional of the process. The simplest thing1 for this is to take
Z t
i
Ct = λi (ξu )du
0

where λi (ξ) is a deterministic function of the chain, which we will refer to as the
(default) intensity (function) of entity i. For simplicity, we shall limit our discussion to
the case where (ξt )t≥0 is a continuous-time finite-state Markov chain. This framework
is already sufficiently flexible for practical purposes, and is simple enough to allow
explicit computation using fast linear algebra routines.
Taking (1) as a starting point, it is easy to derive the individual conditional default
probabilities in closed form. It is also straightforward to compute default correlations.
Moreover, we show how to obtain a fast and reasonably accurate approximation to the
price of CDO tranches and K th -to-default baskets based on a Poisson approximation.
Exact solutions can be obtained by computing the Laplace transform of the portfo-
lio loss distribution and related quantities and then resorting to numerical inversion
techniques.
1
... but as we shall see, not the only thing ...

2
2 Model specification and basic results.

In this section we introduce the main modelling ideas of the paper which will form the
basic building blocks for the pricing of multi-name credit derivatives.
Consider a portfolio of N defaultable securities and assume that there exist a continuous-
time finite-state irreducible Markov chain (ξt )t≥0 with Q-matrix Q, generating a filtra-
tion Ftξ . Assume that conditional on the path of the chain, defaults of the N names
will be independent, the survival probability of the ith reference entity being given by
 
qti = P τ i ≥ t|Ftξ = exp −Cti ,


(2)
where Cti is some additive functional of the chain of the form
Z t X
i
Ct = λi (ξu )du + i
wjk Jjk (t). (3)
0 j6=k

Here, τ i is the default time of the ith name in the portfolio, λi is a deterministic function
of the chain, Jjk (t) denotes the number of jumps by time t from state j to state k, and
i
the wjk are non-negative weights.
In order to gain some intuition, one could think of the chain as representing the state
of health of the economy. If the chain jumps from a state of economic growth to a state
of recession, this may cause the conditional default intensity of some of the reference
entities to go up, increasing the chances of observing a larger number of defaults in the
portfolio. The jump itself may also trigger defaults. Note that the information about
how the various credits in the portfolio are correlated is contained in the λi , the wjk
i
,
and Q. An expression for the dynamic default correlation will be derived in section 3.
Throughout this paper we will also assume that the money market account takes the
following form Z t 
Bt = exp r(ξu )du , (4)
0
where again r is a deterministic function of the chain. Routine calculations allow us
to recover the discounted survival probability of the ith reference entity. In particular,
for any function g we have that
q̃ti (ξ0 ) ≡ E 1{τ i ≥t} Bt−1 g(ξt ) | ξ0
 
" Z t ! #
X
= E exp − (λi + r)(ξu )du − i
wjk Jjk (t) g(ξt )
0 j6=k

= exp(tQ̃i )g(ξ0 ), (5)

3
where

Q̃ijk = qjj − (λij + rj ) (j = k);

i
= exp(−wjk )qjk (j 6= k).

Taking g = 1, the vector all of whose entries are 1, gives the discounted survival
probabilities E 1{τ i ≥t} Bt−1 | ξ0 . Note that survival probabilities depend on the current
state of the chain ξ0 .
When we move to a portfolio of N defaultable securities, we need to be able to find the
distributions of more complicated random variables. For example, if `i ≡ Ai (1 − Ri )
denotes the loss on the ith name, in terms of the notional Ai and the (possibly random)
recovery rate Ri , then the portfolio cumulative loss process
N
X
Lt ≡ `i I{τi ≤t} (6)
i=1

is an object of interest. Apart from Monte Carlo, the only tools available to find the
law of Lt are based on transforms. By conditioning firstly on the path of the chain, it
is easy to see that the (discounted) Laplace transform of Lt is given by
Z t " Z t N
#
Y
(1 − qti )ζi (α) + qti


E exp(− r(ξs )ds − αLt ) = E exp(− r(ξs )ds) , (7)
0 0 i=1

where qti is given by (2) and ζi (α) = E[e−α`i ]. This is the key relation linking our
modelling approach at an abstract level to the kinds of calculation needed to price
credit derivatives of various sorts. The (numerical) inversion of the Laplace transform
(7) is the common first step; the method of Hosono [Ho], popularised by Abate &
Whitt [AW], [AW1], is a fast and accurate solution. We discuss numerical approaches
in Section 4, but before that we record the form of default correlation given by this
approach.

3 Default correlation

In the previous Section, we obtained an expression for the survival probability of a single
obligor. The argument used extends easily to more than one obligor, so we present
here the corresponding results, and derive the correlation of defaults from that.

4
Assume for example we want to compute the joint survival probability of obligors i
and j. Using the independence of default times given ξ, we obtain

q̃tij (ξ0 ) ≡ P (τ i ≥ t, τ j ≥ t | ξ0 )
= exp(tQ̃ij )1(ξ0 ),

where

Q̃ij i
kl = qkk − λk − λk
j
(k = l);
i j
= exp(−wkl − wkl )qkl (k 6= l).

Elementary algebraic calculations allow us to recover the default correlation ρT (ξt )

of i and j from the joint survival probability function and the individual survival
probabilities:
q̃tij (ξ0 ) − q̃ti (ξ0 )q̃tj (ξ0 )
ρt (ξ0 ) = p q (8)
i i j j
q̃t (ξ0 )(1 − q̃t (ξ0 )) q̃t (ξ0 )(1 − q̃t (ξ0 ))
where
q̃ti (ξ0 ) = exp(tQ̃i )1(ξ0 ) (9)
as at (5).
Note that the correlation of defaults is obtained endogenously from the model, rather
than being exogenously imposed as in the copula-based industry approach to default
correlation.

4 Computational approaches

Our attention now focuses on the exact expression (7) for the discounted Laplace
transform of the cumulative loss at time t. We have good techniques for inverting the
transform, but first we have to be able to calculate it (or some approximation), and in
this Section we discuss three possible approaches.

4.1 Exact method

The approach here is to multiply out the product on the right-hand side of (7). The
individual terms are all quite easy to deal with, because each is the exponential of some
additive functional of the Markov chain, and we are able to compute these expectations

5
using fast linear algebra routines. The problem with this approach comes when the
number N of names gets too big; with N names there are 2N terms in the product
when multiplied out, and each of these needs to be evaluated and inverted separately.
When N = 10 there are 1024 such calculations, and typically we need to be able to
handle values of N that are an order of magnitude bigger. Thus the ‘exact’ calculation
method will be too cumbersome for general use.

4.2 Poisson approximation

The expression (2) for the survival probability of name i can be understood in terms
of a standard Poisson process ν independent of the chain ξ. If the jump times of ν are
denoted S1 < S2 < . . ., then we may set
τ i ≡ inf{t : Cti > S1 },
and then the relation (2) holds. The Poisson approximation we propose here is to allow
name i to default more than once, at times
i
τm ≡ inf{t : Cti > Sm }, m = 0, 1, . . . .
By doing this, we arrive at an expression L̄t for the portfolio cumulative loss which
overestimates Lt , because it includes (non-existent) second and subsequent losses of
each of the names. The error we are committing by this is of the same order as the
default probabilities themselves; typically this would be of the order of a few percent,
which would be comparable to the error we could expect from a Monte Carlo approach.
Some more sophisticated method for the estimation and control of this error is needed,
but the virtues of this simplified approach are evident on inspection of the expression
for the Laplace transform of the cumulative loss L̄t :
Z t " Z t N
#
X
E exp(− r(ξs )ds − αL̄t ) = E exp exp(− r(ξs )ds) (ζi (α) − 1)Cti . (10)
0 0 i=1

For each α, we are computing the mean of the exponential of an additive functional of
the chain, and this is a simple and rapid calculation.

4.3 Monte Carlo

Another approach to calculating (7) is to use Monte Carlo simulation to evaluate the
right-hand side, and then invert the transform. We do not discuss this in detail, but
restrict ourselves to a few remarks concerning pitfalls to be avoided.

6
Firstly, we do not generate paths by discretising the time interval into a large number
of subintervals and then simulating the (many) individual steps of the chain; rather,
we use the jump-hold construction of the Markov chain, starting from the embedded
discrete-time jump chain with exponentially-distributed residence times in the states
passed through. This is far more efficient, and makes the calculation of additive func-
tionals of the path a triviality.
Secondly, inversion of the Laplace transform will require evaluation of the transform
at many different values of α; we do not of course simulate a different chain for each
value of α, but keep the same chain for all evaluations.
Thirdly, if the Monte Carlo approach is to be used for calibration, we will also be varying
the parameters of the model, including the parameters of the chain; our simulation
must be done in such as way as to keep the dependence on the parameters as smooth
as possible.

5 Example: K th-to-default basket

To illustrate the methods presented, we will show how to price a K th -to-default basket
under the simplifying assumption that there is a homogeneous recovery rate R.
PN
The cumulative loss process Lt = i=1 1{τ i ≤t} is now just the number of defaults
incurred by time t, with τK denoting the time of the K th default. Let ∆i = Ti − Ti−1
be the year fraction between dates Ti−1 and Ti . As with a plain vanilla CDS, a default
basket consists of a premium leg (a protection fee stream paid up til the K th default)
and a default leg (the loss induced by the K th default).

5.1 Premium leg

If for simplicity we ignore accrued payments at default, the risky PV01 of the premium
leg of a K th -to-default basket (that is, the present value of a contract paying 1bp at each
payment date prior to default) maturing at time T , and paying at dates T1 , . . . , TM , is

7
given by
M
X   Z Ti  
K
P V 01 = ∆i E exp − r(ξu )du ; τ ≥ Ti (11)
i=1 0

XM   Z Ti  
= ∆i E exp − r(ξu )du ; LTi ≤ K − 1 (12)
i=1 0

and calculating such expressions amounts to inverting the discounted Laplace transform
(7). Since Lt has a discrete distribution, we can get a better level of accuracy by using
the methodology developed by Abate-Whitt for probability generating functions (see
[AW1]).

5.2 Default leg

The default leg of a K th -to-default basket is the present value DL of the expected
losses caused by the K th default. Since we have assumed all names in the portfolio are
characterized the same (random) recovery, we do not need to know the identity of the
K th firm-to-default, which simplifies the calculation significantly. Standard calculations
show that
 Z τK 
DL
= E exp(− r(ξs )ds); τK < T
E(1 − R) 0
 Z τK ∧T   Z T 
= E exp(− r(ξs )ds) − E exp(− r(ξs )ds); τK > T
0 0
Z τK ∧T 
r(ξs )Bs ds − E BT−1 ; LT < K
−1
 
= 1−E
0
Z T 
r(ξs )Bs I{Ls < K}ds − E BT−1 ; LT ≤ K − 1 .
−1
 
= 1−E
0

All of the terms appearing in this expression can in principle be computed from invert-
ing the discounted Laplace transform (7), though we will need to do an integral with
respect to s just to finish things off.
Finally the par-spread S of a K th -to-default basket is then given by the ratio of the
default leg and the risky PV01, i.e.

DL
S= . (13)
P V 01

8
6 Example: synthetic CDOs

We turn now our attention to the problem of pricing a CDO tranche, and find the
techniques developed so far will again serve. As before, we derive first the value of the
premium leg and then the value of default leg.

6.1 Premium leg

Let L+ and L− be the upper and lower attachment points of the tranche respectively.
At each payment date, investors receive a coupon which is proportional to the notional
of the tranche, net of the losses suffered by the credit portfolio up to that point. The
tranche PV01 is equal to

M
X   Z t  
P V 01 = ∆i E exp − r(ξu )du Φ(LTj ) , (14)
j=1 0

where
1 h
+

+ −

+ i
Φ(x) = + L −x − L −x , (15)
L − L−
and M is the number of total payments occurring at dates T1 , . . . , TM . In order to
evaluate the PV01, we need to calculate the price of a portfolio of put options on
the portfolio cumulative losses at each payment date Tj . In particular, Φ(x) is the
difference of two put options with strike L+ and L− . Elementary manipulations show
that,

Pt (K) = E BT−1 (K − Lt )+
 
  Z t  
= KE exp − r(ξu )du ; Lt ≤ K −
0
  Z t  
−E exp − r(ξu )du Lt ; Lt ≤ K
0

As before, once we can evaluate the discounted Laplace transform of the cumulative
loss Lt (7), we are able to compute both of these terms.

9
6.2 Default leg

The value of the default leg of a CDO tranche is the expected present value of the
tranche’s losses. More precisely, define
1 h
− +


i
+ +
Ξ(x) = + x − L − x − L .
L − L−
The value of the default leg of the tranche, is then given by
Z T  Z t  
DL = E exp − r(ξu )du dΞ(Lu ) .
0 0

Integrating by parts and noting that Ξ(x) = 1 − Φ(x), we can simplify the previous
expression to
  Z T   Z T  Z t  
DL = 1−E exp − r(ξu )du Φ(LT ) −E r(ξu ) exp − r(ξu )du Φ(Lu )du .
0 0 0
(16)
Again the integral in (16) can be calculated numerically. Note that the basic elements
needed to calculate the default leg are the same as the ones we derived when calculating
the premium leg, with some minor modification to account for the term r(ξ) appearing
in the second expectation of (16).
The tranche spread is recovered as usual by dividing the default leg by the PV01 of
the premium leg.
Remark. All calculations simplify if we assume interest rates the Markov chain ξ are
independent. Then all we need to do is substitute the relevant discount factor Bt−1 ,
with the corresponding riskless zero-coupon bond B(0, t) which we can observe in the
market. This assumption, albeit crude, can be useful to simplify the calibration.

7 Numerical results

We shall now present some numerical result to provide the reader with intuition about
the behavior of the model. As it should be clear from the previous sections, the
distribution of the number of defaults and the cumulative portfolio losses plays a key
role in the pricing of multi-name credit derivatives. As a consequence most of the
results in this section will be concerned with those two distributions.

10
In what follows we will assume that (ξt )t≥0 is a three state Markov chain with Q-matrix

 
−4 2 2
Q =  2 −4 2  (17)
2 2 −4

Figures 1 to 3 give a flavor of the accuracy of the Poisson approximation in the case of
homogeneous losses at default of unit size, as a function of the number of credits in the
portfolio. The conditional default intensities were taken to be flat at 200 bps across
chain states and obligors. The three graphs on the left picture the exact distribution
function and its Poisson approximation for portfolio of increasing sizes (N = 10, 20 and
40 respectively). Graphs on the right show the absolute error one incurs when using the
Poisson approximation in the three cases above mentioned. As one would expect, the
accuracy of the approximation decreases as the size of the portfolio increases. However
in all the cases considered, the absolute error is at most 36 basis points.
Figure 4 to 6 show how the shape of the cumulative number of default density function
changes with the level of the conditional intensities λi . Again we assumed λi were
constant across states and issuers. Figure 4 refers to a relatively safe portfolio, char-
acterized by low conditional intensities. As intensities go up (see Figure 5 and 6) the
density moves to the right (inducing a higher number of expected defaults) and the
peak of the density decreases (i.e. extreme events become more likely). This is due
to the combined effect of higher individual default probabilities and change in default
correlation induced by increasing levels of λi .
So far we have not exploited the main feature of the model, i.e. the possibility of
having different conditional intensities for different states of the chain. Figure 7 shows
the density function of a portfolio of 40 names as a function of the state of the chain.
We choose λi (ξ1 ) = 500 bps, λi (ξ2 ) = 20000 bps, λi (ξ3 ) = 50000 bps for all is. Result
were computed using the Laplace transform method. Recovery rates were assumed to
be uniform (0, 1) independent random variables. Intuitively, ξ1 is the safest state and
ξ3 the riskiest. As we move from ξ1 (black line) to ξ2 (red line) the density function
shifts on the right because the individual default probabilities are now higher and
the expected loss increases. However, the tails of the distribution shrink, because the
move is accompanied by a significant drop in default correlation ρ from 3.63 to 2.85.
Note that the two effects tend to compensate each other. When moving from state ξ2
to state ξ3 (blue line), correlation remains pretty much unchanged but the portfolio
becomes more risky and the density function shifts on the right and it peak decreases
as expected.

11
References
[AW] J. Abate and W. Whitt: Numerical inversion of Laplace transforms of prob-
ability distributions, ORSA Journal on Computing 7, 36–43, 1995.
[AW1] J. Abate and W. Whitt: Numerical inversion of probability generating func-
tions, Operations Research Letters 7, No. 12, 245–251, 1992.
[BHJ] A. D. Barbour, L. Holst, S. Janson: Poisson Approximation, Claredon press,
Oxford, 1992.
[BCJR] T. R. Bielecki, S. Crepey, M. Jeanblanc, M. Rutkowski: Valuation of basket
credit derivatives in the credit migrations environment. Working paper.
[BR] T.R. Bielecki and M. Rutkowski: Credit Risk: Modelling, Valuation and
Hedging, Springer-Verlag, Berlin, 2002.
[CM] P. Carr and D. Madan: Option pricing and the fast Fourier transform, Journal
of Computational Finance 2, 61-73, 1999.
[DL] M. Davis, V. Lo: Infectious Defaults. Working paper, Imperial College, 1999.
[DGR1] G. Di Graziano, L.C.G. Rogers: Equity with Markov-modulated dividends.
Working paper, University of Cambridge, 2004.
[DGR2] G. Di Graziano, L.C.G. Rogers: Barrier options for assets with Markov mod-
ulated dividends. Working paper, University of Cambridge, 2005.
[FB] R. Frey, J. Backhaus: Portfolio credit risk models with interacting default
intensities: a Markovian approach. Working paper, University of Leipzig,
2004.
[GS] S. Galiani, Copula functions and their application in pricing and risk man-
aging multi-name credit derivatives products. Working paper, King’s College
London, 2003.
[Ho] T. Hosono: Numerical inversion of Laplace transform and some applications
to wave optics, Radio Science 16, 1015–1019.
[HK] T. R. Hurd, A. Kuznetsov: Fast CDO computations in affine Markov chains
models. Working paper, McMaster University, 2005.
[KS] I. Karatzas and S. E. Shreve: Methods of Mathematical Finance, Springer-
Verlag, New York, 1998.

12
[LG] J.P. Laurent and J. Gregory: Basket default swaps, CDOs and factor copulas.
Working paper, University of Lyon, 2003.

[L] R. W. Lee: Option pricing by transform methods: extensions, unification

and error control, Journal of Computational Finance 7, 51–86, 2004.

[Pe] A. Pelsser: Pricing double barrier options using Laplace transform, Finance
and Stochastics 7,January 2000, Vol. 4, Iss 1, 95–104.

[P] P. Protter: Stochastic Integration and Differential Equations, Springer-

Verlag, New York, 1990.

[RW] L.C.G. Rogers and D. Williams: Diffusions, Markov Processes and Martin-
gales: Volume 2 , Itô Calculus, Cambridge University Press, Cambridge,
2000.

[RZ] L.C.G. Rogers and O. Zane: Saddle-point approximations to option prices,

Annals of Applied Probability 9, 493–503, 1999.

13
Figure 1: Accuracy of the Poisson approximation, N=10, T=1, λi (ξ) = 200 bps
for all i and ξ

14
Figure 2: Accuracy of the Poisson approximation, N=20, T=1, λi (ξ) = 200 bps
for all i and ξ

15
Figure 3: Accuracy of the Poisson approximation, N=40, T=1, λi (ξ) = 200 bps
for all i and ξ

16
Figure 4: Number of defaults density, N=40, T=1, low λi

17
Figure 5: Number of defaults density, N=40, T=1, mid λi

18
Figure 6: Number of defaults density, N=40, T=1, high λi

19
Figure 7: Portfolio default density as a function of the chain’s initial state,
N=40, T=1

20