Modeling Correlated Defaults: First Passage Model under

Stochastic Volatility

Jean-Pierre Fouque

Brian C. Wignall

Xianwen Zhou

Submitted: July 3, 2006

Revised November 6, 2007
Abstract
Default dependency structure is crucial in pricing multi-name credit derivatives as well as
in credit risk management. In this paper, we extend the rst passage model for one name
with stochastic volatility (Fouque-Sircar-Slna, Applied Mathematical Finance 2006) to the
multi-name case. Correlation of defaults is generated by correlation between the Brownian
motions driving the individual names as well as through common stochastic volatility factors.
A numerical example for the loss distribution of a portfolio of defaultable bonds is examined
after stochastic volatility is incorporated.
1 Introduction
Default dependency structure is a crucial issue in pricing multi-name credit derivatives as well as
in credit risk management. For a multi-name credit derivative, the default dependency structure
among the underlying portfolio of reference entities is as important as, and in many cases even
more important than, the individual term structures of default probabilities.
On the other hand, it is well documented in the nance literature that stock returns exhibit
stochastic volatility. It has been shown in Fouque, Papanicolaou and Sircar (2000) that asymptotic
methods are very ecient in capturing the eects of stochastic volatility in simple robust corrections
to the constant volatility formulas. Recently, by incorporating stochastic volatility to the rst
passage model developed by Black and Cox (1976) in modeling defaultable bonds, Fouque, Sircar

This work is based on the third authors PhD dissertation where the uncorrelated case is treated. Work
supported by NSF grant DMS-0455982

Department of Statistics and Applied Probability, University of California, Santa Barbara, CA 93106-3110,
fouque@pstat.ucsb.edu.

Department of Statistics and Applied Probability, University of California, Santa Barbara, CA 93106-3110,
wignall@pstat.ucsb.edu.

Fixed Income Division, Lehman Brothers, NY, stephenok@hotmail.com.

1
and Slna (2006) obtained much higher yield spreads for short maturity bonds than those under
constant volatility. In fact those low yield spreads at short maturities is exactly what the original
rst passage model in fact all rms value models have been criticized for. By using models
incorporating fast and slow stochastic volatility factors, and a combination of singular and regular
perturbations techniques they obtain reasonable ts to defaultable bonds data.
In this paper, we extend the rst passage model to model default dependency in two directions:
by extending to multi-dimension and by incorporating stochastic volatility. We derive approxima-
tions for the joint survival probabilities and subsequently for the distribution of number of defaults
in a basket of names. Since we are not considering in this paper particular structured products
such as CDOs, we do not specify all the features of the defaultable bonds except that an individual
bond defaults at the rst time that the underlying rms value goes to or below some exogenously
prespecied level the default threshold. Also, since we are mainly interested in the number
of defaults before maturity, recovery does not play a role. If we were pricing derivatives, such as
tranches of CDOs, which is outside the scope of the paper, we would have to incorporate recovery.
This can be done following the remark in Fouque et al. (2006) (Section 2) by considering more
general boundary value conditions.
The rest of this paper is organized as follows. We rst set up in Section 2 the class of models that
we consider. We study rst the case where the Brownian motions driving the names are independent
so that the correlation of defaults is only due to the common factors of stochastic volatility. Then in
Section 3, using combined regular and singular perturbation techniques, as in Fouque et al. (2003)
for a single name, we obtain approximations for the joint survival probabilities. Specically, we give
a formula for the leading order term and we characterize the corrections terms as solutions of partial
dierential equations. The derivation of these equations is presented in Appendix A and explicit
formulas are given in Appendix B. In Section 4, we examine the loss distribution for a portfolio of
defaultable bonds. The general case is examined briey, and then a special case homogeneous
portfolio case is examined in detail. Following this, some numerical results illustrating the eect
of stochastic volatility are presented for the homogeneous portfolio case. Finally in Section 5 we
generalize our model to the case where the Brownian motions driving the names are correlated. We
show that an expansion around the independent case can be performed and gives a tractable way to
combine this source of default correlation with the one coming from stochastic volatility. Our result
shows that for a given maturity the dependency generated by correlating the Brownian motions
driving the names is of the same nature as the one generated by stochastic volatility. However they
dier across maturities and therefore we conclude that both sources of default correlation should
be taken into account when pricing CDOs tranches for instance.
2 Model Setup
We consider a pool of n defaultable bonds whose underlying rms value processes
_
X
(i)
t
_
n
i=1
exhibit
the following multi-factor stochastic volatility dynamics under the physical probability measure P
the real world probability measure:
dX
(1)
t
=
1
X
(1)
t
dt +f
1
(Y
t
, Z
t
)X
(1)
t
dW
(1)
t
,
dX
(2)
t
=
2
X
(2)
t
dt +f
2
(Y
t
, Z
t
)X
(2)
t
dW
(2)
t
,
2

dX
(n)
t
=
n
X
(n)
t
dt +f
n
(Y
t
, Z
t
)X
(n)
t
dW
(n)
t
,
dY
t
=
1

(m
Y
Y
t
)dt +

Y

dW
(Y )
t
,
dZ
t
= (m
Z
Z
t
)dt +
Z

2dW
(Z)
t
,
where W
(i)
t
s are standard Brownian motions and we consider rst the uncorrelated case corre-
sponding to dW
(i)
, W
(j)
)
t

ij
dt = 0 for i ,= j. The stochastic volatility correlation structure is
given by:
dW
(Y )
, W
(i)
)
t
=
iY
dt, dW
(Z)
, W
(i)
)
t
=
iZ
dt, dW
(Y )
, W
(Z)
)
t
=
Y Z
dt,
with all s being constant numbers between 1 and 1. Note that

n
i=1

2
iY
1 and

n
i=1

2
iZ
1
must be satised if the Brownian motions W
(i)
t
s are to be independent. Also
i
, , , m
Y
,
Y
, m
Z
,
Z
are all constant numbers with > 0 and > 0 both being small so that stochastic volatilities are
driven by two Ornstein-Uhlenbeck (OU) processes, Y
t
being fast mean-reverting with rate of mean-
reversion 1/ and the invariant distribution ^(m
Y
,
2
Y
), and Z
t
being slowly mean-reveting with
rate of mean-reversion and the invariant distribution ^(m
Z
,
2
Z
). The function f
i
s are positive
functions, smooth with respect to the slow variable z, and are assumed here, for instance, to be
bounded above and below away from zero. If the stochastic volatility is turned o by choosing
f
i
= 0, i = 1, . . . , n, then the defaults become independent and the default of a given rm follows
the model developed by Black and Cox (1976).
Remark. We start by assuming independence among the Brownian motions W
(i)
t
s for
i = 1, 2, . . . , n basically for two reasons. Firstly, we try to avoid the intractability caused by the
interdependence, as can be seen in Zhou (2001) where results on the joint distribution of two hitting
times are derived. This dependency will be restored in Section 5 and made tractable by using a
perturbation argument around the uncorrelated case. Secondly, we argue that the dependence
among the defaultable names introduced through stochastic volatilities is as important as the
dependence generated by the interdependence among the Brownian motions driving them. This
important issue will be discussed further in Section 5.
Under the risk-neutral probability measure

P, chosen by the market through derivatives trading,
the dynamics becomes
dX
(1)
t
= rX
(1)
t
dt +f
1
(Y
t
, Z
t
)X
(1)
t
d

W
(1)
t
,
dX
(2)
t
= rX
(2)
t
dt +f
2
(Y
t
, Z
t
)X
(2)
t
d

W
(2)
t
,

dX
(n)
t
= rX
(n)
t
dt +f
n
(Y
t
, Z
t
)X
(n)
t
d

W
(n)
t
,
dY
t
=
_
1

(m
Y
Y
t
)

Y

1
(Y
t
, Z
t
)
_
dt +

Y

d

W
(Y )
t
,
dZ
t
=
_
(m
Z
Z
t
)
Z

2
2
(Y
t
, Z
t
)
_
dt +
Z

2d

W
(Z)
t
,
where r is the riskfree interest rate (assumed constant here),

W
(i)
t
s are standard Brownian motions
with d

W
(i)
,

W
(j)
)
t
= 0 for i ,= j, and
d

W
(Y )
,

W
(i)
)
t
=
iY
dt, d

W
(Z)
,

W
(i)
)
t
=
iZ
dt, d

W
(Y )
,

W
(Z)
)
t
=
Y Z
dt.
3
The quantity of interest are the joint survival probabilities of n given rms since they are the
building blocks when one wants to compute the probability distribution of the number of defaults
as explained in Section 4.
For xed time T > 0, our objective is to nd the joint (risk-neutral) survival probability
u
,
(t, x, y, z)

P
_

(1)
t
> T, . . . ,
(n)
t
> T [ X
t
= x, Y
t
= y, Z
t
= z
_
, (1)
where t < T, X
t
(X
(1)
t
, . . . , X
(n)
t
), x (x
1
, . . . , x
n
), and
(i)
t
is the default time of rm i, dened
as follows,

(i)
t
= inf
_
s t, X
(i)
s
B
i
(s)
_
,
where B
i
(t) is the exogenously pre-specied default threshold at time t for rm i. Here we follow
Black and Cox (1976) and we assume that
B
i
(t) = K
i
e

i
t
,
with K
i
> 0 and
i
0, all being constant numbers. It is very common that credit derivatives
have long maturities. Therefore it is more realistic to assume time varying default thresholds
(exponentially growing in our case) than constant ones. Observe that u
,
is zero whenever x
i

B
i
(t) for some i and therefore we only need to focus on the case where x
i
> B
i
(t) for all i =
1, 2, . . . , n.
3 Approximated Joint Survival Probabilities
We derive in this section an approximation for the joint survival probability u
,
dened in (1).
We rst write a PDE representation for it and we then perform singular and regular perturbations
with respect to the small parameters and .
3.1 PDE Representation
In terms of partial dierential equations (PDE), u
,
is the solution to the following boundary value
problem:
L
,
u
,
= 0, x
i
> B
i
(t), for all i, t < T,
u
,
(t, x
1
, x
2
, . . . , x
n
, y, z) = 0, i 1, , n, x
i
= B
i
(t), t T,
u
,
(T, x
1
, x
2
, . . . , x
n
, y, z) = 1, x
i
> B
i
(t), for all i,
where the operator L
,
has the following decomposition in terms of powers of

and

:
L
,
=
1

L
0
+
1

L
1
+L
2
+

/
1
+/
2
+
_

/
3
, (2)
4
with the notations
L
0
= (m
Y
y)

y
+
2
Y

2
y
2
,
L
1
=
Y

2
_
n

i=1

iY
f
i
(y, z)x
i

2
x
i
y

1
(y, z)

y
_
,
L
2
=

t
+
n

i=1
_
1
2
f
2
i
(y, z)x
2
i

2
x
2
i
+rx
i

x
i
_
, (3)
/
1
=
Z

2
_
n

i=1

iZ
f
i
(y, z)x
i

2
x
i
z

2
(y, z)

z
_
,
/
2
= (m
Z
z)

z
+
2
Z

2
z
2
,
/
3
= 2
Y Z

2
yz
.
As in Fouque et al. (2003), we expand u
,
in terms of powers of

and

:
u
,
= u
0
+

u
1,0
+

u
0,1
+ u
2,0
+

u
1,1
+ u
0,2
+ , (4)
and retain
u u
0
+

u
1,0
+

u
0,1
(5)
as our approximation for u
,
. In the Appendix A we present the formal expansion argument leading
to the characterization of the leading order term u
0
and the two correction terms

u
1,0
and

u
0,1
.
In the following sections we compute these terms in the context of our multidimensional boundary
value problem.
3.2 Leading Order Term u
0
Following Fouque et al. (2003) as explained in the Appendix A, the leading order term u
0
is
independent of y and is determined by the following PDE system with respect to the variables
(t, x), the variable z being simply a parameter:
L
2
)u
0
= 0, x
i
> B
i
(t), for all i, t < T, (6)
u
0
(t, x
1
, x
2
, . . . , x
n
) = 0, i 1, , n, x
i
= B
i
(t), t T,
u
0
(T, x
1
, x
2
, . . . , x
n
) = 1, x
i
> B
i
(t), for all i,
where, from the denition (3) of L
2
, we have
L
2
) =

t
+
n

i=1
_
1
2
f
2
i
(, z))x
2
i

2
x
2
i
+rx
i

x
i
_
. (7)
5
The variable z appears only as a parameter in the averaged diusion coecients f
2
i
(, z)) with
respect to the invariant distribution ^(m
Y
,
2
Y
) of the process Y
t
under the real world measure P,
i.e.,
f
2
i
(, z)) =
_
f
2
i
(y, z)
1

2
exp
_

(y m
Y
)
2
2
2
Y
_
dy.
Proposition 1 The leading order term u
0
in the approximation (5) is given by:
u
0
=
n

i=1
Q
i

n

i=1
_
N
_
d
+
2(i)
_

_
x
i
B
i
(t)
_
p
i
N
_
d

2(i)
_
_
, (8)
where N() is the standard cumulative normal distribution function, and
d

2(i)

ln
x
i
B
i
(t)
+
_
r
i

2
i
(z)
2
_
(T t)

i
(z)

T t
,

i
(z)
_
f
2
i
(, z)), (eective volatility of rm i)
p
i
1
2(r
i
)

2
i
(z)
.
Proof From (6) and (7), u
0
admits the probabilistic representation:
u
0
=

E
_
n

i=1
1
n
inf
tsT

X
(i)
s
/B
i
(s) >1
o
[

X
s
= x
_
,
where, under the probability measure

P, we have
d

X
(1)
t
= r

X
(1)
t
dt +
1
(z)

X
(1)
t
d

W
(1)
t
,
d

X
(2)
t
= r

X
(2)
t
dt +
2
(z)

X
(2)
t
d

W
(2)
t
,

d

X
(n)
t
= r

X
(n)
t
dt +
n
(z)

X
(n)
t
d

W
(n)
t
,
with

W
(i)
s being independent standard Brownian motions. In other words,
_

X
(i)
t
_
, i = 1, 2, . . . , n,
are independent Geometric Brownian motions and it follows that
u
0
=
n

i=1

P
_
inf
tsT

X
(i)
s
/B
i
(s) > 1 [

X
(i)
t
= x
i
_
.
By doing change of variables to make the corresponding boundary constant (instead of exponentially
growing) and then using the appropriate formula in Borodin and Salminen (2002) obtained by the
method of images or equivalently the reection principle in one dimension, one can reach the
conclusion in the proposition.
6
3.3 Correction Term

u
1,0
From Appendix A the term u
1,0
is determined by the following PDE system:
L
2
)u
1,0
= /u
0
, x
i
> B
i
(t), for all i, t < T, (9)
u
1,0
(t, x
1
, x
2
, . . . , x
n
) = 0, i 1, , n, x
i
= B
i
(t), t T,
u
1,0
(T, x
1
, x
2
, . . . , x
n
) = 0, x
i
> B
i
(t), for all i,
where the operator / is given by
/ L
1
L
1
0
(L
2
L
2
)))
=

Y

2
_
_
n

i=1
n

j=1

iY
_
f
i

j
y
_
x
i

x
i
_
x
2
j

2
x
2
j
_

j=1
_

j
y
_
x
2
j

2
x
2
j
_
_
=

2
n

i=1
_

i
y
_
x
2
i

2
x
2
i
+

Y

2
n

i=1

iY
_
f
i

i
y
_
x
i

x
i
_
x
2
i

2
x
2
i
_
+

2
n

i,j=1
i=j

iY
_
f
i

j
y
_
x
i

x
i
_
x
2
j

2
x
2
j
_
. (10)
Here
i
(y, z), for i = 1, 2, . . . , n, denote solutions of reasonable growth at innity to the following
Poisson equations with respect to the variable y:
L
0

i
(y, z) = f
2
i
(y, z) f
2
i
(, z)), (11)
and ) denotes the average with respect to the invariant distribution ^(m
Y
,
2
Y
) of Y .
The following result shows that the multi-dimensional problem (9) can be reduced to many
uncoupled one- and two-dimensional problems.
Proposition 2 The correction term

u
1,0
is given by

u
1,0
=
n

i=1
R
(2)
i
w
(2)
i
n

j=1
j=i
Q
j
+
n

i=1
R
(3)
i
w
(3)
i
n

j=1
j=i
Q
j
+
n

i,j=1
i=j
R
(3)
ij
w
(3)
ij
n

k=1
k=i,j
Q
k
, (12)
where the coecients R
(2)
i
, R
(3)
i
, R
(3)
ij
depend on the parameter z and are given by
R
(2)
i
=

2
_

1
(, z)

i
y
(, z)
_
(13)
R
(3)
i
=

Y

iY
_
f
i
(, z)

i
y
(, z)
_
(14)
R
(3)
ij
=

Y

iY
_
f
i
(, z)

j
y
(, z)
_
, i ,= j, (15)
7
with
i
s given by (11) and Q
i
s given in Proposition 1, and where the functions w
(2)
i
(t, x
i
), w
(3)
i
(t, x
i
)
and w
(3)
ij
(t, x
i
, x
j
) depend on the parameter z and are given by the following problems:
_

t
+
1
2

2
i
(z)x
2
i

2
x
2
i
+rx
i

x
i
_
w
(2)
i
= x
2
i

2
Q
i
x
2
i
, x
i
> B
i
(t), t < T, (16)
w
(2)
i
(t, B
i
(t)) = 0, t T,
w
(2)
i
(T, x
i
) = 0, x
i
> B
i
(t),
_

t
+
1
2

2
i
(z)x
2
i

2
x
2
i
+rx
i

x
i
_
w
(3)
i
= x
i

x
i
_
x
2
i

2
Q
i
x
2
i
_
, x
i
> B
i
(t), t < T, (17)
w
(3)
i
(t, B
i
(t)) = 0, t T,
w
(3)
i
(T, x
i
) = 0, x
i
> B
i
(t),
_

t
+
1
2

2
i
(z)x
2
i

2
x
2
i
+
1
2

2
j
(z)x
2
j

2
x
2
j
+rx
i

x
i
+rx
j

x
j
_
w
(3)
ij
=
_
x
i
Q
i
x
i
_
_
x
2
j

2
Q
j
x
2
j
_
, (18)
x
i
> B
i
(t), x
j
> B
j
(t), t < T,
w
(3)
ij
(t, x
i
, x
j
) = 0, if x
i
= B
i
(t) or x
j
= B
j
(t), t T,
w
(3)
ij
(T, x
i
, x
j
) = 0, x
i
> B
i
(t), x
j
> B
j
(t),
with
i
(z) given in Proposition 1.
Proof Note that with the form (10) of / and the denitions (13, 14, 15) we have:

/ =
n

i=1
R
(2)
i
x
2
i

2
x
2
i
+
n

i=1
R
(3)
i
x
i

x
i
_
x
2
i

2
x
2
i
_
+
n

i,j=1
i=j
R
(3)
ij
x
i

x
i
_
x
2
j

2
x
2
j
_
.
By linearity of (9), it is enough to check that
L
2
)
_
_
w
(2)
i
n

j=1,j=i
Q
j
_
_
= x
2
i

2
u
0
x
2
i
, (19)
L
2
)
_
_
w
(3)
i
n

j=1,j=i
Q
j
_
_
= x
i

x
i
_
x
2
i

2
u
0
x
2
i
_
, (20)
L
2
)
_
_
w
(3)
ij
n

k=1,k=i,j
Q
k
_
_
= x
i

x
i
_
x
2
j

2
u
0
x
2
j
_
. (21)
Using the form (7) of L
2
) and u
0
=

n
i=1
Q
i
one can easily check that (19), (20), and (21) are
satised. The boundary and terminal conditions for the correction

u
1,0
are directly inherited
from the boundary and terminal conditions for the functions w
(2)
i
s, w
(3)
i
s, w
(3)
ij
s, and Q
i
s.
The problems (16), (17) and (18) are boundary value problems with sources. In Appendix B
we show how to transform them into boundary value problems without source leading to explicit
formulas (up to Gaussian integrals).
8
3.4 Correction Term

u
0,1
From Appendix A the term u
0,1
is determined by the following PDE system:
L
2
)u
0,1
= /
1
)u
0
, x
i
> B
i
(t), for all i, t < T, (22)
u
0,1
(t, x
1
, x
2
, . . . , x
n
) = 0, i 1, , n, x
i
= B
i
(t), t T,
u
0,1
(T, x
1
, x
2
, . . . , x
n
) = 0, x
i
> B
i
(t), for all i.
where the operator /
1
) is given by:
/
1
) =
Z

2
_
n

i=1

iZ
f
i
(, z))x
i

2
x
i
z

2
(, z))

z
_
.
The following result shows that, as for u
1,0
, the multi-dimensional problem (22) for u
0,1
can be
reduced to many uncoupled one- and two-dimensional problems.
Proposition 3 The correction term

u
0,1
is given by

u
0,1
=
n

i=1
R
(0)
i
w
(0)
i
n

j=1
j=i
Q
j
+
n

i=1
R
(1)
i
w
(1)
i
n

j=1
j=i
Q
j
+
n

i,j=1
i=j
R
(1)
ij
w
(1)
ij
n

k=1
k=i,j
Q
k
, (23)
where the coecients R
(0)
i
, R
(1)
i
, R
(1)
ij
depend on the parameter z and are given by
R
(0)
i
=
Z

2
2
(, z))

i
(z) (24)
R
(1)
i
=
Z

2
iZ
f
i
(, z))

i
(z) (25)
R
(1)
ij
=
Z

2
iZ
f
i
(, z))

j
(z), i ,= j, (26)

i
(z)s and Q
i
s given in Proposition 1,

i
= d
i
/dz, and where the functions w
(0)
i
(t, x
i
; z), w
(1)
i
(t, x
i
; z)
and w
(1)
ij
(t, x
i
, x
j
; z) depend on the parameter z and are given by the following problems:
_

t
+
1
2

2
i
(z)x
2
i

2
x
2
i
+rx
i

x
i
_
w
(0)
i
=
Q
i

i
, x
i
> B
i
(t), t < T, (27)
w
(0)
i
(t, B
i
(t)) = 0, t T,
w
(0)
i
(T, x
i
) = 0, x
i
> B
i
(t),
_

t
+
1
2

2
i
(z)x
2
i

2
x
2
i
+rx
i

x
i
_
w
(1)
i
= x
i

x
i
_
Q
i

i
_
, x
i
> B
i
(t), t < T, (28)
w
(1)
i
(t, B
i
(t)) = 0, t T,
w
(1)
i
(T, x
i
) = 0, x
i
> B
i
(t),
9
_

t
+
1
2

2
i
(z)x
2
i

2
x
2
i
+
1
2

2
j
(z)x
2
j

2
x
2
j
+rx
i

x
i
+rx
j

x
j
_
w
(1)
ij
=
_
x
i
Q
i
x
i
__
Q
j

j
_
, (29)
x
i
> B
i
(t), x
j
> B
j
(t), t < T,
w
(1)
ij
(t, x
i
, x
j
) = 0, if x
i
= B
i
(t) or x
j
= B
j
(t), t T,
w
(1)
ij
(T, x
i
, x
j
) = 0, x
i
> B
i
(t), x
j
> B
j
(t).
Proof The proof is very similar to that of Proposition 2. Since u
0
depends on z only through

i
(z)s, we have
u
0
z
=
n

j=1

j
(z)
u
0

j
,
and therefore

/
1
)u
0
=
Z

2
_
_
n

i=1

iZ
f
i
(, z))x
i

x
i
_
_
n

j=1

j
(z)
u
0

j
_
_

2
(, z))
n

j=1

j
(z)
u
0

j
_
_
=
Z

2
_
_
n

i=1
n

j=1

iZ
f
i
(, z))

j
(z)x
i

x
i
_
u
0

j
_

2
(, z))
n

i=1

i
(z)
u
0

i
_
_
=
n

i=1
R
(0)
i
u
0

i
+
n

i=1
R
(1)
i
x
i

x
i
_
u
0

i
_
+
n

i,j=1
i=j
R
(1)
ij
x
i

x
i
_
u
0

j
_
,
where we have used the denitions (24, 25, 26) of (R
(0)
i
, R
(1)
i
, R
(1)
ij
). By linearity of (22), it is enough
to check that
L
2
)
_
_
w
(1)
i
n

j=1,j=i
Q
j
_
_
=
u
0

i
, (30)
L
2
)
_
_
w
(1)
i
n

j=1,j=i
Q
j
_
_
= x
i

x
i
_
u
0

i
_
, (31)
L
2
)
_
_
w
(1)
ij
n

k=1,k=i,j
Q
k
_
_
= x
i

x
i
_
u
0

j
_
. (32)
Using the form (7) of L
2
) and u
0
=

n
i=1
Q
i
one can easily check that (30), (31), and (32) are
satised. The boundary and terminal conditions for the correction

u
0,1
are directly inherited
from the boundary and terminal conditions for the functions w
(0)
i
s, w
(1)
i
s, w
(1)
ij
s, and Q
i
s.
As for (16), (17) and (18) we show in Appendix B that the boundary value problems with sources
(27), (28), and (29) for w
(0)
i
s, w
(1)
i
s, w
(1)
ij
s can be transformed into boundary value problems
without source leading to explicit formulas (up to Gaussian integrals).
10
3.5 Summary of the Approximation
Combining the results of Propositions 1, 2 and 3 we get that the approximation u in (5) is given
by
u =
n

i=1
Q
i
+
n

i=1
R
(2)
i
w
(2)
i
n

j=1
j=i
Q
j
+
n

i=1
R
(3)
i
w
(3)
i
n

j=1
j=i
Q
j
+
n

i,j=1
i=j
R
(3)
ij
w
(3)
ij
n

k=1
k=i,j
Q
k
+
n

i=1
R
(0)
i
w
(0)
i
n

j=1
j=i
Q
j
+
n

i=1
R
(1)
i
w
(1)
i
n

j=1
j=i
Q
j
+
n

i,j=1
i=j
R
(1)
ij
w
(1)
ij
n

k=1
k=i,j
Q
k
, (33)
where (R
(2)
i
, R
(3)
i
, R
(3)
ij
) are small of order

, (R
(0)
i
, R
(1)
i
, R
(1)
ij
) are small of order

, and they all

depend on the parameter z. The functions Q
i
, w
(2)
i
, w
(3)
i
, w
(0)
i
, w
(1)
i
depend on the variable x
i
, the
functions w
(3)
ij
, w
(1)
ij
depend on the variables (x
i
, x
j
), and they all depend on the parameter z.
The accuracy of approximation (33) is given at the end of Appendix A.
Remark The (approximated) survival probability of a single bond/rm (treated in Fouque et
al. (2006)) is a particular case of the result (33). It can be obtained by xing an index and
eliminating all those terms that involve any one of the other indices, which includes all the two-
index cross-terms. For example, the (approximated) survival probability of bond 1 only involves
Q
1
, w
(0)
1
, w
(1)
1
, w
(2)
1
and w
(3)
1
.
3.6 Numerical Illustration of the Accuracy of Approximation
In order to illustrate the quality of the approximation of the joint survival probability given by
(33) we have conducted the following numerical experiments. For n = 10 names (Table 1), and
for n = 25 names (Table 2), we compute the zero-order approximation u
0
given by (8), and the
rst order approximation u given by (33) with the explicit formulas derived in Appendix B. We
present the results for four sets of values of the small parameters and . Since there is no explicit
formulas for the true value, we obtain it by Monte Carlo simulations with a very large number of
realizations (10
5
) and using an Euler scheme with a very small time-step (10
4
) in order to ensure
accuracy of the true value proxy denoted by u
MC
. The absolute and relative errors are shown in
the last columns. In all cases we have used the following parameter values:
X
(i)
0
= 20, r = 5%,
i
= 6%, K
i
= 10,

ij
= 0,
iY
=
iZ
= 1/(2

n),
Y Z
= 0,
m
Y
= m
Z
= 30%,
Y
=
Z
= 10%, Y
0
= Z
0
= 30%,
1
=
2
= 0,
f
i
(y, z) = 30%exp(y +z)/ exp(m
Y
+m
Z
+
2
Y
+
2
Z
), T = 1.
As expected the rst order approximation u converges to the (simulated) true value u
MC
as
(, ) goes to (0, 0). In fact, as often observed in homogenization, the approximation remains very
11
u
0
u u
MC
Absolute (relative) error
1/100 1/50 0.740389 0.75079 0.7502 0.0006 (0.08%)
1/50 1/20 0.740389 0.756015 0.7529 0.003 (0.4%)
1/20 1/10 0.740389 0.763647 0.7567 0.007 (0.9%)
1 1 0.740389 0.82833 .7653 0.063 (8.2%)
Table 1: Joint survival probability for ten rms (n = 10).
u
0
u u
MC
Absolute (relative) error
1/100 1/50 0.471683 0.481506 0.4789 0.003 (0.5%)
1/50 1/20 0.471683 0.486892 0.4803 0.006 (1.4%)
1/20 1/20 0.471683 0.488478 0.4854 0.003 (0.6%)
1/20 1/10 0.471683 0.493648 0.4843 0.009 (1.9%)
Table 2: Joint survival probability for twenty ve rms (n = 25).
accurate even in regimes where these parameters are not so small. We also observe that, in the
present case with
ij
= 0, the volatility-name correlations
iY
,
iZ
are small (since

n
i=1

2
iY
1
and

n
i=1

2
iZ
1), and the eect of the correction due to stochastic volatility is also relatively
small (u
0
is already close to u
MC
). This will not be the case with name-name correlations
ij
,= 0
as shown in Section 5.
4 Loss Distribution
We consider now a portfolio consisting of N defaultable bonds and we denote by a
i
the number of
bond i in this portfolio. Assuming zero recovery rate from default for each bond, the loss at time
t of this portfolio is given by the random variable
L(t)
N

i=1
a
i

i
(t),
where
i
(t) takes on 1 if bond i defaults before t and 0 otherwise. Our objective is to study the
distribution of L(T) for a maturity T smaller than all the bond maturities.
4.1 General Case
Since the portfolio consists only of a nite number of bonds, L must have a discrete distribution
function. It hence suces to nd the probability value that a subset of names in the portfolio
default. It turns out that this can be done recursively. For example,
P

only bond 1 defaults in the portfolio before time T

= P

1
T,
2
> T, ,
N
> T
= P

2
> T, ,
N
> T P

1
> T,
2
> T, ,
N
> T,
12
while
i
is the default time of bond i. Both terms on the right hand side of the last equality are
known (approximately) by formula (33). In general, if we denote
I i
1
, i
2
, . . . , i
n
J j
1
, j
2
, . . . , j
m


N 1, 2, . . . , N,
D
|I|
|J|
(I; J) P

(bond i defaults, i I) (bond j survives, j JI),

then
D
n
m
(i
1
, . . . , i
n
; J) = P

(
i
T, i i
1
, . . . , i
n1
, i
n
) (
j
> T, j JI)
= P

(
i
T, i i
1
, . . . , i
n1
) (
j
> T, j JI)
P

(
i
T, i i
1
, . . . , i
n1
) (
j
> T, j Ji
1
, . . . , i
n1
) .
Therefore,
D
n
m
(i
1
, . . . , i
n
; j
1
, . . . , j
m
) = D
n1
m1
(i
1
, . . . , i
n1
; j
1
, . . . , j
m
i
n
)
D
n1
m
(i
1
, . . . , i
n1
; j
1
, . . . , j
m
) . (34)
Formula (34) is recursive and by implementing it once, one can reduce by one the superscript n
(= the number of defaults) on D. This can be done repeatedly until one reduces the superscript
to zero, implying a total survival of an appropriate subset of names in the portfolio. Finally the
probability of total survival is (approximately) u given by (33), with names/indices 1, 2, . . . , n
replaced by the appropriate set of indices/names.
It can be shown that in order to compute D
n
m
(, ), one has to evaluate u-like forms 2
n
times.
This may be computationally expensive for a portfolio of big size, say 50 names or more. For a
smaller size portfolio, say N = 20, however, the computational cost is acceptable to nd for instance
the probability of 50% loss, namely D
10
20
(, ). To be more precise, the computation of D
10
20
(, ) would
require to compute 2
10
= 1024 of u-like forms given by (33). However, every item in equation (33)
is in closed-form (or up to double integrals), and therefore can be computed fairly fast when an
appropriate programming language is chosen (say, C/C++). This can be done within 0.05 to 0.5
seconds on a 2GHz CPU with 2GB RAM PC. Therefore the whole computation of D
10
20
(, ) can be
done within 51.2 to 512 seconds, which is what we mean by acceptable.
4.2 Special Case: Homogeneous Portfolio
We now consider a fully homogeneous portfolio, that is
f
i
(y, z) = f(y, z),
iY
=
Y
,
iZ
=
Z
, a
i
= a, X
(i)
0
= x
i
= x, (35)
for all i = 1, , N. We rst recall a classical result.
Lemma 1 Let
i
be the default time of bond i, and suppose that under a probability P
P(
i
> T, i i
1
, i
2
, . . . , i
m
) = P(
j
> T, j j
1
, j
2
, . . . , j
m
),
for any two equal-size sets of indices i
1
, i
2
, . . . , i
m
and j
1
, j
2
, . . . , j
m
chosen from 1, 2, . . . , N
with 1 m N. Dene
S
m
P(
i
> T, i 1, 2, . . . , m), 1 m N. (36)
13
Then, for 0 k < N, the probability that exactly k bonds, among the N bonds, default before T is
given by
F
(N)
k
P
_
N

i=1

i
(T) = k
_
=
_
N
k
_
k

j=0
_
k
j
_
(1)
j
S
N+jk
. (37)
Proof We provide here a short proof which consists in computing the moment generating function
of the number of survivals. Denote by

i
1
i
the event rm i survives after T.
On one hand we have
(z) E
_
z
(
P
N
i=1

i
)
_
= E
_
z
(N
P
N
i=1

i
)
_
=
N

k=0
z
(Nk)
F
(N)
k
=
N

k=0
z
k
F
(N)
Nk
. (38)
On the other hand we have
(z) = E
_
N

i=1
z

i
_
= E
_
N

i=1
_
1 + (z 1)

i
_
_
= E
_
_

(i
1
,,in),n=0,,N
(z 1)
n

i
1

in
_
_
=
N

n=0
(z 1)
n

(i
1
,,in)
P(
i
> T, i i
1
, i
2
, . . . , i
n
) =
N

n=0
(z 1)
n
_
N
n
_
S
n
=
N

n=0
_
n

k=0
(1)
nk
_
n
k
_
z
k
_
_
N
n
_
S
n
=
N

k=0
z
k
_
N

n=k
(1)
nk
_
n
k
__
N
n
_
S
n
_
. (39)
The function (z) being the polynomial in z given by (38) and (39) we deduce that
F
(N)
Nk
=
N

n=k
(1)
nk
_
n
k
__
N
n
_
S
n
=
_
N
N k
_
N

n=k
(1)
nk
_
N k
n k
_
S
n
.
We nally obtain (37):
F
(N)
k
=
_
N
k
_
N

n=Nk
(1)
nN+k
_
k
n N +k
_
S
n
=
_
N
k
_
k

j=0
(1)
j
_
k
j
_
S
Nk+j
,
by setting j = n N +k.
With the homogeneity assumption, we set q Q
1
(t, x) for the survival probability after t of
one given bond, and we obtain from the results in Section 3 that
u
0
=
n

i=1
Q
i
= q
n
,
14

u
1,0
=
n

i=1
R
(2)
i
w
(2)
i
n

j=1
j=i
Q
j
+
n

i=1
R
(3)
i
w
(3)
i
n

j=1
j=i
Q
j
+
n

i,j=1
i=j
R
(3)
ij
w
(3)
ij
n

k=1
k=i,j
Q
k
= nR
(2)
1
w
(2)
1
(t, x)q
n1
+nR
(3)
1
w
1
(t, x)
(3)
q
n1
+n(n 1)R
(3)
12
w
(3)
12
(t, x, x)q
n2
,

u
0,1
=
n

i=1
R
(0)
i
w
(0)
i
n

j=1
j=i
Q
j
+
n

i=1
R
(1)
i
w
(1)
i
n

j=1
j=i
Q
j
+
n

i,j=1
i=j
R
(1)
ij
w
(1)
ij
n

k=1
k=i,j
Q
k
= nR
(0)
1
w
(0)
1
(t, x)q
n1
+nR
(1)
1
w
(1)
1
(t, x)q
n1
+n(n 1)R
(1)
12
w
(1)
12
(t, x, x)q
n2
,
where we have used the fact that the Rs and ws do not depend on a particular choice of names
(i, j).
We dene the quantities
A
3

k=0
R
(k)
1
w
(k)
1
(t, x), (40)
B R
(1)
12
w
(1)
12
(t, x, x) +R
(3)
12
w
(3)
12
(t, x, x), (41)
which also depend on the parameter z, and we rewrite the joint survival probabilities (36) as
S
n
u u
0
+

u
1,0
+

u
0,1
= q
n
+Anq
n1
+Bn(n 1)q
n2
, n 2.
Note that S
1
= q +A, and hence the previous formula for S
n
is actually valid for all n 1, as well
as for n = 0 with S
0
1.
The approximated loss distribution is now given by
P

(L = k) =
_
N
k
_
k

j=0
_
k
j
_
(1)
j
S
N+jk
=
_
N
k
_
k

i=0
_
k
i
_
(1)
ki
S
Ni

_
N
k
_
k

i=0
_
k
i
_
(1)
ki
_
q
Ni
+A(N i)q
Ni1
+B(N i)(N i 1)q
Ni2

=
_
N
k
_
k

i=0
_
k
i
_
(1)
ki
q
Ni
+A
_
N
k
_
k

i=0
_
k
i
_
(1)
ki
(N i)q
Ni1
+B
_
N
k
_
k

i=0
_
k
i
_
(1)
ki
(N i)(N i 1)q
Ni2
I
0
+AI
1
+BI
2
,
where I
0
, I
1
, and I
2
are obtained by straightforward calculation:
I
0
=
_
N
k
_
k

i=0
_
k
i
_
(1)
ki
q
Ni
=
_
N
k
_
(1 q)
k
q
Nk
, (42)
I
1
=
_
N
k
_
_
k

i=0
_
k
i
_
(1)
ki
s
Ni
_

s=q
=
_
N
k
_
_
s
Nk
(1 s)
k
_

s=q
=
_
N
k
_
_
(N k)q
Nk1
(1 q)
k
kq
Nk
(1 q)
k1
_
15
=
_
N k
q

k
1 q
_
I
0
, (43)
I
2
=
_
N
k
_
_
k

i=0
_
k
i
_
(1)
ki
s
Ni
_

s=q
=
_
N
k
_
_
s
Nk
(1 s)
k
_

s=q
=
_
N
k
_
_
(N k)(N k 1)q
Nk2
(1 q)
k
2k(N k)q
Nk1
(1 q)
k1
+k(k 1)q
Nk
(1 q)
k2
_
=
_
(N k)(N k 1)
q
2

2k(N k)
q(1 q)
+
k(k 1)
(1 q)
2
_
I
0
. (44)
To summarize, for 0 k < N,
P

(L = k) I
0
+AI
1
+BI
2
, (45)
where I
0
, I
1
and I
2
are explicitly given by (42, 43, 44), and A and B dened in (40, 41) are small
of order max

.
Note that I
0
corresponds to the case where all the underlying assets are mutually indepen-
dent, which gives rise to the classic binomial distribution. Since the method used in this paper is
perturbation, we call (45) the perturbed binomial formula.
4.3 Numerical Illustration
In the homogeneous-portfolio case discussed above, we implemented some numerical computation
to illustrate the eect of introducing stochastic volatilities. We compute the approximated loss
distribution given by (45) where only four parameters (N, q, A, B) are needed. The purpose of this
computation is to show stylized features of the loss distribution generated by the correlation of
defaults due to the presence of stochastic volatility.
The parameters used in Figure 1 are as follows:
N = 100, q = 0.9, A = 0.00, B = 0.0006.
The upper graph presents the probability mass function and the lower one presents the cumulative
distribution function for the loss. Note that in order to make comparison between our result and
the classic binomial distribution, we chose A = 0 because a non-zero A would make single name
survival probabilities distinct under these two scenarios. By choosing A = 0, the single name
survival probability is q = 0.9 = 90%. Note also that B is chosen very small to ensure that the
approximation is in its range of validity.
It can be observed from Figure 1 that:
the probability, with stochastic volatilities, that the total portfolio loss is less than 3% is
bigger than that from the classic binomial distribution;
16
0 10 20 30 40 50 60 70 80 90 100
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
number of defaults
p
r
o
b
a
b
ilit
y
binomial
perturbed
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
number of defaults
c
u
m
u
la
t
iv
e

p
r
o
b
a
b
ilit
y
binomial
perturbed
Figure 1: Perturbed Binomial Loss Distribution
the probability, with stochastic volatilities, that the total portfolio loss is bigger than 15% is
bigger than that from the classic binomial distribution.
In terms of CDO tranches (we refer to Due and Singleton (2003) for an introduction), this
implies that the expected loss and hence the fair spread of the equity tranche become less, while
the expected loss and hence the fair spread of the senior tranche become bigger, than in the classic
binomial case. This phenomenon is consistent with what has been documented so far in CDO
literature if one increases the default correlation between the underlying bonds and keeps other
things unchanged. For example, see Due and Garleanu (2001). Note that a positive number
B indicates positive correlation between the underlying bonds. Here we assume that the equity
tranche absorbs 0-3% loss and the senior absorbs 15%-100% loss.
Furthermore, the binomial probability mass function achieves its maximum 0.1319 at k = 11,
while the probability mass function with stochastic volatilities achieves its maximum 0.1047 at
k = 8.
5 Models with Name-Name Correlation
We now consider the correlated case, namely where the Brownian motions

W
(i)
s driving the names
are correlated. Using the notation of Section 2, this means taking d

W
(i)
,

W
(j)
)
t
=
ij
dt for i ,= j
17
under the risk-neutral probability measure

P with [
ij
[ < 1. Observe that for non-zero
ij
s the
conditions

n
i=1

2
iY
1 and

n
i=1

2
iZ
1 are no longer needed. We denote by u
,,
the joint
survival probability dened by
u
,,
(t, x, y, z)

P
_

(1)
t
> T, . . . ,
(n)
t
> T [ X
t
= x, Y
t
= y, Z
t
= z
_
. (46)
In terms of PDEs, u
,,
is the solution to the boundary value problem
L
,,
u
,,
= 0, x
i
> B
i
(t), for all i, t < T,
u
,,
(t, x
1
, x
2
, . . . , x
n
, y, z) = 0, i 1, , n, x
i
= B
i
(t), t T,
u
,,
(T, x
1
, x
2
, . . . , x
n
, y, z) = 1, x
i
> B
i
(t), for all i,
where the operator L
,,
can be written as
L
,,
= L
,
+
n

i<j

ij
L
(ij)

, (47)
where we have used
ij
=
ji
, the denition (2) of the operator L
,
, and the notation
L
(ij)

= f
i
(y, z)f
j
(y, z)x
i
x
j

2
x
i
x
j
. (48)
The leading order term in the small and small expansion carried out in the previous sections
would correspond to the correlated multi-name case with constant volatility. As explained in the
introduction this is not a tractable model because of the lack of simple formulas for the joint distri-
bution of hitting times. This leads us to perform an additional expansion around the independent
case where the
ij
s are zero. We therefore consider the case where the
ij
s are small and of the
same order.
Expanding u
,,
in powers of

,

, and
ij
we get:
u
,,
= u
,
+
n

i<j

ij
_
u
(ij)
0,0,1
+

u
(ij)
1,0,1
+

u
(ij)
0,1,1
+
_
+ , (49)
and retain the lowest order terms,
u u
0
+

u
1,0
+

u
0,1
+
n

i<j

ij
u
(ij)
0,0,1
, (50)
as our approximation for u
,,
where the rst three terms have been computed in Section 3.
5.1 Correction Terms
ij
u
(ij)
0,0,1
From Appendix A the term u
(ij)
0,0,1
is determined by the following PDE system:
L
2
)u
(ij)
0,0,1
= L
(ij)

)u
0
, x
l
> B
l
(t), for all l, t < T, (51)
u
(ij)
0,0,1
(t, x
1
, x
2
, . . . , x
n
) = 0, l 1, , n, x
l
= B
l
(t), t T,
u
(ij)
0,0,1
(T, x
1
, x
2
, . . . , x
n
) = 0, x
l
> B
l
(t), for all l,
18
where the operator L
(ij)

) is given by
L
(ij)

) = f
i
(, z)f
j
(, z))x
i
x
j

2
x
i
x
j
.
The following result shows that each u
(ij)
0,0,1
corresponds to a single two-dimensional problem.
Proposition 4 The correction term
ij
u
(ij)
0,0,1
is given by

ij
u
(ij)
0,0,1
= R
(4)
ij
w
(4)
ij
n

k=1
k=i,j
Q
k
, (52)
where the coecient R
(4)
ij
depends on the parameter z and is given by
R
(4)
ij
=
ij
f
i
(, z)f
j
(, z)), i ,= j, (53)
and where the function w
(4)
ij
(t, x
i
, x
j
) depends on the parameter z and is given by the following
problem:
_

t
+
1
2

2
i
(z)x
2
i

2
x
2
i
+
1
2

2
j
(z)x
2
j

2
x
2
j
+rx
i

x
i
+rx
j

x
j
_
w
(4)
ij
=
_
x
i
Q
i
x
i
__
x
j
Q
j
x
j
_
, (54)
x
i
> B
i
(t), x
j
> B
j
(t), t < T,
w
(4)
ij
(t, x
i
, x
j
) = 0, if x
i
= B
i
(t) or x
j
= B
j
(t), t T,
w
(4)
ij
(T, x
i
, x
j
) = 0, x
i
> B
i
(t), x
j
> B
j
(t),
with
i
(z) given in Proposition 1.
Proof The proof is very similar to that of Proposition 2. With the denition (53) of R
(4)
ij
, we
have

ij
L
(ij)

)u
0
= R
(4)
ij
x
i

x
i
_
x
j
u
0
x
j
_
. (55)
It is therefore enough to check that
L
2
)
_
_
w
(4)
ij
n

k=1,k=i,j
Q
k
_
_
= x
i

x
i
_
x
j
u
0
x
j
_
. (56)
Using the form (7) of L
2
) and u
0
=

n
i=1
Q
i
one can easily check that (56) is satised.
The boundary and terminal conditions for the correction
ij
u
(ij)
0,0,1
are directly inherited from the
boundary and terminal conditions for the functions w
(4)
ij
s, and Q
i
s.
As for (18) we show in Appendix B that the boundary value problem with source (54) for w
(4)
ij
can be transformed into boundary value problems without source leading to explicit formulas (up
to Gaussian integrals).
19
5.2 Summary of Approximation with Name-Name Correlation
Combining the results of Propositions (1)(4), we get that the approximation u in (50) is given by
u =
n

i=1
Q
i
+
n

i=1
R
(2)
i
w
(2)
i
n

j=1
j=i
Q
j
+
n

i=1
R
(3)
i
w
(3)
i
n

j=1
j=i
Q
j
+
n

i,j=1
i=j
R
(3)
ij
w
(3)
ij
n

k=1
k=i,j
Q
k
+
n

i=1
R
(0)
i
w
(0)
i
n

j=1
j=i
Q
j
+
n

i=1
R
(1)
i
w
(1)
i
n

j=1
j=i
Q
j
+
n

i,j=1
i=j
R
(1)
ij
w
(1)
ij
n

k=1
k=i,j
Q
k
+
n

i<j
R
(4)
ij
w
(4)
ij
n

k=1
k=i,j
Q
k
, (57)
where R
(4)
ij
are small of order
ij
and depend on the parameter z, and the functions w
(4)
ij
depend
on the variables (x
i
, x
j
) and the parameter z.
5.3 Loss Distribution (Homogeneous Portfolio)
For simplicity we only consider the homogeneous portfolio case. In addition to the homogeneity
conditions (35) we also assume that
ij
= for all (i, j) and [[ 1.
Using the same notation as in Section 4, the name-name correction term becomes:
n

i<j

ij
u
(ij)
0,0,1
=
n

i<j
R
(4)
ij
w
(4)
ij
n

k=1
k=i,j
Q
k
=
1
2
n(n 1)R
(4)
12
w
(4)
12
(t, x, x)q
n2
,
where from (53), R
(4)
12
=
2
(z).
In particular this implies that the approximation for the loss distribution given by (45) still
holds if we replace the quantity B in (41) by B +B

where
B

1
2
R
(4)
12
w
(4)
12
(t, x, x), (58)
so that (45) becomes
P

(L = k) I
0
+AI
1
+ (B +B

)I
2
. (59)
Note that for a single maturity T the correlations generated by stochastic volatility and name-
name correlation are of the same form to leading order. However if one looks at the term structure
of correlation across several maturities, an important aspect of CDO tranches, then the shape of
the function w
(4)
12
is dierent from the shapes of w
(1)
12
and w
(3)
12
and therefore the nature of the
correlation plays a role.
20
5.4 Numerical Illustration
In order to illustrate the quality of the approximation of the joint survival probability given by (57)
we have conducted the following numerical experiments. For n = 25 names (Table 3), we compute
the zero-order approximation u
0
given by (8), and the rst order approximation u given by (57)
and the explicit formulas derived in Appendix B. We present the results for = 0.02, = 0.5,
and ve cases of
ij
= . The true value proxy u
MC
is obtained by Monte Carlo simulations with
10
5
realizations and using an Euler scheme with time-step of 10
4
as in Section 3.6. The other
parameters are as in Section 3.6.

ij
u
0
u u
MC
Absolute (relative) error
1/50 1/20 0 0.471683 0.486892 0.4803 0.006 (1.4%)
1/50 1/20 0.05 0.471683 0.518151 0.5119 0.006 (1.2%)
1/50 1/20 0.1 0.471683 0.549409 0.5426 0.007 (1.3%)
1/50 1/20 0.2 0.471683 0.611926 0.5986 0.013 (2.2%)
1/50 1/20 0.4 0.471683 0.736961 0.6937 0.043 (6.2%)
Table 3: Joint survival probability for twenty ve rms (n = 25) with stochastic volatility and
name-name correlations.
The presence of a small name-name correlation enhances the importance of the correction
( u u
0
) derived in this paper, as can be seen by comparing u
0
and u
MC
in Table 2 and in Table 3.
Table 4 gives a loss distribution generated with stochastic volatility and name-name correlation,
as computed by (59). The results are plotted in Figure 2. The parameters are taken from Section
3.6, with n = 100, = 1/50, = 1/20,
ij
= 1/10. With these parameters, the coecients in (59)
are: A = 6.607 10
4
, B = 0.014 10
4
, and B

= 2.08 10
4
. The binomial distribution
plotted in Figure 2 is obtained by setting A = B = B

= 0 in (59).
0 1 2 3 4
0.16 0.26 0.17 0.062 0.047
5 6 7 8 9
0.078 0.086 0.065 0.037 0.017
10 11 12 13 14
6.5 10
3
2.2 10
3
6.2 10
4
1.6 10
4
3.7 10
5
Table 4: Loss distribution for 100 rms (n = 100) with stochastic volatility and name-name corre-
lations. All omitted probabilities are less than 10
5
. The expected number of defaults is 2.88 when
stochastic volatility is present, while it is 2.96 in the independent constant volatility case.
The loss distribution shown in Figure 2 clearly has a bimodal structure. With the given pa-
rameters, one can see that the event of having 25 defaults is signicantly less likely than if the
stocks were independent.
21
0 2 4 6 8 10 12 14
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Figure 2: Plot of the loss distribution () given in Table 4. The mass function of the corresponding
binomial distribution () is superimposed for comparison.
A Formal Expansion
Following Fouque et al. (2003), we present the formal expansion used in Section 3, and we discuss
its accuracy. We compute L
,
u
,
where the operator L
,
is given by (2), and the function u
,
is
rst expanded in powers of

corresponding to a regular perturbation:

u
,
= u

0
+

1
+ u

2
+ (60)
In order to cancel the terms of order zero in and the terms of order

in L
,
u
,
= 0 , we get
_
1

L
0
+
1

L
1
+L
2
_
u

0
= 0 (61)
_
1

L
0
+
1

L
1
+L
2
_
u

1
+/
1
u

0
+
1

/
3
u

0
= 0, (62)
with zero boundary value conditions at x
i
= B
i
(t) and a terminal condition one (resp. zero) at t = T
for u

0
(resp. u

1
). We now expand u

0
in powers of

corresponding to a singular perturbation:
u

= u
0
+

u
1,0
+ u
2,0
+

u
3,0

so that (61) becomes:
1

L
0
u
0
+
1

(L
0
u
1,0
+L
1
u
0
) + (L
0
u
2,0
+L
1
u
1,0
+L
2
u
0
)
+

(L
0
u
3,0
+L
1
u
2,0
+L
2
u
1,0
) + = 0
By choosing u
0
and u
1,0
independent of the variable y the terms in 1/ and 1/

cancel. The
operator L
1
taking derivatives with respect to y, the next term becomes
L
0
u
2,0
+L
2
u
0
= 0,
22
which is a Poisson equation in u
2,0
with respect to the variable y, leading to the solvability condition
L
2
)u
0
= 0,
and the problem (6) characterizing u
0
. Subsequently u
2,0
is given by L
1
0
(L
2
L
2
)) u
0
, and
the terms of order

, as a Poisson equation in u
3,0
with respect to the variable y, lead to the
solvability condition
L
2
)u
1,0
= L
1
L
1
0
(L
2
L
2
)))u
0
,
and the problem (9) characterizing u
1,0
.
In order to obtain u
0,1
needed in the approximation (5), we only need the leading order term in
u

1
= u
0,1
+

u
1,1
+ since u

1
is multiplied by

in (60). By choosing u
0,1
and u
1,1
independent
of the variable y, and using the fact that /
3
takes derivatives with respect to y, to leading order
equation (62) becomes:
L
0
u
2,1
+L
2
u
0,1
+/
1
u
0
= 0,
which is again a Poisson equation in u
2,1
with respect to y. Its solvability condition
L
2
)u
0,1
= /
1
)u
0
leads to the problem (22) characterizing u
0,1
.
The accuracy of the the approximation (5) is the same as obtained in Fouque et al. (2006),
that is O(
2/3
log [[ +).
The computation of L
,,
u
,,
, used in Section 5, proceeds along the same lines. The operator
L
,,
is given by (47), and the function u
,,
is rst expanded in powers of
ij
corresponding to a
regular perturbation:
u
,,
= u
,
0
+
n

i<j

ij
u
,,(ij)
1
+
n

i<j
n

k<l

ij

kl
u
,,(ij;kl)
2
+ (63)
Canceling the terms of order zero in
ij
and the terms of order
ij
in L
,,
u
,,
= 0, we get
L
,
u
,
0
= 0 (64)
L
,
u
,
1
= L
(ij)

u
,
0
, (65)
with zero boundary conditions at x
i
= B
i
(t) and a terminal condition one (resp. zero) at t = T for
u
,
0
(resp. u
,
1
) where L
(ij)

is given by (48). Equation (64) has been discussed above. Equation

(65) has a source which is of order zero in and in , so that after expanding rst in

and then
in

, u
(ij)
0,0,1
and u
(ij)
1,0,1
can be chosen independent of y to make the terms in 1/ and 1/

cancel.
After dropping terms which dierentiate u
(ij)
0,0,1
or u
(ij)
1,0,1
with respect to y, to leading order equation
(65) becomes
L
0
u
(ij)
2,0,1
+L
2
u
(ij)
0,0,1
= L
(ij)

u
0,0,0
.
This Poisson equation has the solvability condition
L
2
)u
(ij)
0,0,1
= L
(ij)

)u
0,0,0
,
which is the problem (51) characterizing u
(ij)
0,0,1
.
23
B Explicit Formulas
In order to implement the main formulas (33) and (58) for the approximation u, one needs to
compute the functions w
(l)
i
for each i 1, 2, , n, l 0, 1, 2, 3, and w
(l)
ij
for each pair (i, j)
1, 2, , n 1, 2, , n with i ,= j, and l 1, 3, 4, dened in (16, 17, 18, 27, 28, 29, 54).
We present in this Appendix an ecient method for computing these quantities. The formulas are
given below in Lemmas 2 and 3.
Lemma 2 For each i 1, 2, , n:
a) w
(2)
i
is given by
w
(2)
i
(t, x
i
) = w
(2)
i
(t, x
i
) (T t)F
(2)
i
(t, x
i
),
where
F
(2)
i
(t, x
i
) x
2
i

2
Q
i
x
2
i
= N

_
d
+
2(i)
_
_

d
+
2(i)

2
i
(T t)

1

T t
_
N
_
d

2(i)
_
p
i
(p
i
1)
_
x
i
B
i
(t)
_
p
i
+N

_
d

2(i)
_
_
d

2(i)

2
i
(T t)
+
2p
i
1

T t
_
_
x
i
B
i
(t)
_
p
i
,
w
(2)
i
(t, x
i
) =
_
T
t
g
(2)
i
(s)h
i
(s, x
i
)ds,
g
(2)
i
(t) (T t) lim
x
i
B
i
(t)
F
(2)
i
(t, x
i
)
= N

(d
i
)
2(p
i
1)

T t N(d
i
)p
i
(p
i
1)(T t),
h
i
(s, x
i
) =
ln
x
i
K
i

i
t

i
_
2(s t)
3
exp
_

_
ln
x
i
K
i

i
t
p
i

2
i
2
(s t)
_
2
2
2
i
(s t)
_

_
,
d
i
=
(r
i

2
i
/2)(T t)

T t
=
p
i

i
2

T t,
with N

() being the standard normal density function, and

i

i
(z) when no possible confusion
is caused.
b) w
(3)
i
is given by
w
(3)
i
(t, x
i
) = w
(3)
i
(t, x
i
) (T t)F
(3)
i
(t, x
i
),
where
F
(3)
i
(t, x
i
) x
i

x
i
_
x
2
i

2
Q
i
x
2
i
_
24
= N

_
d
+
2(i)
_
_

_
_
d
+
2(i)
_
2
1

3
i
(T t)
3/2
+
d
+
2(i)

2
i
(T t)
_

_ + N
_
d

2(i)
_
p
2
i
(1 p
i
)
_
x
i
B
i
(t)
_
p
i
+N

_
d

2(i)
_
_

_
_
d

2(i)
_
2
1

3
i
(T t)
3/2
+
d

2(i)
(3p
i
1)

2
i
(T t)
+
p
i
(3p
i
2)

T t
_

_
_
x
i
B
i
(t)
_
p
i
,
w
(3)
i
(t, x
i
) =
_
T
t
g
(3)
i
(s)h
i
(s, x
i
)ds,
g
(3)
i
(t) (T t) lim
x
i
B
i
(t)
F
(3)
i
(t, x
i
)
= N

(d
i
)
1

3
i
_
2

T t
+ 4p
i
r

T t
_
+ N(d
i
)p
2
i
(1 p
i
)(T t).
c) w
(0)
i
is given by
w
(0)
i
(t, x
i
) = w
(0)
i
(t, x
i
) + (T t)F
(0)
i
(t, x
i
)
1
2
(T t)
2

i
F
(2)
i
(t, x
i
),
where
F
(0)
i
(t, x
i
)
Q
i

i
= N

_
d
+
2(i)
_
_

d
+
2(i)

T t
_
+ N

_
d

2(i)
_
_
d

2(i)

i
+

T t
_
_
x
i
B
i
(t)
_
p
i
+N
_
d

2(i)
_
_
2

i
(p
i
1) ln
_
x
i
B
i
(t)
__ _
x
i
B
i
(t)
_
p
i
,
w
(0)
i
(t, x
i
) =
_
T
t
g
(0)
i
(s)h
i
(s, x
i
)ds,
g
(0)
i
(t) (T t) lim
x
i
B
i
(t)
F
(0)
i
(t, x
i
) +
1
2
(T t)
2

i
lim
x
i
B
i
(t)
F
(2)
i
(t, x
i
)
= N

(d
i
)(p
i
1)(T t)
3/2
N(d
i
)

i
2
p
i
(p
i
1)(T t)
2
.
d) w
(1)
i
is given by
w
(1)
i
(t, x
i
) = w
(1)
i
(t, x
i
) + (T t)F
(1)
i
(t, x
i
)
1
2
(T t)
2

i
F
(3)
i
(t, x
i
),
where
F
(1)
i
(t, x
i
) x
i

x
i
_
Q
i

i
_
= N

_
d
+
2(i)
_
_

_
_
d
+
2(i)
_
2
1

2
i

T t
+
d
+
2(i)

i
_

_
+ N

_
d

2(i)
_
_

_
_
d

2(i)
_
2
1

2
i

T t
25
+(p
i
+ 1)
d

2(i)

i
+p
i

T t +
2(1 p
i
)

2
i

T t
ln
_
x
i
B
i
(t)
_
_
_
x
i
B
i
(t)
_
p
i
+N
_
d

2(i)
_
_
2

i
(p
i
1)
_
1 +p
i
ln
_
x
i
B
i
(t)
___ _
x
i
B
i
(t)
_
p
i
,
w
(1)
i
(t, x
i
) =
_
T
t
g
(1)
i
(s)h
i
(s, x
i
)ds,
g
(1)
i
(t) (T t) lim
x
i
B
i
(t)
F
(1)
i
(t, x
i
) +
1
2
(T t)
2

i
lim
x
i
B
i
(t)
F
(3)
i
(t, x
i
)
= N

(d
i
)
1

2
i
_

T t 2p
i
r(T t)
3/2
_
N(d
i
)
_
2

i
(p
i
1)(T t) +

i
2
p
2
i
(p
i
1)(T t)
2
_
.
Proof We use the same technique as in Fouque, Sircar and Solna (2005), but here we need to deal
with the time varying hitting boundary B
i
(t), whereas they considered a constant hitting boundary.
a) If we dene
L
(i)
2


t
+
1
2
f
2
i
(y, z)x
2
i

2
x
2
i
+rx
i

x
i
,
then L
(i)
2
)x
k
i

k
Q
i
x
k
i
= 0 for k 1, where we used the fact that L
(i)
2
) commutes with x
k
i

k
x
k
i
and the
fact that L
(i)
2
)Q
i
= 0. Now we dene
w
(2)
i
w
(2)
i
+ (T t)x
2
i

2
Q
i
x
2
i
,
which satises
L
(i)
2
) w
(2)
i
(t, x
i
) = 0, x
i
> B
i
(t), t < T,
w
(2)
i
(t, B
i
(t)) = (T t) lim
x
i
B
i
(t)
x
2
i

2
Q
i
x
2
i
g
(2)
i
(t), t T,
w
(2)
i
(T, x
i
) = 0, x
i
> B
i
(t).
Change of variables: dene

i

1

i
ln
x
i
B
i
(t)
,
i
=
r
i

2
i
/2

i
=
p
i

i
2
, w
(2)
i
(t,
i
) w
(2)
i
(t, x
i
).
Then w
(2)
i
(t,
i
) satises
w
(2)
i
t
+
1
2

2
w
(2)
i

2
i
+
i
w
(2)
i

i
= 0,
i
> 0, t < T,
w
(2)
i
(t, 0) = g
(2)
i
(t), t T,
w
(2)
i
(T,
i
) = 0,
i
> 0.
26
It follows that, see, for example, Karatzas and Shreve (1991), w
(2)
i
admits the probabilistic repre-
sentation:
w
(2)
i
(t,
i
) = E
Q
_
g
(2)
i
(

)1
{

T}

(i)
t
=
i
_
,
where
(i)
is a Brownian motion with drift
i
under probability measure Q, and

is the rst time

that
(i)
hits 0. By plugging in the distribution of

, which is well known and can be found, for

example, in Borodin and Salminen (2002), we obtain
w
(2)
i
(t,
i
) =
_
T
t
g
(2)
i
(s)

i
_
2(s t)
3
exp
_

[
i
+
i
(s t)]
2
2(s t)
_
ds
=
_
T
t
g
(2)
i
(s)h
i
(s)ds.
So in summary,
w
(2)
i
= w
(2)
i
(T t)x
2
i

2
Q
i
x
2
i
= w
(2)
i
(T t)F
(2)
i
(t, x
i
).
b) The same as that for part a).
c) Note that
L
(i)
2
)
Q
i

i
=
i
x
2
i

2
Q
i
x
2
i
,
which can be obtained by taking derivative with respect to
i
on equation L
(i)
2
)Q
i
= 0. Now dene
w
(0)
i
= w
(0)
i
(T t)
Q
i

i
+
1
2
(T t)
2

i
x
2
i

2
Q
i
x
2
i
,
then w
(0)
i
(t, x
i
) satises
L
(i)
2
) w
(0)
i
(t, x
i
) = 0, x
i
> B
i
(t), t < T,
w
(0)
i
(t, B
i
(t)) = g
(0)
i
(t), t T,
w
(0)
i
(T, x
i
) = 0, x
i
> B
i
(t).
The rest of the proof is as in part a).
d) The same as that for part c), using the fact that
L
(i)
2
)x
i

x
i
_
Q
i

i
_
=
i
x
i

x
i
_
x
2
i

2
Q
i
x
2
i
_
.
All the calculations of the formulas within this lemma are straightforward. The proof is com-
plete.
27
Lemma 3 For each pair (i, j) 1, 2, , n 1, 2, , n with i ,= j:
a) w
(3)
ij
is given by
w
(3)
ij
(t, x
i
, x
j
) = w
(3)
ij
(t, x
i
, x
j
) (T t)F
(3)
ij
(t, x
i
, x
j
),
where
F
(3)
ij
(t, x
i
, x
j
)
_
x
i
Q
i
x
i
_
_
x
2
j

2
Q
j
x
2
j
_
= F
(2)
j
(t, x
j
)
_
N

_
d
+
2(i)
_
1

T t
+N

_
d

2(i)
_
1

T t
_
x
i
B
i
(t)
_
p
i
N
_
d

2(i)
_
p
i
_
x
i
B
i
(t)
_
p
i
_
,
w
(3)
ij
(t, x
i
, x
j
) =
_
T
t
_

0

G
(j)
ij
(s, )
j
(s, , x
j
)h
ij
(s, x
i
, x
j
)dds
+
_
T
t
_

0

G
(i)
ij
(s, )
i
(s, , x
i
)h
ji
(s, x
j
, x
i
)dds,

j
(s, , x
j
) =
2
_
2(s t)
sinh
_
ln
x
j
B
j
(t)

j
(s t)
_
exp
_

2
2(s t)

p
j

j
2

_
,
h
ij
(s, x
i
, x
j
) = h
i
(s, x
i
) exp
_

_
ln
x
j
K
j

j
t
p
j

2
j
2
(s t)
_
2
2
2
j
(s t)
_

_
,

G
(j)
ij
(t, ) (T t)F
(3)
ij
_
t, B
i
(t), B
j
(t)e

_
=
_
N

(d
i
)
2

T t N(d
i
)p
i
(T t)
_

_
N

d
+
2(j)
_
_

d
+
2(j)

2
j
(T t)

1

T t
_
N
_

2(j)
_
p
j
(p
j
1)e
p
j

+N

2(j)
_
_

d

2(j)

2
j
(T t)
+
2p
j
1

T t
_
e
p
j

_
,

G
(i)
ij
(t, ) (T t)F
(3)
ij
_
t, B
i
(t)e

, B
j
(t)
_
=
_
N

(d
j
)
2(p
j
1)

T t N(d
j
)p
j
(p
j
1)(T t)
_

_
N

d
+
2(i)
_
1

T t
+ N

2(i)
_
e
p
i

T t
N
_

2(i)
_
p
i
e
p
i

_
,

2(i)
=
p
i

i
2

T t.
b) w
(1)
ij
is given by
w
(1)
ij
(t, x
i
, x
j
) = w
(1)
ij
(t, x
i
, x
j
) + (T t)F
(1)
ij
(t, x
i
, x
j
)
1
2
(T t)
2

j
F
(3)
ij
(t, x
i
, x
j
),
28
where
F
(1)
ij
(t, x
i
, x
j
)
_
x
i
Q
i
x
i
__
Q
j

j
_
= F
(0)
j
(t, x
j
)
_
N

_
d
+
2(i)
_
1

T t
+N

_
d

2(i)
_
1

T t
_
x
i
B
i
(t)
_
p
i
N
_
d

2(i)
_
p
i
_
x
i
B
i
(t)
_
p
i
_
,
w
(1)
ij
(t, x
i
, x
j
) =
_
T
t
_

0

H
(j)
ij
(s, )
j
(s, , x
j
)h
ij
(s, x
i
, x
j
)dds
+
_
T
t
_

0

H
(i)
ij
(s, )
i
(s, , x
i
)h
ji
(s, x
j
, x
i
)dds,

H
(j)
ij
(t, ) (T t)F
(1)
ij
_
t, B
i
(t), B
j
e

_
+
1
2
(T t)
2

j
F
(3)
ij
_
t, B
i
(t), B
j
e

_
=
_
N

(d
i
)
2

T t N(d
i
)p
i
(T t)
_

_
N

d
+
2(j)
_
_
d
+
2(j)
2
j
+
1
2

T t
_
+ N

2(j)
_
_

2(j)
2
j
+ (p
j

3
2
)

T t
_
e
p
j

N
_

2(j)
_
(p
j
1)
_
2 +
p
j

j
2
(T t)
_
e
p
j

_
,

H
(i)
ij
(t, ) (T t)F
(1)
ij
_
t, B
i
(t)e

, B
j
(t)
_
+
1
2
(T t)
2

j
F
(3)
ij
_
t, B
i
(t)e

, B
j
(t)
_
=
_
N

(d
j
)(p
j
1)(T t)
3/2
N(d
j
)
p
j

j
2
(p
j
1)(T t)
2
_

_
N

d
+
2(i)
_
1

T t
+ N

2(i)
_
e
p
i

T t
N
_

2(i)
_
p
i
e
p
i

_
.
c) w
(4)
ij
is given by
w
(4)
ij
(t, x
i
, x
j
) = w
(4)
ij
(t, x
i
, x
j
) (T t)F
(4)
ij
(t, x
i
, x
j
),
where
F
(4)
ij
(t, x
i
, x
j
) =
_
x
i
Q
i
x
i
__
x
j
Q
j
x
j
_
=
_
N

_
d
+
2(i)
_
1

T t
+ N

_
d

2(i)
_
1

T t
_
x
i
B
i
(t)
_
p
i
N
_
d

2(i)
_
p
i
_
x
i
B
i
(t)
_
p
i
_ _
N
_
d

2(j)
_
p
j
_
x
j
B
j
(t)
_
p
j
+N

_
d
+
2(j)
_
1

T t
+ N

_
d

2(j)
_
1

T t
_
x
j
B
j
(t)
_
p
j
_
,
w
(4)
ij
(t, x
i
, x
j
) =
_
T
t
_

0

K
(j)
ij
(s, )
j
(s, , x
j
)h
ij
(s, x
i
, x
j
)dds
+
_
T
t
_

0

K
(i)
ij
(s, )
i
(s, , x
i
)h
ji
(s, x
j
, x
i
)dds,
29

K
(j)
ij
(t, ) (T t)F
(4)
ij
_
t, B
i
(t), B
j
(t)e

_
,

K
(i)
ij
(t, ) (T t)F
(4)
ij
_
t, B
i
(t)e

, B
j
(t)
_
.
Proof
a) If we dene
L
(ij)
2


t
+
1
2
f
2
i
(y, z)x
2
i

2
x
2
i
+
1
2
f
2
j
(y, z)x
2
j

2
x
2
j
+rx
i

x
i
+rx
j

x
j
,
then L
(ij)
2
)x
k
i

k
Q
i
x
k
i
= 0 for k 1. Next we dene
w
(3)
ij
w
(3)
ij
+ (T t)
_
x
i
Q
i
x
i
_
_
x
2
j

2
Q
j
x
2
j
_
,
which satises
L
(ij)
2
) w
(3)
ij
(t, x
i
, x
j
) = 0, x
i
> B
i
(t), x
j
> B
j
(t), t < T,
w
(3)
ij
(t, B
i
(t), x
j
) = (T t)F
(3)
ij
(t, B
i
(t), x
j
) G
(j)
ij
(t, x
j
), x
j
> B
j
(t), t T,
w
(3)
ij
(t, x
i
, B
j
(t)) = (T t)F
(3)
ij
(t, x
i
, B
j
(t)) G
(i)
ij
(t, x
i
), x
i
> B
i
(t), t T,
w
(3)
ij
(T, x
i
, x
j
) = 0, x
i
> B
i
(t), x
j
> B
j
(t).
Change of variables: dene

i

1

i
ln
x
i
B
i
(t)
,
j

1

j
ln
x
j
B
j
(t)
,

i

r
i

2
i
/2

i
=
p
i

i
2
,
j

r
j

2
j
/2

j
=
p
j

j
2
,
and also
w
(3)
ij
(t,
i
,
j
) w
(3)
ij
(t, x
i
, x
j
),

G
(j)
ij
(t,
j
) G
(j)
ij
(t, x
j
) = G
(j)
ij
_
t, B
j
(t)e

j
_
,

G
(i)
ij
(t,
i
) G
(i)
ij
(t, x
i
) = G
(i)
ij
_
t, B
i
(t)e

i
_
.
Then we have
w
(3)
ij
t
+
1
2

2
w
(3)
ij

2
i
+
1
2

2
w
(3)
ij

2
j
+
i
w
(3)
ij

i
+
j
w
(3)
ij

j
= 0,
i
> 0,
j
> 0, t < T,
w
(3)
ij
(t, 0,
j
) =

G
(j)
ij
(t,
j
),
j
> 0, t T,
w
(3)
ij
(t,
i
, 0) =

G
(i)
ij
(t,
i
),
i
> 0, t T,
w
(3)
ij
(T,
i
,
j
) = 0,
i
> 0,
j
> 0, t < T.
30
The function w
(3)
ij
admits the probabilistic representation:
w
(3)
ij
(t,
i
,
j
) = E
Q
_

G
(j)
ij
_
,
(j)

_
1
{ =
i
}
1
{ T}
+

G
(i)
ij
_
,
(i)

_
1
{ =
j
}
1
{ T}

(i)
t
=
i
,
(j)
t
=
j
_
,
where
(i)
and
(j)
are independent Brownian motions with drifts
i
and
j
, respectively, under
probability measure Q,
i
(
j
) is the rst time that
(i)
(
(j)
) hits 0, and = min
i
,
j
. From
Borodin and Salminen (2002), we have

h
ij
(s, ;
i
,
j
) Q
_
ds, =
i
,
(j)

d

(i)
t
=
i
,
(j)
t
=
j
_
=
2
_
2(s t)
sinh
_

j
s t
_
exp
_

j


2
2(s t)
_
exp
_

[
j
+
j
(s t)]
2
2(s t)
_

i
_
2(s t)
3
exp
_

[
i
+
i
(s t)]
2
2(s t)
_
dds.
Therefore,
w
(3)
ij
(t,
i
,
j
) =
_
T
t
_

0

G
(j)
ij
(s, )

h
ij
(s, ;
i
,
j
)dds
+
_
T
t
_

0

G
(i)
ij
(s, )

h
ji
(s, ;
j
,
i
)dds.
Now change (
i
,
j
) back to (x
i
, x
j
) to obtain the desired result.
b) Dene
w
(1)
ij
w
(1)
ij
(T t)
_
x
i
Q
i
x
i
__
Q
j

j
_
+
1
2
(T t)
2

j
_
x
i
Q
i
x
i
_
_
x
2
j

2
Q
j
x
2
j
_
.
Then w
(1)
ij
satises
L
(ij)
2
) w
(1)
ij
(t, x
i
, x
j
) = 0, x
i
> B
i
(t), x
j
> B
j
(t), t < T,
w
(1)
ij
(t, B
i
(t), x
j
) = (T t)F
(1)
ij
(t, B
i
(t), x
j
) +
1
2
(T t)
2

j
F
(3)
ij
(t, B
i
(t), x
j
),
x
j
> B
j
(t), t T,
w
(1)
ij
(t, x
i
, B
j
(t)) = (T t)F
(1)
ij
(t, x
i
, B
j
(t)) +
1
2
(T t)
2

j
F
(3)
ij
(t, x
i
, B
j
(t)),
x
i
> B
i
(t), t T,
w
(1)
ij
(T, x
i
, x
j
) = 0, x
i
> B
i
(t), x
j
> B
j
(t).
The rest of the proof is similar to that of part a).
c) Dene
w
(4)
ij
w
(4)
ij
(T t)
_
x
i
Q
i
x
i
__
x
j
Q
j
x
j
_
.
31
Then w
(4)
ij
satises
L
(ij)
2
) w
(4)
ij
(t, x
i
, x
j
) = 0, x
i
> B
i
(t), x
j
> B
j
(t), t < T,
w
(1)
ij
(t, B
i
(t), x
j
) = (T t)F
(4)
ij
(t, B
i
(t), x
j
),
x
j
> B
j
(t), t T,
w
(4)
ij
(t, x
i
, B
j
(t)) = (T t)F
(4)
ij
(t, x
i
, B
j
(t)),
x
i
> B
i
(t), t T,
w
(4)
ij
(T, x
i
, x
j
) = 0, x
i
> B
i
(t), x
j
> B
j
(t).
The rest of the proof is similar to that of part a). The proof is complete.
Acknowledgment
The rst author would like to thank Ronnie Sircar and Knut Slna for the many helpful discussions
on the subject. The authors thank an anonymous referee for helpful comments which improved an
earlier version of the paper.
References
[1] F. Black and J. C. Cox (1976). Valuing Corporate Securities: Some Eects of Bond Indenture
Provisions. Journal of Finance, 31, 351367.
[2] A. N. Borodin and P. Salminen (2002). Handbook of Brownian motion: Facts and Formulae,
2nd ed., Birkahauser Verlag.
[3] D. Due and K. Singleton (2003) Credit Risk, Princeton University Press.
[4] D. Due and N. Garleanu (2001). Risk and Valuation of Collateralized Debt Obligations.
Financial Analysts Journal, 57(1) 4159.
[5] J.P. Fouque, G. Papanicolaou and R. Sircar (2000). Derivatives in Financial Markets with
Stochastic Volatility, Cambridge University Press.
[6] J.P. Fouque, G. Papanicolaou, R. Sircar and K. Solna (2003). Multiscale Stochastic Volatility
Asymptotics. SIAM Journal on Multiscale Modeling and Simulation, 2(1) 2242.
[7] J.P. Fouque, R. Sircar and K. Solna (2006). Stochastic Volatility Eects on Defaultable Bonds.
Applied Mathematical Finance, 13(3) 215244.
[8] I. Karatzas and S. Shreve (1991). Brownian Motion and Stochastic Calculus, 2nd ed., Springer-
Verlag.
[9] C. Zhou (2001). An Analysis of Default Correlations and Multiple Defaults. Review of Finan-
cial Studies, 14(2) 555576.
32