Lip Ton 2010

CVA for CDS via SDM and Related Issues
Analytical and Numerical Results

Alexander Lipton and Artur Sepp
Bank of America Merrill Lynch
March 2010
A Lipton and A Sepp (Bank of America Merrill Lynch) Credit Modeling 02/03 1 / 74
Plan
A brief history of credit modeling.
How to model CDSs?
Reduced, structural, and hybrid single-name models.
Examples.
Multi-name structural models.
Examples.
Plan
How to model CDSs?
Examples.
Examples.
Plan
How to model CDSs?
Examples.
Examples.
Plan
How to model CDSs?
Examples.
Examples.
Plan
How to model CDSs?
Examples.
Examples.
Plan
How to model CDSs?
Examples.
Examples.
Acknowledgements
We am also grateful to other Merrill Lynch colleagues, especially S.
Inglis, A. Rennie, D. Shelton, I. Savescu, and members of the Credit
Analytics Team and Structured Credit Team for all their help.
References
Selected references:
Reduced form model: Jarrow & Turnbull (2000), Hughston &
Turnbull (2001), Lando (1998), Madan & Unal (1994).
Structural model: Black & Cox (1976), Hilberink & Rogers (2002),
Hyer et al (1998), Kiesel & Scherer (2007), Leland (1994), Longsta
& Schwartz (1995), Lipton (2002), Merton (1974), Sepp (2004,
2006), Zhou (2001a), Zhou (2001a).
Counterparty risk: Blanchet & Patras (2008), Brigo & Chourdakis
(2008), Crepey, Jeanblanc & Zargari (2009), Leung & Kwok (2005),
Haworth, Reisisnger & Shaw (2006), Hull & White (2000), (2001),
Misisrpashayev (2008), Turnbull (2005), Valuzis (2008).
Numerical methods: Andersen & Andreasen (2000), Boyarchenko &
Levendorsky (2000), Cont & Tankov (2004), Cont & Voltchkova
(2005), Clift & Forsyth (2008), dHalluin, Forsyth & Vetzal (2005),
Feng & Linetsky (2008), Lipton (2003).
References
2006), Zhou (2001a), Zhou (2001a).
References
2006), Zhou (2001a), Zhou (2001a).
References
2006), Zhou (2001a), Zhou (2001a).
References
2006), Zhou (2001a), Zhou (2001a).
Moodys number of defaults gure
All rated, Source Moody's
0
50
100
150
200
250
1
9
2
0
1
9
2
4
1
9
2
8
1
9
3
2
1
9
3
6
1
9
4
0
1
9
4
4
1
9
4
8
1
9
5
2
1
9
5
6
1
9
6
0
1
9
6
4
1
9
6
8
1
9
7
2
1
9
7
6
1
9
8
0
1
9
8
4
1
9
8
8
1
9
9
2
1
9
9
6
2
0
0
0
2
0
0
4
Moodys table of recoveries
Source: Moody's
Year Sr. Secured Sr. Secured Sr. Unsecured Sr. Subordinated Subordinated Jr. Subordinated All
Bank Loans Bonds Bonds Bonds Bonds Bonds Bonds
1982 NA $72.50 $35.79 $48.09 $29.99 NA $35.57
1983 NA $40.00 $52.72 $43.50 $40.54 NA $43.64
1984 NA NA $49.41 $67.88 $44.26 NA $45.49
1985 NA $83.63 $60.16 $30.88 $39.42 $48.50 $43.66
1986 NA $59.22 $52.60 $50.16 $42.58 NA $48.38
1987 NA $71.00 $62.73 $44.81 $46.89 NA $50.48
1988 NA $55.40 $45.24 $33.41 $33.77 $36.50 $38.98
1989 NA $46.54 $43.81 $34.57 $26.36 $16.85 $32.31
1990 $75.25 $33.81 $37.01 $25.64 $19.09 $10.70 $25.50
1991 $74.67 $48.39 $36.66 $41.82 $24.42 $7.79 $35.53
1992 $61.13 $62.05 $49.19 $49.40 $38.04 $13.50 $45.89
1993 $53.40 NA $37.13 $51.91 $44.15 NA $43.08
1994 $67.59 $69.25 $53.73 $29.61 $38.23 NA $45.57
1995 $75.44 $62.02 $47.60 $34.30 $41.54 NA $43.28
1996 $88.23 $47.58 $62.75 $43.75 $22.60 NA $41.54
1997 $78.75 $75.50 $56.10 $44.73 $35.96 $30.58 $49.39
1998 $51.40 $48.14 $41.63 $44.99 $18.19 $62.00 $39.65
1999 $75.82 $43.00 $38.04 $28.01 $35.64 NA $34.33
2000 $68.32 $39.23 $23.81 $20.75 $31.86 $15.50 $25.18
2001 $66.16 $37.98 $21.45 $19.82 $15.94 $47.00 $22.21
2002 $58.80 $48.37 $29.69 $23.21 $24.51 NA $30.18
2003 $73.43 $63.46 $41.87 $37.27 $12.31 NA $40.69
2004 $87.74 $73.25 $54.25 $46.54 $94.00 NA $59.12
2005 $82.07 $71.93 $54.88 $26.06 $51.25 NA $55.97
2006 $76.02 $74.63 $55.02 $41.41 $56.11 NA $55.02
2007 $67.74 $80.54 $51.02 $54.47 NA NA $53.53
CDS gure
0
50
100
150
200
250
300
350
400
29/3/07 28/5/07 27/7/07 25/9/07 24/11/07 23/1/08 23/3/08
IBM 5Y GE 5Y MER 5Y DB 5Y DCX 5Y SIEM 5Y
Figure: Time series of 5Y CDS spreads (in bps) for 6 typical companies (IBM, GE,
Merrill Lynch, Deutsche Bank, Daimler, and Siemens). Source: Merrill Lynch.
Recent trends
The volume of CDSs grew by a factor of 100 between 2001 and 2007.
Lately, CDS dealers reduced their market by 38% in one year.
According to the most recent ISDA survery published on April 22nd,
the notional value of CDSs outstanding decreased to $38.6 trillion as
of Dec 31st 2008 from $54.6 trillion mid-year and $62.2 trillion as of
Dec 31st 2007.
Recent trends
Dec 31st 2007.
Recent trends
Dec 31st 2007.
CDS modeling
First, we need to explain how to price CDSs.
Then we need to extend our theory to cover indices, tranches,
baskets, etc.
Three complementary approaches to pricing CDSs:
reduced form,
structural,
hybrid
All of them have pros and cons. Reduced form and structural
approaches dominate.
CDS modeling
baskets, etc.
reduced form,
structural,
hybrid
CDS modeling
baskets, etc.
reduced form,
structural,
hybrid
CDS modeling
baskets, etc.
reduced form,
structural,
hybrid
CDS modeling
baskets, etc.
reduced form,
structural,
hybrid
CDS modeling
baskets, etc.
reduced form,
structural,
hybrid
CDS modeling
baskets, etc.
reduced form,
structural,
hybrid
Simple non-risky model
First, we consider non-risky bonds and introduce the short interest
rate r (t) which is driven by the SDE:
dr (t) = f
1
(t, r (t)) dt +g
1
(t, r (t)) dW
1
(t) r (0) = r
0
Let D(t, T) be time t price of the non-risky zero-coupon bond
maturing at time T. This price can be written as follows:
D (t, T) = V
1
(t, r (t)) D (0, T) = V
1
(0, r
0
)
where V
1
(t, r ) is the solution of the following backward Kolmogorov
problem
V
1,t
+
1
2
g
2
1
V
1,rr
+f
1
V
1,r
rV
1
= 0 V
1
(T, r ) = 1
Functions f
1
, g
1
are calibrated to the market to match the yield curve
and some swaption volatilities. Alternatively,
D (t, T) = E
t
_
exp
_
_
T
t
r
_
t
/
_
dt
/
__
= exp
_
_
T
t
r
_
t, t
/
_
dt
/
_
where r is the forward rate.
dr (t) = f
1
(t, r (t)) dt +g
1
(t, r (t)) dW
1
(t) r (0) = r
0
D (t, T) = V
1
(t, r (t)) D (0, T) = V
1
(0, r
0
)
where V
1
problem
V
1,t
+
1
2
g
2
1
V
1,rr
+f
1
V
1,r
rV
1
= 0 V
1
(T, r ) = 1
Functions f
1
, g
1
D (t, T) = E
t
_
exp
_
_
T
t
r
_
t
/
_
dt
/
__
= exp
_
_
T
t
r
_
t, t
/
_
dt
/
_
dr (t) = f
1
(t, r (t)) dt +g
1
(t, r (t)) dW
1
(t) r (0) = r
0
D (t, T) = V
1
(t, r (t)) D (0, T) = V
1
(0, r
0
)
where V
1
problem
V
1,t
+
1
2
g
2
1
V
1,rr
+f
1
V
1,r
rV
1
= 0 V
1
(T, r ) = 1
Functions f
1
, g
1
D (t, T) = E
t
_
exp
_
_
T
t
r
_
t
/
_
dt
/
__
= exp
_
_
T
t
r
_
t, t
/
_
dt
/
_
Reduced form model formulation
The name defaults at the rst time a Cox process jumps from 0 to 1.
SDE for the default intensity (hazard rate) X (t) is:
dX (t) = f (t, X (t)) dt +g (t, X (t)) dW (t) +JdN (t) , X (0) = X
0
where W (t) is a standard Wiener processes, N (t) is a Poisson
process with intensity (t), and J is a positive jump distribution;
W, N, J are mutually independent.
We impose the following constraints
f (t, 0) _ 0, f (t, ) < 0, g (t, 0) = 0
For analytical convenience (rather than for deeper reasons) it is
customary to assume that X is governed by the square-root stochastic
dierential equation (SDE):
dX (t) = ( (t) X (t)) dt +
_
X (t)dW (t) +JdN (t) , X (0) = X
0
with exponential (or hyper-exponential) jump distribution.
0
f (t, 0) _ 0, f (t, ) < 0, g (t, 0) = 0
dX (t) = ( (t) X (t)) dt +
_
X (t)dW (t) +JdN (t) , X (0) = X
0
0
f (t, 0) _ 0, f (t, ) < 0, g (t, 0) = 0
dX (t) = ( (t) X (t)) dt +
_
X (t)dW (t) +JdN (t) , X (0) = X
0
Survival probability
For practical purposes it is more convenient to consider discrete jump
distributions with jump values J
m
> 0, 1 _ m _ M, occurring with
probabilities
m
> 0; such distributions are more exible than
parametric ones because they allow one to place jumps where they are
needed.
In this framework, the survival probability of the name from time 0 to
time T has the form
q (0, T) = E
0
_
e
_
T
0
X(t
/
)dt
/
_
= E
0
_
e
Y (T)
_
Here Y (t) is governed by the following degenerate SDE:
dY (t) = X (t) dt, Y (0) = 0
More generally, the survival probability from time t to time T
conditional on no default before time t has the form
q (t, T[ X (t) , Y (t)) = e
Y (t)
I
(>t)
E
t
_
e
Y (T)
X (t) , Y (t)
_
m
probabilities
m
needed.
time T has the form
q (0, T) = E
0
_
e
_
T
0
X(t
/
)dt
/
_
= E
0
_
e
Y (T)
_
dY (t) = X (t) dt, Y (0) = 0
q (t, T[ X (t) , Y (t)) = e
Y (t)
I
(>t)
E
t
_
e
Y (T)
X (t) , Y (t)
_
m
probabilities
m
needed.
time T has the form
q (0, T) = E
0
_
e
_
T
0
X(t
/
)dt
/
_
= E
0
_
e
Y (T)
_
dY (t) = X (t) dt, Y (0) = 0
q (t, T[ X (t) , Y (t)) = e
Y (t)
I
(>t)
E
t
_
e
Y (T)
X (t) , Y (t)
_
m
probabilities
m
needed.
time T has the form
q (0, T) = E
0
_
e
_
T
0
X(t
/
)dt
/
_
= E
0
_
e
Y (T)
_
dY (t) = X (t) dt, Y (0) = 0
q (t, T[ X (t) , Y (t)) = e
Y (t)
I
(>t)
E
t
_
e
Y (T)
X (t) , Y (t)
_
Calculation of expectations
Expectations of the form E
t
_
e
Y (T)
X (t) , Y (t)
_
, can be
computed by solving the following augmented partial dierential
equation (PDE)
(
t
+/) V (t, T, X, Y) +XV
Y
(t, T, X, Y) = 0
V (T, T, X, Y) = e
Y
where
/V = ( (t) X) V
X
+
1
2
2
XV
XX
+
m

m
[V (X +J
m
) V (X)]
Specically, the following relation holds
E
t
_
e
Y (T)
X (t) , Y (t)
_
= V (t, T, X (t) , Y (t))
Calculation of expectations
Expectations of the form E
t
_
e
Y (T)
X (t) , Y (t)
_
, can be
computed by solving the following augmented partial dierential
equation (PDE)
(
t
+/) V (t, T, X, Y) +XV
Y
(t, T, X, Y) = 0
V (T, T, X, Y) = e
Y
where
/V = ( (t) X) V
X
+
1
2
2
XV
XX
+
m

m
[V (X +J
m
) V (X)]
Specically, the following relation holds
E
t
_
e
Y (T)
X (t) , Y (t)
_
= V (t, T, X (t) , Y (t))
Ane ansatz
The corresponding solution can be written in the so-called ane form:
V (t, T, X, Y) = e
a(t,T,)+b(t,T,)XY
where a, b are functions of time governed by the following system of
ordinary dierential equations (ODEs):
_
_
_
da(t,T,)
dt
= (t) b (t, T, )
m

m
_
e
J
m
b(t,T,)
1
_
db(t,T,)
dt
= +b (t, T, )
1
2
2
b
2
(t, T, )
a (T, T, ) = 0, b (T, T, ) = 0
This system cannot be solved analytically, it is very easy to solve it
numerically. The survival probability q (0, T) and default probability
p (0, T) have the form
q (0, T) = e
a(0,T,1)+b(0,T,1)X
0
,
p (0, T) = 1 q (0, T) = 1 e
a(0,T,1)+b(0,T,1)X
0
Ane ansatz
The corresponding solution can be written in the so-called ane form:
V (t, T, X, Y) = e
a(t,T,)+b(t,T,)XY
where a, b are functions of time governed by the following system of
ordinary dierential equations (ODEs):
_
_
_
da(t,T,)
dt
= (t) b (t, T, )
m

m
_
e
J
m
b(t,T,)
1
_
db(t,T,)
dt
= +b (t, T, )
1
2
2
b
2
(t, T, )
a (T, T, ) = 0, b (T, T, ) = 0
This system cannot be solved analytically, it is very easy to solve it
numerically. The survival probability q (0, T) and default probability
p (0, T) have the form
q (0, T) = e
a(0,T,1)+b(0,T,1)X
0
,
p (0, T) = 1 q (0, T) = 1 e
a(0,T,1)+b(0,T,1)X
0
CDS valuation 1
The value U of a credit default swap (CDS) paying an up-front
amount and a coupon s in exchange for receiving 1 R (where R is
the default recovery) on default as follows:
U = +V (0, X
0
)
Here V (t, X) solves the following pricing problem
(
t
+/) V (t, X) (r +X) V (t, X) = s (1 R) X
V (T, X) = 0
Using Duhamels principle, we obtain the following expression for V:
V (t, X) = s
_
T
t
D
_
t, t
/
_
e
a(t,t
/
,1)+b(t,t
/
,1)X
dt
/
(1 R)
_
T
t
D
_
t, t
/
_
d
_
e
a(t,t
/
,1)+b(t,t
/
,1)X
_
where D (t, t
/
) is the discount factor between two times t, t
/
.
CDS valuation 1
U = +V (0, X
0
)
(
t
+/) V (t, X) (r +X) V (t, X) = s (1 R) X
V (T, X) = 0
V (t, X) = s
_
T
t
D
_
t, t
/
_
e
a(t,t
/
,1)+b(t,t
/
,1)X
dt
/
(1 R)
_
T
t
D
_
t, t
/
_
d
_
e
a(t,t
/
,1)+b(t,t
/
,1)X
_
where D (t, t
/
/
.
CDS valuation 1
U = +V (0, X
0
)
(
t
+/) V (t, X) (r +X) V (t, X) = s (1 R) X
V (T, X) = 0
V (t, X) = s
_
T
t
D
_
t, t
/
_
e
a(t,t
/
,1)+b(t,t
/
,1)X
dt
/
(1 R)
_
T
t
D
_
t, t
/
_
d
_
e
a(t,t
/
,1)+b(t,t
/
,1)X
_
where D (t, t
/
/
.
CDS valuation 2
Accordingly,
U = s
_
T
0
D
_
0, t
/
_ _
1 p
_
0, t
/
__
dt
/
+ (1 R)
_
T
0
D
_
0, t
/
_
dp
_
0, t
/
_
For a given up-front payment , we can represent the corresponding
par spread s (i.e. the spread which makes the value of the
corresponding CDS zero) as follows:
s (T) =
+ (1 R)
_
T
0
D (0, t
/
) dp (0, t
/
)
_
T
0
D (0, t
/
) (1 p (0, t
/
)) dt
/
Conversely, for a given spread we can represent the par up-front
payment in the form
= s (T)
_
T
0
D
_
0, t
/
_ _
1 p
_
0, t
/
__
dt
/
+ (1 R)
_
T
0
D
_
0, t
/
_
dp
_
0, t
/
_
In these formulas we implicitly assumed that the corresponding CDS
is fully collateralized, so that in the event of default 1 R is readily
available.
CDS valuation 2
Accordingly,
U = s
_
T
0
D
_
0, t
/
_ _
1 p
_
0, t
/
__
dt
/
+ (1 R)
_
T
0
D
_
0, t
/
_
dp
_
0, t
/
_
s (T) =
+ (1 R)
_
T
0
D (0, t
/
) dp (0, t
/
)
_
T
0
D (0, t
/
) (1 p (0, t
/
)) dt
/
payment in the form
= s (T)
_
T
0
D
_
0, t
/
_ _
1 p
_
0, t
/
__
dt
/
+ (1 R)
_
T
0
D
_
0, t
/
_
dp
_
0, t
/
_
available.
CDS valuation 2
Accordingly,
U = s
_
T
0
D
_
0, t
/
_ _
1 p
_
0, t
/
__
dt
/
+ (1 R)
_
T
0
D
_
0, t
/
_
dp
_
0, t
/
_
s (T) =
+ (1 R)
_
T
0
D (0, t
/
) dp (0, t
/
)
_
T
0
D (0, t
/
) (1 p (0, t
/
)) dt
/
payment in the form
= s (T)
_
T
0
D
_
0, t
/
_ _
1 p
_
0, t
/
__
dt
/
+ (1 R)
_
T
0
D
_
0, t
/
_
dp
_
0, t
/
_
available.
Two name correlation 1
It is very tempting to extend the above framework to cover several
correlated names. For example, consider two credits, A, B and assume
for simplicity that their default intensities coincide,
X
A
(t) = X
B
(t) = X (t)
and both names have the same recovery R
A
= R
B
= R.
For a given maturity T, the default event correlation is dened as
follows
(0, T) =
p
AB
(0, T) p
A
(0, T) p
B
(0, T)
_
p
A
(0, T) (1 p
A
(0, T)) p
B
(0, T) (1 p
B
(0, T))
where
A
,
B
are the default times, and
p
A
(0, T) = P (
A
_ T) , p
B
(0, T) = P (
B
_ T)
p
AB
(0, T) = P (
A
_ T,
B
_ T)
It is very tempting to extend the above framework to cover several
correlated names. For example, consider two credits, A, B and assume
for simplicity that their default intensities coincide,
X
A
(t) = X
B
(t) = X (t)
and both names have the same recovery R
A
= R
B
= R.
For a given maturity T, the default event correlation is dened as
follows
(0, T) =
p
AB
(0, T) p
A
(0, T) p
B
(0, T)
_
p
A
(0, T) (1 p
A
(0, T)) p
B
(0, T) (1 p
B
(0, T))
where
A
,
B
are the default times, and
p
A
(0, T) = P (
A
_ T) , p
B
(0, T) = P (
B
_ T)
p
AB
(0, T) = P (
A
_ T,
B
_ T)
It is clear that
p
A
(0, T) = p
B
(0, T) = p (0, T) = 1 e
a(0,T,1)+b(0,T,1)X
0
Simple calculation yields
p
AB
(0, T) = E
0
_
e
_
T
0
(X
A
(t
/
)+X
B
(t
/
))dt
/
_
+p
A
(0, T) +p
B
(0, T) 1
= E
0
_
e
2
_
T
0
X(t
/
)dt
/
_
+ 2p (0, T) 1
Thus
(0, T) =
e
a(0,T,2)+b(0,T,2)X
0
e
2a(0,T,1)+2b(0,T,1)X
0
_
1 e
a(0,T,1)+b(0,T,1)X
0
_
e
a(0,T,1)+b(0,T,1)X
0
In the absence of jumps, the corresponding event correlation is very
low. If large positive jumps are added (while overall survival
probability is preserved), then correlation can increase all the way to
one. Assuming that T = 5y, = 0.5, = 7%, and J = 5.0, we
illustrate this observation below.
It is clear that
p
A
(0, T) = p
B
(0, T) = p (0, T) = 1 e
a(0,T,1)+b(0,T,1)X
0
p
AB
(0, T) = E
0
_
e
_
T
0
(X
A
(t
/
)+X
B
(t
/
))dt
/
_
+p
A
(0, T) +p
B
(0, T) 1
= E
0
_
e
2
_
T
0
X(t
/
)dt
/
_
+ 2p (0, T) 1
Thus
(0, T) =
e
a(0,T,2)+b(0,T,2)X
0
e
2a(0,T,1)+2b(0,T,1)X
0
_
1 e
a(0,T,1)+b(0,T,1)X
0
_
e
a(0,T,1)+b(0,T,1)X
0
It is clear that
p
A
(0, T) = p
B
(0, T) = p (0, T) = 1 e
a(0,T,1)+b(0,T,1)X
0
p
AB
(0, T) = E
0
_
e
_
T
0
(X
A
(t
/
)+X
B
(t
/
))dt
/
_
+p
A
(0, T) +p
B
(0, T) 1
= E
0
_
e
2
_
T
0
X(t
/
)dt
/
_
+ 2p (0, T) 1
Thus
(0, T) =
e
a(0,T,2)+b(0,T,2)X
0
e
2a(0,T,1)+2b(0,T,1)X
0
_
1 e
a(0,T,1)+b(0,T,1)X
0
_
e
a(0,T,1)+b(0,T,1)X
0
It is clear that
p
A
(0, T) = p
B
(0, T) = p (0, T) = 1 e
a(0,T,1)+b(0,T,1)X
0
p
AB
(0, T) = E
0
_
e
_
T
0
(X
A
(t
/
)+X
B
(t
/
))dt
/
_
+p
A
(0, T) +p
B
(0, T) 1
= E
0
_
e
2
_
T
0
X(t
/
)dt
/
_
+ 2p (0, T) 1
Thus
(0, T) =
e
a(0,T,2)+b(0,T,2)X
0
e
2a(0,T,1)+2b(0,T,1)X
0
_
1 e
a(0,T,1)+b(0,T,1)X
0
_
e
a(0,T,1)+b(0,T,1)X
0
Event correlation Figure
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
0.00% 0.50% 1.00% 1.50% 2.00% 2.50% 3.00%
lambda
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
3.50%
corr (left axis) theta (right axis)
Figure: Correlation and mean-reversion level = X
0
as functions of jump
intensity . Other parameters are as follows: T = 5y, = 0.5, = 7%, and
J = 5.0.
FTD-STD
In the two-name portfolio, we can dene two types of CDSs which
depend on the correlation: (A) the rst-to-default swap (FTD); (B)
the second-to-default swap (STD). The corresponding par spreads
(assuming that there are no up-front payments) are
s
1
(T) =
(1 R)
_
T
0
D (0, t
/
) d
_
1 e
a(0,t
/
,2)+b(0,t
/
,2)X
0
_
_
T
0
D (0, t
/
) e
a(0,t
/
,2)+b(0,t
/
,2)X
0
dt
/
s
2
(T) =
(1 R)
_
T
0
D (0, t
/
) d
_
1
_
2e
a(t
/
,1)+b(t
/
,1)X
0
e
a(t
/
,2)+b(t
/
,2)X
0
__
_
T
0
D (0, t
/
)
_
2e
a(t
/
,1)+b(t
/
,1)X
0
e
a(t
/
,2)+b(t
/
,2)X
0
_
dt
/
It is clear that the relative values of s
1
, s
2
very strongly depend on
whether or not jumps are present in the model.
FTD-STD
In the two-name portfolio, we can dene two types of CDSs which
depend on the correlation: (A) the rst-to-default swap (FTD); (B)
the second-to-default swap (STD). The corresponding par spreads
(assuming that there are no up-front payments) are
s
1
(T) =
(1 R)
_
T
0
D (0, t
/
) d
_
1 e
a(0,t
/
,2)+b(0,t
/
,2)X
0
_
_
T
0
D (0, t
/
) e
a(0,t
/
,2)+b(0,t
/
,2)X
0
dt
/
s
2
(T) =
(1 R)
_
T
0
D (0, t
/
) d
_
1
_
2e
a(t
/
,1)+b(t
/
,1)X
0
e
a(t
/
,2)+b(t
/
,2)X
0
__
_
T
0
D (0, t
/
)
_
2e
a(t
/
,1)+b(t
/
,1)X
0
e
a(t
/
,2)+b(t
/
,2)X
0
_
dt
/
It is clear that the relative values of s
1
, s
2
very strongly depend on
whether or not jumps are present in the model.
FTD-STD Figure
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
3.50%
4.00%
0.00% 0.50% 1.00% 1.50% 2.00% 2.50% 3.00%
lambda
s_1 s_2 s
Figure: FTD spread s
1
, STD spread s
2
, and single name CDS spread s as
functions of jump intensity . Other parameters are the same as in Fig.1. It is
clear that jumps are necessary to have s
1
and s
2
of similar magnitudes.
Counter-party eects
An important application of the above model is to the evaluation of
counter-party eects on fair CDS spreads. Let us assume that name
A has written a CDS on reference name B. It is clear that the pricing
problem for the value of the uncollateralized CDS

V can be written as
follows
/
V (t, X) (r + 2X)

V (t, X)
= s (1 R) X (RV
+
(t, X) +V
(t, X)) X
where V is the value of a fully collateralized CDS on name B.
It is clear that the discount rate is increased from r +X to r + 2X,
since there are two cases when the uncollateralized CDS can be
terminated due to default: when the reference name B defaults; when
the issuer A defaults.
Although this equation is no longer analytically solvable, it can be
solved numerically via, say, an appropriate modication of the
classical Crank-Nicholson method. It turns out that in the presence of
jumps the value of the fair par spread goes down dramatically.
Counter-party eects

V can be written as
follows
/
V (t, X) (r + 2X)

V (t, X)
= s (1 R) X (RV
+
(t, X) +V
(t, X)) X
Counter-party eects

V can be written as
follows
/
V (t, X) (r + 2X)

V (t, X)
= s (1 R) X (RV
+
(t, X) +V
(t, X)) X
Counter-party Figure
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
0.00% 0.50% 1.00% 1.50% 2.00% 2.50% 3.00%
lambda
Fraction of Default Leg Value
Figure: Reduction in default leg value as function of jump intensity .
Mean-reversion level = X
0
is chosen in order to preserve survival probability.
Other parameters are as follows: T = 5y, = 0.5, = 7%, and J = 5.0.
Typical structural model without jumps formulation
A typical structural model without jumps for the evolution of the
log-value of the rm has the form:
dx = dt + dW (t) , x (0) = x
0
x (t) = ln
_
v (t)
B (t)
_
= ln
_
v (t)
B (0) exp ((r d) t)
_
=
2
2
This value is governed by a Wiener process. The rm defaults if the
value x(t) crosses zero barrier b(t) = 0.
Typical structural model without jumps formulation
A typical structural model without jumps for the evolution of the
dx = dt + dW (t) , x (0) = x
0
x (t) = ln
_
v (t)
B (t)
_
= ln
_
v (t)
B (0) exp ((r d) t)
_
=
2
2
This value is governed by a Wiener process. The rm defaults if the
value x(t) crosses zero barrier b(t) = 0.
Assuming that all the relevant parameters are constant, the
corresponding formula for Q (t, T) is
Q (t, T) = I (0, t)
_
N
_
x (t)
1
2
_
e
x(t)
N
_
x (t)
1
2
__
where N is the cumulative normal distribution. It is easy to verify that
Q (0, 0) = 1, Q
T
(0, 0) = 0
The latter fact is rather disconcerting, since it implies that the par
short term CDS spread vanishes when T 0.
Typical structural model with jumps formulation
A typical structural model with jumps for the evolution of the
dx = dt + dW (t) +j
+
dN
+
+j
dN
, x (0) = x
0
x = ln
_
v
B
_
=
2
2

+
+
1
+

+ 1
This value is governed by a combination of a Wiener process and a
Poisson process with exponentially distributed jumps. The rm
defaults if the value x(t) crosses zero barrier b(t) = 0. It was realized
early on that without jumps (or/and curvilinear or uncertain barriers)
it is impossible to explain the short end of the CDS curve within the
structural framework. When all the relevant parameters are constant,
the problem can be solved analytically via the Laplace transform.
However, in general this approach does not work.
Typical structural model with jumps formulation
A typical structural model with jumps for the evolution of the
dx = dt + dW (t) +j
+
dN
+
+j
dN
, x (0) = x
0
x = ln
_
v
B
_
=
2
2

+
+
1
+

+ 1
This value is governed by a combination of a Wiener process and a
Poisson process with exponentially distributed jumps. The rm
defaults if the value x(t) crosses zero barrier b(t) = 0. It was realized
early on that without jumps (or/and curvilinear or uncertain barriers)
it is impossible to explain the short end of the CDS curve within the
structural framework. When all the relevant parameters are constant,
the problem can be solved analytically via the Laplace transform.
However, in general this approach does not work.
Unbounded PDF 1
When the barrier is absent, can be found via a simple recursion
(knowing an analytical solution is useful for benchmarking purposes).
Vis-a-vis the Gaussian distribution, has fat tails and a narrow peak.
Let
_
t
= (x t) /
_
t
0,0
(x) = n () /
_
t
1,0
(x) =
+
P(,
+
)
0,1
(x) =
P(,
)
where P(a, b) = exp
_
ab +b
2
/2
_
N(a +b)
Unbounded PDF 2
Then:
k,0
(x) =

+
((+
+
)
k1,0
+
+
k2,0
)
k 1
0,l
(x) =

((
)
0,l 1
+
0,l 2
)
l 1
k,l
(x) =

+
k1,l
+
k,l 1
+
+
w
k,l
=
e
(
+
+
)t
(
+
t)
k
(
t)
l
k!l !
(x) =

k=0,l =0
w
k,l
k,l
(x)
No barrier PDF gure
sigma=0.2, lambda=0.1, alpha=1.0, mu=0.0
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
-10.00 -8.00 -6.00 -4.00 -2.00 0.00 2.00 4.00
X
P
D
F
Fokker-Planck equation
We can solve the barrier problem by using the forward Fokker-Planck
equation for the t.p.d.f. and putting probabilities below the barrier to
zero. This equation has the form:
/
=
t

1
2
xx
+
x
+
_
0
(t, x +j ) e
j
dj = 0
(0, x) = (x )
(t, x) = 0 if < 0
When parameters are time-independent, we can nd via the
Laplace transform.
Fokker-Planck equation
We can solve the barrier problem by using the forward Fokker-Planck
equation for the t.p.d.f. and putting probabilities below the barrier to
zero. This equation has the form:
/
=
t

1
2
xx
+
x
+
_
0
(t, x +j ) e
j
dj = 0
(0, x) = (x )
(t, x) = 0 if < 0
When parameters are time-independent, we can nd via the
Laplace transform.
Laplace transform
The Laplace transform
1
2
xx
x
( +p)

+
_

0

(p, x +j ) e
j
dj = (x )
(p, 0) = 0
Forward characteristic equation:
1
2
2
( +p)

= 0
This equation has three roots of which two are positive and one
negative. We denote them as
i
.
Laplace transform
1
2
xx
x
( +p)

+
_

0

(p, x +j ) e
j
dj = (x )
(p, 0) = 0
1
2
2
( +p)

= 0
i
.
Laplace transform
1
2
xx
x
( +p)

+
_

0

(p, x +j ) e
j
dj = (x )
(p, 0) = 0
1
2
2
( +p)

= 0
i
.
Greens function construction 1
Hence the overall solution has the form:
(p, x) =
_
C
3
e
3
(x)
; x _
D
1
e
1
(x)
+D
2
e
2
(x)
+D
3
e
3
(x)
; x _
Matching conditions
D
1
+D
2
+ (D
3
C
3
) = 0
1
D
1
+
2
D
2
+
3
(D
3
C
3
) =
2
2
D
1
e
+D
2
e
+D
3
e
= 0
One more condition is obtained from the following observations:
/
_
e
i
(x)
H (x )
_
=

+
i
e
(x)
H ( x)
/
_
e
i
(x)
H ( x)
_
=

+
i
e
(x)
H ( x)
Here /
is the corresponding dierential operator and H (.) is the

Heaviside function.
(p, x) =
_
C
3
e
3
(x)
; x _
D
1
e
1
(x)
+D
2
e
2
(x)
+D
3
e
3
(x)
; x _
Matching conditions
D
1
+D
2
+ (D
3
C
3
) = 0
1
D
1
+
2
D
2
+
3
(D
3
C
3
) =
2
2
D
1
e
+D
2
e
+D
3
e
= 0
/
_
e
i
(x)
H (x )
_
=

+
i
e
(x)
H ( x)
/
_
e
i
(x)
H ( x)
_
=

+
i
e
(x)
H ( x)
Here /

Heaviside function.
(p, x) =
_
C
3
e
3
(x)
; x _
D
1
e
1
(x)
+D
2
e
2
(x)
+D
3
e
3
(x)
; x _
Matching conditions
D
1
+D
2
+ (D
3
C
3
) = 0
1
D
1
+
2
D
2
+
3
(D
3
C
3
) =
2
2
D
1
e
+D
2
e
+D
3
e
= 0
/
_
e
i
(x)
H (x )
_
=

+
i
e
(x)
H ( x)
/
_
e
i
(x)
H ( x)
_
=

+
i
e
(x)
H ( x)
Here /

Heaviside function.
(p, x) =
_
C
3
e
3
(x)
; x _
D
1
e
1
(x)
+D
2
e
2
(x)
+D
3
e
3
(x)
; x _
Matching conditions
D
1
+D
2
+ (D
3
C
3
) = 0
1
D
1
+
2
D
2
+
3
(D
3
C
3
) =
2
2
D
1
e
+D
2
e
+D
3
e
= 0
/
_
e
i
(x)
H (x )
_
=

+
i
e
(x)
H ( x)
/
_
e
i
(x)
H ( x)
_
=

+
i
e
(x)
H ( x)
Here /

Heaviside function.
The corresponding prole has the form
(p, x) =
_
C
3
e
3
(x)
; x _
D
1
e
_
e
1
x
e
3
x
+D
2
e
_
e
2
x
e
3
x
; x <
D
1
=
2
2
( +
1
)
(
1
2
) (
1
3
)
D
2
=
2
2
( +
2
)
(
2
1
) (
2
3
)
D
3
=
2
2
_
e
(
1
3
)
( +
1
)
(
1
2
) (
1
3
)
+
e
(
2
3
)
( +
2
)
(
2
1
) (
2
3
)
_
C
3
= D
1
+D
2
+D
3
Stehfest algorithm
Inverse Laplace transform yields (t, x). We use Stehfest algorithm
to evaluate Q:
(t, x) = p
N
k=1
(1)
k
St
N
k

(kp, x)
p =
ln 2
t
Coecients St
N
k
are very sti. We typically choose N = 20. For small
t inversion can be numerically unstable unless computation is carried
with many signicant digits.
-15
-10
-5
0
5
10
15
20
25
30
35
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Stehfest algorithm
Inverse Laplace transform yields (t, x). We use Stehfest algorithm
to evaluate Q:
(t, x) = p
N
k=1
(1)
k
St
N
k

(kp, x)
p =
ln 2
t
Coecients St
N
k
are very sti. We typically choose N = 20. For small
t inversion can be numerically unstable unless computation is carried
with many signicant digits.
-15
-10
-5
0
5
10
15
20
25
30
35
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Hitting time density
The rst hitting time density f (t) has the form
f (t) =
2
2

x
(t, 0)
Its Laplace transform is given by
f (p) =
e
( +
2
) e
( +
1
)
1
For the regular diusion we obtain
f (p) = e
f (t) =
2
2

x
(t, 0) =
e
(+t)
2
/2
2
t
_
2
2
t
3
f (t) =
2
2

x
(t, 0)
f (p) =
e
( +
2
) e
( +
1
)
1
f (p) = e
f (t) =
2
2

x
(t, 0) =
e
(+t)
2
/2
2
t
_
2
2
t
3
f (t) =
2
2

x
(t, 0)
f (p) =
e
( +
2
) e
( +
1
)
1
f (p) = e
f (t) =
2
2

x
(t, 0) =
e
(+t)
2
/2
2
t
_
2
2
t
3
Hitting time density gure
0.000
0.002
0.004
0.006
0.008
0.010
0.012
0.014
0.016
0.018
0.020
0.00 5.00 10.00 15.00 20.00 25.00
Time
P
r
o
b
Jump-Diffusion Inversion Regular Diffusion Inversion Regular Diffusion Analytical
Greens function expansion
Perturbative calculation of Greens function in 1D. For brevity we
assume that all the parameters are time independent.
Time-dependent case can be solved in a similar way. To start with,
we make the following transform
(t, x) = exp
_
_

2
2
2
+
_
t +

2
(x )
_

(t, x)
The modied Greens function solves the following propagation
problem
t

(t, x)
1
2
2

(t, x)
_

0

(t, x +j ) e
j
dj = 0
(0, x) = (x )

(t, 0) = 0,

(t, ) = 0
where =
_

2
_
, =
1
2
2
+

+1
We assume that 1 and represent

as follows
(t, x) =

(0)
(t, x) +

(1)
(t, x) + ...
(t, x) = exp
_
_

2
2
2
+
_
t +

2
(x )
_

(t, x)
problem
t

(t, x)
1
2
2

(t, x)
_

0

(t, x +j ) e
j
dj = 0
(0, x) = (x )

(t, 0) = 0,

(t, ) = 0
where =
_

2
_
, =
1
2
2
+

+1

as follows
(t, x) =

(0)
(t, x) +

(1)
(t, x) + ...
(t, x) = exp
_
_

2
2
2
+
_
t +

2
(x )
_

(t, x)
problem
t

(t, x)
1
2
2

(t, x)
_

0

(t, x +j ) e
j
dj = 0
(0, x) = (x )

(t, 0) = 0,

(t, ) = 0
where =
_

2
_
, =
1
2
2
+

+1

as follows
(t, x) =

(0)
(t, x) +

(1)
(t, x) + ...
Zero order term
A simple calculation yields (with =
2
t)
(0)
(t, x) =
1
_
_
n
_
x
_
_
n
_
x +
_
__
First order rhs
H
(1)
(t, x) = P
_
x
_
,
_
_
P
_
x +
_
,
_
_
We use Duhamels principle and represent

(1)
as follows
(1)
(t, x) =
_
t
0
_

0

(0)
(t, x; s,
1
) H
(1)
(s,
1
) dsd
1
Zero order term
2
t)
(0)
(t, x) =
1
_
_
n
_
x
_
_
n
_
x +
_
__
First order rhs
H
(1)
(t, x) = P
_
x
_
,
_
_
P
_
x +
_
,
_
_

(1)
as follows
(1)
(t, x) =
_
t
0
_

0

(0)
(t, x; s,
1
) H
(1)
(s,
1
) dsd
1
Zero order term
2
t)
(0)
(t, x) =
1
_
_
n
_
x
_
_
n
_
x +
_
__
First order rhs
H
(1)
(t, x) = P
_
x
_
,
_
_
P
_
x +
_
,
_
_

(1)
as follows
(1)
(t, x) =
_
t
0
_

0

(0)
(t, x; s,
1
) H
(1)
(s,
1
) dsd
1
First order term
A very elaborate calculation yields
(1)
(t, x) =

2
1
_
P
_
x
_
,
_
_
+xP
_
x +
_
,
_
_
(x ) P
_
x +
_
,
_
_
(x + ) e

P
_
x
_
,
_
_
+ (x ) e

P
_
x
_
,
_
__
The formula can be independently checked via the method of heat
potentials.
First order term
A very elaborate calculation yields
(1)
(t, x) =

2
1
_
P
_
x
_
,
_
_
+xP
_
x +
_
,
_
_
(x ) P
_
x +
_
,
_
_
(x + ) e

P
_
x
_
,
_
_
+ (x ) e

P
_
x
_
,
_
__
The formula can be independently checked via the method of heat
potentials.
Quality of approximation gure
pdf vs. x
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.000 1.000 2.000 3.000 4.000 5.000 6.000
x
pdf_ana pdf_exp_0 pdf_exp_1 pdf_num
Figure: G (t, x) for t = 20, = 20%, = 2%, = 4, = 1.
Barrier PDF gure
sigma=0.2, lambda=0.1, alpha=1.0, mu=0.0, barrier=-2.0
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
-5.00 -4.00 -3.00 -2.00 -1.00 0.00 1.00 2.00 3.00 4.00 5.00
X
P
r
o
b
Series PDE Laplace
sigma=0.2,
lambda=0.1,
alpha=1.0,
mu=0.0
Hybrid model
Modeling of the value of the company is not always the best way of
dealing with default. It is sometimes thought that modeling the stock
price instead is a better choice.
A typical model of this kind can be written as
dS = r (t) Sdt + (t, S) dW (t)
where (t, 0) ,= 0. From that angle, one can argue that the Bachelier
model is actually better than Black-Scholes!!
Alternatively, one can add a terminal (Samuelson-style) jump to
default and model S as follows
dS = [r (t) + (S)] Sdt + (t, S) dW (t) SdN (t)
Hybrid model
Hybrid model
Model Comparison
Comparison of numerical and analytical t.p.d.f. as well as the solution
of the barrier problem with non-constant barrier is given below.
By bootstrapping the barrier, we can reproduce term structure of
CDS for most names. In addition, we can price equity derivatives and
produce a respectable volatility skew.
The barrier can assumed to be stochastic as in Due and Lando
(2001) or CreditGrades (2000). Reduction of the information set
makes reduced form and structural modeling almost identical, Jarrow,
Protter, Yildirim (2004).
Model Comparison
Model Comparison
Survival probability gure
55%
70%
85%
100%
0 1 2 3 4 5 6 7 8 9 10
RF MBC LFKY L AL
Figure: Typical survival probabilities Q (0, T) as functions of T (in years) for the
reduced-form, Merton-Black-Cox, Lardy et al., Lipton, and Atlan-Lelanc models.
Notice that Q
MBC
(0, T) and Q
AL
(0, T) are too at when T 0, while
Q
LFHY
(0, 0) < 1. Q
RF
(0, T) and Q
L
(0, T) behave properly.
Greens function without jumps
Without jumps, we need to nd Greens function for the correlated heat
equation in the quarter-plane, i.e., to solve the following problem
t

1
2
_
2
1
x
1
x
1
+ 2
1
x
1
x
2
+
2
2
x
2
x
2
_
+
1
x
1
+
2
x
2
= 0
(t, x
1
, 0) = 0, (t, 0, x
2
) = 0
(t, x
1
, x
2
) (x
1
1
) (x
2
2
)
Standard transform
= exp (t +
1
(x
1
1
) +
2
(x
2
2
))

removes drift terms provided that

= , =
1
2
1

Transforms
A sequence of linear transforms allows us to rewrite the pricing problem as
follows
(2)
t

1
2
_

(2)
x
(2)
1
x
(2)
1
+

(2)
x
(2)
2
x
(2)
2
_
= 0
w
_
t, x
(2)
1
, 0
_
= 0, w
_
t, x
(2)
2
/ , x
(2)
2
_
= 0
w
_
t, x
(2)
1
, x
(2)
2
_

_
x
(2)
1

(2)
1
_
_
x
(2)
2

(2)
2
_
where =
_
1
2
. We now have the standard heat equation in an angle
and can use our previous results. This angle is formed by the horizontal
axis
(2)
2
= 0 and a sloping line
(2)
1
=
(2)
2
/ . It is acute when < 0
and blunt otherwise. The size of this angle is =atan( /).
Eigenfunction expansion
The solution of the above problem via the method of images was
independently introduced in nance by He et al., Zhou and Lipton. We
want to nd the Greens function for the following parabolic equation
(2)
t

1
2
_

(2)
rr
+
1
r

(2)
r
+
1
r
2

(2)
_
supplied with the boundary conditions of the form
(2)
(t, r , )
r 0
C < ,

(2)
(t, r , )
r
0,

(2)
(t, r , 0) = 0,

(2)
(t, r , ) = 0
and the initial condition
(2)
(t, r , )
t0
(r r
/
) (
/
)
r
/
The fundamental solution has the form
_
t, r , [0, r
/
,
/
_
=
2e
(
r
2
+r
/2
)
/2t
t
n=1
I
n/
_
rr
/
t
_
sin
_
n
/
_
sin
_
n
_
Non-periodic Greens function
It turns out that we can nd the fundamental solution via the method of
images too. This approach seems to be new. To start with, we nd
non-periodic solution of the heat equation written in polar co-ordinates in
the positive half-plane 0 < r < , < < :
_
t, r , [0, r
/
,
/
_
=
1
_
t, r , [0, r
/
,
/
_
2
_
t, r , [0, r
/
,
/
_
where =
/
, s
=sign( ),
1
_
t, r , [0, r
/
,
/
_
=
e
(
r
2
+r
/2
)
/2t
2t
(s
+
+s
)
2
e
(rr
/
/t) cos
2
_
t, r , [0, r
/
,
/
_
=
e
(
r
2
+r
/2
)
/2t
2
2
t
_
0
_
s
+
e
(rr
/
/t) cosh((+))
+s
e
(rr
/
/t) cosh(())
_
2
+ 1
d
Graph of Greens function
0.000
0.005
0.010
0.015
0.020
0.025
0.030
-20.00 -15.00 -10.00 -5.00 0.00 5.00 10.00 15.00 20.00
0.05 0.6 1.1 1.6 2.1 2.6 3.1 3.6 4.1 4.6
Figure: Cross-section of non-periodic Greens function. Parameters are as follows
t = 5, r
/
= 1.5,
/
= 0.
Method of images
Simple balancing of terms shows that
2
_
t, r , [0, r
/
,
/
_
= O
_
2
_
Next, we represent the fundamental solution in the form
_
t, r , [0, r
/
,
/
_
=

n=
_
_
t, r , [0, r
/
,
/
+ 2n
_
_
t, r , [0, r
/
,
/
+ 2n
_
Indeed, it is clear that the sum converges, every term solves the parabolic
equation, only one term has a pole inside the angle, and
(0) =
() = 0.
Correlated PDF gure
0
.
0
0
0
.
5
0
1
.
0
0
1
.
5
0
2
.
0
0
2
.
5
0
3
.
0
0
3
.
5
0
4
.
0
0
4
.
5
0
5
.
0
0
5
.
5
0
6
.
0
0
6
.
5
0
7
.
0
0
7
.
5
0
8
.
0
0
8
.
5
0
9
.
0
0
9
.
5
0
1
0
.
0
0
0.00
1.20
2.40
3.60
4.80
6.00
7.20
8.40
9.60
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
Figure: PDF in the positive quarter-plane
Greens function with jumps
With jumps, we need to nd Greens function for the correlated
jump-diusion equation in the quarter-plane. In the simplest case we have
to solve the following problem
t

1
2
_
2
1
x
1
x
1
+ 2
1
x
1
x
2
+
2
2
x
2
x
2
_
+
1
x
1
+
2
x
2
_
0
_
0
(x
1
+j
1
, x
2
+j
2
) e
1
j
1
2
j
2
dj
1
dj
2
+ = 0
(t, x
1
, 0) = 0, (t, 0, x
2
) = 0
(t, x
1
, x
2
) (x
1
1
) (x
2
2
)
This problem has been recently solved (numerically) by Lipton and Sepp.
Its analytical solution is not known (??).
Multi-name case
Consider N companies and assume that their asset values are driven
by the following SDEs
dV
i
(t)
V
i
(t)
= (r d
i

i
i
(t)) dt +
i
(t) dW
i
(t) +
_
e
J
i
1
_
dN
i
(t) V
i
(0) = v
i
i
= E
_
e
J
i
1
_
Default boundaries are driven by the ODEs of the form
dB
i
(t)
B
i
(t)
= (r d
i
) dt B
i
(0) = b
i
In log coordinates
x
i
(t) = ln
_
V
i
(t)
B
i
(t)
_
dx
i
(t) =
i
(t) dt +
i
(t) dW
i
(t) +J
i
dN
i
(t) x
i
(0) = ln
_
v
i
b
i
_
=
i
i
=
1
2
2
i

i
i
The corresponding default boundaries are now at
x
i
= 0
Multi-name case
dV
i
(t)
V
i
(t)
= (r d
i

i
i
(t)) dt +
i
(t) dW
i
(t) +
_
e
J
i
1
_
dN
i
(t) V
i
(0) = v
i
i
= E
_
e
J
i
1
_
dB
i
(t)
B
i
(t)
= (r d
i
) dt B
i
(0) = b
i
In log coordinates
x
i
(t) = ln
_
V
i
(t)
B
i
(t)
_
dx
i
(t) =
i
(t) dt +
i
(t) dW
i
(t) +J
i
dN
i
(t) x
i
(0) = ln
_
v
i
b
i
_
=
i
i
=
1
2
2
i

i
i
x
i
= 0
Multi-name case
dV
i
(t)
V
i
(t)
= (r d
i

i
i
(t)) dt +
i
(t) dW
i
(t) +
_
e
J
i
1
_
dN
i
(t) V
i
(0) = v
i
i
= E
_
e
J
i
1
_
dB
i
(t)
B
i
(t)
= (r d
i
) dt B
i
(0) = b
i
In log coordinates
x
i
(t) = ln
_
V
i
(t)
B
i
(t)
_
dx
i
(t) =
i
(t) dt +
i
(t) dW
i
(t) +J
i
dN
i
(t) x
i
(0) = ln
_
v
i
b
i
_
=
i
i
=
1
2
2
i

i
i
x
i
= 0
Multi-name case
dV
i
(t)
V
i
(t)
= (r d
i

i
i
(t)) dt +
i
(t) dW
i
(t) +
_
e
J
i
1
_
dN
i
(t) V
i
(0) = v
i
i
= E
_
e
J
i
1
_
dB
i
(t)
B
i
(t)
= (r d
i
) dt B
i
(0) = b
i
In log coordinates
x
i
(t) = ln
_
V
i
(t)
B
i
(t)
_
dx
i
(t) =
i
(t) dt +
i
(t) dW
i
(t) +J
i
dN
i
(t) x
i
(0) = ln
_
v
i
b
i
_
=
i
i
=
1
2
2
i

i
i
x
i
= 0
Correlation structure
We introduce correlation between diusions in the usual way and
assume that
dW
i
(t) dW
j
(t) =
ij
(t) dt
We introduce correlation between jumps following the Marshall-Olkin
idea. Let
(N)
be the set of all subsets of N names except for the
empty subset ?, and its typical member. With every we
associate a Poisson process N
(t) with intensity
(t), and
represent N
i
(t) as follows
N
i
(t) =

(N)
1
i
N
(t)
i
(t) =

(N)
1
i
(t)
Thus, we assume that there are both common and idiosyncratic jump
sources.
assume that
dW
i
(t) dW
j
(t) =
ij
(t) dt
idea. Let
(N)
(t) with intensity
(t), and
represent N
i
(t) as follows
N
i
(t) =

(N)
1
i
N
(t)
i
(t) =

(N)
1
i
(t)
sources.
assume that
dW
i
(t) dW
j
(t) =
ij
(t) dt
idea. Let
(N)
(t) with intensity
(t), and
represent N
i
(t) as follows
N
i
(t) =

(N)
1
i
N
(t)
i
(t) =

(N)
1
i
(t)
sources.
Multi-name pricing problem
We now formulate a typical pricing equation in the positive cone
R
(N)
+
. We have
t
U (t,~x) +/
(N)
U (t,~x) = (t,~x)
U (t,~x
0,k
) =
0,k
(t,~y) , U (t,~x
,k
) =
,k
(t,~y)
U (T,~x) = (~x)
where ~x
0,k
, ~x
,k
, ~y
k
are N and N 1 dimensional vectors,
respectively,
~x
0,k
=
_
x
1
, ..., 0
k
, ...x
N
_
, ~x
,k
=
_
x
1
, ...,
k
, ...x
N
_
, ~y
k
= (x
1
, ..., x
N
)
The integro-dierential operator /
(N)
can be written as
/
(N)
f (~x) =
1
2
i

2
i

2
i
f (~x) +

i ,j ,j >i
a
ij
j
f (~x) +
i

i
i
f (~x)
f (~x) +

(N)
i
f (~x)
Here a
ij
=
i
ij
, = +
(N)

.
R
(N)
+
. We have
t
U (t,~x) +/
(N)
U (t,~x) = (t,~x)
U (t,~x
0,k
) =
0,k
(t,~y) , U (t,~x
,k
) =
,k
(t,~y)
U (T,~x) = (~x)
where ~x
0,k
, ~x
,k
, ~y
k
respectively,
~x
0,k
=
_
x
1
, ..., 0
k
, ...x
N
_
, ~x
,k
=
_
x
1
, ...,
k
, ...x
N
_
, ~y
k
= (x
1
, ..., x
N
)
(N)
can be written as
/
(N)
f (~x) =
1
2
i

2
i

2
i
f (~x) +

i ,j ,j >i
a
ij
j
f (~x) +
i

i
i
f (~x)
f (~x) +

(N)
i
f (~x)
Here a
ij
=
i
ij
, = +
(N)

.
R
(N)
+
. We have
t
U (t,~x) +/
(N)
U (t,~x) = (t,~x)
U (t,~x
0,k
) =
0,k
(t,~y) , U (t,~x
,k
) =
,k
(t,~y)
U (T,~x) = (~x)
where ~x
0,k
, ~x
,k
, ~y
k
respectively,
~x
0,k
=
_
x
1
, ..., 0
k
, ...x
N
_
, ~x
,k
=
_
x
1
, ...,
k
, ...x
N
_
, ~y
k
= (x
1
, ..., x
N
)
(N)
can be written as
/
(N)
f (~x) =
1
2
i

2
i

2
i
f (~x) +

i ,j ,j >i
a
ij
j
f (~x) +
i

i
i
f (~x)
f (~x) +

(N)
i
f (~x)
Here a
ij
=
i
ij
, = +
(N)

.
Jump terms
In the case of negative exponential jumps,
J
i
f (~x) =
_
0
x
i
f (x
1
, ..., x
i
+j , ...x
N
)
i
(j ) dj
while in the case of discrete negative jumps
J
i
f (~x) = H(x
i
+J
i
) f (x
1
, ..., x
i
+J
i
, ...x
N
)
and H is the Heaviside function.
We naturally split the operator /
(N)
into the local (dierential) and
non-local (integral) parts:
/
(N)
f (~x) = T
(N)
f (~x) +J
(N)
f (~x)
Jump terms
In the case of negative exponential jumps,
J
i
f (~x) =
_
0
x
i
f (x
1
, ..., x
i
+j , ...x
N
)
i
(j ) dj
while in the case of discrete negative jumps
J
i
f (~x) = H(x
i
+J
i
) f (x
1
, ..., x
i
+J
i
, ...x
N
)
and H is the Heaviside function.
We naturally split the operator /
(N)
into the local (dierential) and
non-local (integral) parts:
/
(N)
f (~x) = T
(N)
f (~x) +J
(N)
f (~x)
Adjoint operator
The corresponding adjoint operator is
/
(N)
g (~x) =
1
2
i

2
i

2
i
g (~x) +

i ,j ,j >i
a
ij
j
g (~x)
i

i
i
g (~x)
g (~x) +

(N)
i
g (~x)
where
J
i
g (~x) =
_

0
g (x
1
, ..., x
i
j , ...x
N
)
i
(j ) dj
J
i
g (~x) = g (x
1
, ..., x
i
J
i
, ...x
N
)
Greens formula
We introduce Greens function (t,~x), or, more explicitly,
_
t
/
,~x; t,
~
_
, such that
t
/
_
t
/
,~x
_
/
(N)
_
t
/
,~x
_
= 0
_
t
/
,~x
0k
_
= 0,
_
t
/
,~x
k
_
= 0, (t,~x) =
_
~x
~
_
It can be shown that
U
_
t,
~
_
=
_
R
(N)
+
(~x)
_
T,~x; t,
~
_
d~x
+
k
_
T
t
_
R
(N1)
+
k
_
t
/
,~y
_
_
1
2
2
k
k
_
t
/
,~y; t,
~
_
_
dt
/
d~y
_
T
t
_
R
(N)
+
_
t
/
,~x
_
_
t
/
,~x; t,
~
_
dt
/
d~x
In other words, instead of solving the backward pricing problem with
nonhomogeneous rhs and boundary conditions, we can solve the
forward propagation problem for Greens function with homogeneous
rhs and boundary conditions.
Greens formula
We introduce Greens function (t,~x), or, more explicitly,
_
t
/
,~x; t,
~
_
, such that
t
/
_
t
/
,~x
_
/
(N)
_
t
/
,~x
_
= 0
_
t
/
,~x
0k
_
= 0,
_
t
/
,~x
k
_
= 0, (t,~x) =
_
~x
~
_
It can be shown that
U
_
t,
~
_
=
_
R
(N)
+
(~x)
_
T,~x; t,
~
_
d~x
+
k
_
T
t
_
R
(N1)
+
k
_
t
/
,~y
_
_
1
2
2
k
k
_
t
/
,~y; t,
~
_
_
dt
/
d~y
_
T
t
_
R
(N)
+
_
t
/
,~x
_
_
t
/
,~x; t,
~
_
dt
/
d~x
In other words, instead of solving the backward pricing problem with
nonhomogeneous rhs and boundary conditions, we can solve the
forward propagation problem for Greens function with homogeneous
rhs and boundary conditions.
Single-name case
We now are in a position to describe the relevant operators in more
detail for N = 1, 2.
When N = 1 we deal with the reference name alone. Since
(1)
consists of just one element,
(1)
= 1
/
(1)
f =
1
2
2
1
2
1
f +
1
1
f f +
1
1
f = T
(1)
f +J
(1)
f
/
(1)
g =
1
2
2
1
2
1
g
1
1
g g +
1
1
g = T
(1)
g +J
(1)
g
= +
1
J
(1)
f (x
1
) =
1
_
0
x
1
f (x
1
+j
1
) e
1
j
1
d (
1
j
1
) =
1
_
x
1
0
f (y
1
) e
1
(x
1
y
1
)
d (
1
y
1
)
J
(1)
g (x
1
) =
1
_
0
g (x
1
j
1
) e
1
j
1
d (
1
j
1
) =
1
_

x
1
g (y
1
) e
1
(x
1
y
1
)
d (
1
y
1
)
J
(1)
f (x
1
) =
1
H(x
1
+J
1
) f (x
1
+J
1
)
J
(1)
f (x
1
) =
1
f (x
1
J
1
)
Single-name case
We now are in a position to describe the relevant operators in more
detail for N = 1, 2.
When N = 1 we deal with the reference name alone. Since
(1)
consists of just one element,
(1)
= 1
/
(1)
f =
1
2
2
1
2
1
f +
1
1
f f +
1
1
f = T
(1)
f +J
(1)
f
/
(1)
g =
1
2
2
1
2
1
g
1
1
g g +
1
1
g = T
(1)
g +J
(1)
g
= +
1
J
(1)
f (x
1
) =
1
_
0
x
1
f (x
1
+j
1
) e
1
j
1
d (
1
j
1
) =
1
_
x
1
0
f (y
1
) e
1
(x
1
y
1
)
d (
1
y
1
)
J
(1)
g (x
1
) =
1
_
0
g (x
1
j
1
) e
1
j
1
d (
1
j
1
) =
1
_

x
1
g (y
1
) e
1
(x
1
y
1
)
d (
1
y
1
)
J
(1)
f (x
1
) =
1
H(x
1
+J
1
) f (x
1
+J
1
)
J
(1)
f (x
1
) =
1
f (x
1
J
1
)
Two-name case 1
When N = 2 we deal with the reference name 1 and protection seller
2, say. Since
(2)
consists of three elements,
(2)
= 1 , 2 , 1, 2
we have
/
(2)
f =
1
2
2
1
2
1
f +
1
2
2
2
2
2
f +
1
12
2
f +
1
1
f +
2
1
f f
+
1
1
f +
2
2
f +
1,2
2
f
= T
(2)
f +J
(2)
f
/
(2)
g =
1
2
2
1
2
1
g +
1
2
2
2
2
2
g +
1
12
2
g
1
1
g
2
1
g g
+
1
1
g +
2
2
g +
1,2
2
g
= T
(2)
g +J
(2)
g
= +
1
+
2
+
1,2
Two-name case 2
Specically,
J
(2)
f (x
1
, x
2
) =
1
_
x
1
0
f (y
1
, x
2
) e
1
(x
1
y
1
)
d (
1
y
1
)
+
2
_
x
2
0
f (x
1
, y
2
) e
2
(x
2
y
2
)
d (
2
y
2
)
+
1,2
_
x
1
0
_
x
2
0
f (y
1
, y
2
) e
1
(x
1
y
1
)
2
(x
2
y
2
)
d (
1
y
1
) d (
2
y
2
)
J
(2)
g (x
1
, x
2
) =
1
_

x
1
g (y
1
, x
2
) e
1
(x
1
y
1
)
d (
1
y
1
)
+
2
_

x
2
g (x
1
, y
2
) e
2
(x
2
y
2
)
d (
2
y
2
)
+
1,2
_

x
1
_

x
2
g (y
1
, y
2
) e
1
(x
1
y
1
)
2
(x
2
y
2
)
d (
1
y
1
) d (
2
y
2
)
Non-risky CDS pricing problem
We are now in a position to formulate the pricing problems of interest.
We start with a CDS on a reference credit 1 which is sold by a
nonrisky seller to a nonrisky buyer. Assuming for simplicity that the
coupon is paid continuously at the rate s, and that the default
boundary is continuous, we write the pricing equation as follows
t
U
(1)
(t, x
1
) +/
(1)
U
(1)
(t, x
1
) = (t, x
1
)
U
(1)
(t, 0) =
0
(t) , U
(1)
(t, ) =
(t)
U
(1)
(T, x
1
) = 0
Here
(t, x
1
) = s
1
_
1

R
1
_
e
1
x
1

R
1
=

1
R
1
1
+ 1
0
(t) = (1 R
1
) ,
(t) = s
_
T
t
D
_
t, t
/
_
dt
/
Non-risky CDS pricing problem
We are now in a position to formulate the pricing problems of interest.
We start with a CDS on a reference credit 1 which is sold by a
nonrisky seller to a nonrisky buyer. Assuming for simplicity that the
coupon is paid continuously at the rate s, and that the default
boundary is continuous, we write the pricing equation as follows
t
U
(1)
(t, x
1
) +/
(1)
U
(1)
(t, x
1
) = (t, x
1
)
U
(1)
(t, 0) =
0
(t) , U
(1)
(t, ) =
(t)
U
(1)
(T, x
1
) = 0
Here
(t, x
1
) = s
1
_
1

R
1
_
e
1
x
1

R
1
=

1
R
1
1
+ 1
0
(t) = (1 R
1
) ,
(t) = s
_
T
t
D
_
t, t
/
_
dt
/
Risky seller CDS pricing problem 1
For a CDS on a reference credit 1 which is sold by a risky seller 2 to a
nonrisky buyer, we write the pricing equation as follows
t
U
(2)
(t, x
1
, x
2
) +/
(2)
U
(2)
(t, x
1
, x
2
) = (t, x
1
, x
2
)
U
(2)
(t, 0, x
2
) =
0,1
(t, x
2
) , U
(2)
(t, , x
2
) =
,1
(t, x
2
)
U
(2)
(t, x
1
, 0) =
0,2
(t, x
1
) , U
(2)
(t, x
1
, ) =
,2
(t, x
1
)
U
(2)
(T, x
1
, x
2
) = 0
Risky seller CDS pricing problem 2
Here
(t, x
1
, x
2
) = s
1
1
(t, x
1
, x
2
)
2
2
(t, x
1
, x
2
)
1,2
1,2
(t, x
1
, x
2
)
1
(t, x
1
, x
2
) =
_
1

R
1
_
e
1
x
1
2
(t, x
1
, x
2
) =

U
(1)
(t, x
1
) e
2
x
2
1,2
(t, x
1
, x
2
) =
_
1

R
1
_
e
1
x
1
_
1 e
2
x
2
_
+
_
0
x
1
U
(1)
(t, x
1
+j
1
) e
1
j
1
d (
1
j
1
) e
2
x
2
+
_
1

R
1
_

R
2
e
1
x
1
2
x
2
U
(1)
(t, x
1
) =

R
2
U
(1)
+
(t, x
1
) +U
(1)
(t, x
1
)
0,1
(t, x
2
) = 1 R
1
,
,1
(t, x
2
) = s
_
T
t
D
_
t, t
/
_
dt
/
0,2
(t, x
1
) = R
2
U
(1)
+
(t, x
1
) +U
(1)
(t, x
1
) ,
,2
(t, x
1
) = U
(1)
(t, x
1
)
Illustrations
s(0) L(0) R l (0) a(0)
{dis

{exp

JPM 39.6 508 40% 203 243 0.1780 0.1780 5.6167 1.63%
MS 29.31 534 40% 214 243 0.1286 0.1286 7.7759 2.41%
Illustrations
CDS Spread Survival Prob Default Leg Annuity Leg
T JPM MS JPM MS JPM MS JPM MS
1y 0.0055 0.0155 0.9909 0.9745 0.0055 0.0153 0.9954 0.9872
2y 0.0055 0.0155 0.9818 0.9496 0.0109 0.0302 1.9818 1.9492
3y 0.0055 0.0151 0.9729 0.9275 0.0163 0.0435 2.9591 2.8876
4y 0.0055 0.0146 0.9640 0.9076 0.0216 0.0555 3.9276 3.8050
5y 0.0059 0.0137 0.9518 0.8928 0.0289 0.0643 4.8858 4.7047
6y 0.0060 0.0126 0.9417 0.8829 0.0350 0.0703 5.8323 5.5922
7y 0.0060 0.0117 0.9323 0.8741 0.0406 0.0756 6.7693 6.4707
8y 0.0060 0.0113 0.9230 0.8613 0.0462 0.0832 7.6969 7.3388
9y 0.0060 0.0112 0.9138 0.8475 0.0517 0.0915 8.6153 8.1932
10y 0.0060 0.0110 0.9048 0.8339 0.0571 0.0997 9.5246 9.0338
Illustrations
JPM
0.00
0.01
0.02
0.03
0.04
1y 2y 3y 4y 5y 6y 7y 8y 9y 10y
T
$\lambda^{dis}$
$\lambda^{exp}$
Spread
MS
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
1y 2y 3y 4y 5y 6y 7y 8y 9y 10y
T
$\lambda^{dis}$
$\lambda^{exp}$
Spread
Figure: Calibrated intensity rates for JMP (lhs) and MS(rhs) in the models with
ENJs and DNJs, respectively.
Illustrations
JPM
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.00 0.10 0.20 0.30 0.40
X
1y
5y
10y
JPM
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.00 0.10 0.20 0.30 0.40
X
1y
5y
10y
Figure: PDF of the driver x(T) for JMP in the model with DNJs (lhs) and ENJs
(rhs) for T = 1y, 5y, and 10 y.
Illustrations
MS
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
0.00 0.10 0.20 0.30 0.40
X
1y
5y
10y
MS
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
0.00 0.10 0.20 0.30 0.40
X
1y
5y
10y
Figure: Same graphs as in Figure 3 but for MS.
Illustrations
JPM
0%
10%
20%
30%
40%
50%
60%
80% 100% 120% 140% 160% 180%
K/S
C
D
S

I
m
p
l
i
e
d

V
o
l
$\lambda^{dis}$
$\lambda^{exp}$
MS
0%
10%
20%
30%
40%
50%
60%
80% 100% 120% 140% 160% 180%
K/S
C
D
S

I
m
p
l
i
e
d

V
o
l
$\lambda^{dis}$
$\lambda^{exp}$
Figure: Log-normal credit default swaption volatility implied from model with
T
e
= 1y, T
t
= 5y as a function of the inverse moneyness K/ c (T
e
, T
e
+T
t
).
Illustrations
JPM
0%
50%
100%
150%
200%
250%
14% 41% 69% 96% 123%
K/S
E
q
u
i
t
y

I
m
p
l
i
e
d

V
o
l
$\lambda^{dis}$
$\lambda^{exp}$
MarketVol
MS
0%
50%
100%
150%
200%
250%
30% 59% 89% 118%
K/S
E
q
u
i
t
y

I
m
p
l
i
e
d

V
o
l
$\lambda^{dis}$
$\lambda^{exp}$
MarketVol
Figure: Log-normal equity volatility implied by the model as a function of inverse
moneyness K/s for put options with T = 0.5y.
Illustrations
rho=0.5
0.00
0.20
0.40
0.60
0.80
1.00
0.25 1.25 2.25 3.25 4.25 5.25 6.25 7.25 8.25 9.25
T
G
a
u
s
s
i
a
n

C
o
r
r
e
l
a
t
i
o
n
$\lambda^{dis}$
$\lambda^{exp}$
rho=0.99
0.00
0.20
0.40
0.60
0.80
1.00
0.25 1.25 2.25 3.25 4.25 5.25 6.25 7.25 8.25 9.25
T
G
a
u
s
s
i
a
n

C
o
r
r
e
l
a
t
i
o
n
$\lambda^{dis}$
$\lambda^{exp}$
Figure: Implied Gaussian correlations for FTDSs.
Illustrations
96
106
116
126
136
146
156
1 2 3 4 5 6 7 8 9 10
T, years
S
p
r
e
a
d
,

b
p
InputSpread
Risky Seller, rho=0.75
Risky Buyer, rho=0.75
Risky Seller, rho=-0.75
Risky Buyer, rho=-0.75
30
35
40
45
50
55
60
1 2 3 4 5 6 7 8 9 10
T, years
S
p
r
e
a
d
,

b
p
InputSpread
Risky Seller, rho=-0.75
Risky Buyer, rho=-0.75
Figure: Equilibrium spread for protection buyer and protection seller of a CDS on
MS with JPM as the counterparty (lhs); same for a CDS on JPM with MS as the
counterparty (rhs).
Conclusions
We gave a broad overview of the state of CDS modeling.
We showed how to build meaningful structural default models and
connected them to reduced form models.
We showed that even the simplest elementary building blocks of
credit universe are dicult to describe properly.
The opinions expressed in this talk are those of the speaker and do
not necessarily reect the views or opinions of Bank of America
Merrill Lynch.
Conclusions
Merrill Lynch.
Conclusions
Merrill Lynch.
Conclusions
Merrill Lynch.

Lip Ton 2010

Загружено:

Сведения о документе

Исходное описание:

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Lip Ton 2010

Загружено:

Авторское право:

Доступные форматы

CVA for CDS via SDM and Related Issues

Analytical and Numerical Results

is the corresponding dierential operator and H (.) is the

is the corresponding dierential operator and H (.) is the

is the corresponding dierential operator and H (.) is the

is the corresponding dierential operator and H (.) is the

removes drift terms provided that

(t) with intensity

(t) with intensity

(t) with intensity

Вам также может понравиться