IEOR E4731: Credit Risk and Credit Derivatives: Lecture 13: A General Picture of Portfolio Credit Risk Modeling
IEOR E4731: Credit Risk and Credit Derivatives: Lecture 13: A General Picture of Portfolio Credit Risk Modeling
Structural models
Reduced-form models
Copula models
Intensity-based models
i
:= inf{t 0 |
_
t
0
i
(u)du > E
i
}
where
i
() is the intensity process adapted to the market
information {F
t
}
t
0 and E
i
is a unit exponential random
variable that is unobservable and independent of the market
information.
At each time t,
i
(t), the probability of obligor i default next
per unit time, is observable. The default time, however, is
unpredictable.
Key assumption: E
i
s are independent among dierent
obligors. As a result, given the intensities
i
s, the default
times are independent of each other.
Therefore, each
i
should depend on some observable
variables in the market.
We can model
i
(t) =
i
(U
i
(t), V (t)) for some function
i
,
where U
i
() is the dynamics of a vector of rm-specic
variables and V () is the dynamics of a vector of market-wide
variables.
Examples of U
i
: rms distance-to-default, rms size, rms
trailing one-year stock return. Examples of V (): the
three-month Treasury bill rate, the ten-year Treasury yield,
the trailing one-year return on S&P 500 index.
i=1
N
i
(1 R
i
)1
{t}
.
If we assume both N
i
s and R
i
s are constant among obligors,
the total loss becomes L(t) = N(1 R)n(t), where
n(t) :=
M
i=1
1
{t}
is the number of default up to time t.
7 / 23
Computing Portfolio Loss Distribution (Contd)
Assume
i
(t) = a
i
m
(t) +
id
i
(t), where
m
is the common
factor,
id
i
is the idiosyncratic factor, and a
i
is some constant.
Assume
m
,
id
i
s are independent.
For instance,
m
(t) =
m
(V (t)) for some function
m
,
id
i
(t) =
id
i
(U
i
(t)) for some function
id
i
, and V () and U
i
()
are independent.
i
> t |
id
(u),
m
(u), u 0
_
|
m
(u), u 0
_
=E
Q
_
e
t
0
(a
i
m
(u)+
id
i
(u))du
|
m
(u), u 0
_
=e
a
i
m
(t)
E
Q
_
e
id
(t)
_
,
where
m
(t) :=
_
t
0
m
(u)du and
id
(t) :=
_
t
0
id
(u)du are
cumulative intensity processes.
Denote by q
c
i
(t, z) the survival probability of
i
> t
conditioning on Z = z, i.e.,
q
c
i
(t, z) := e
a
i
z
E
Q
_
e
id
(t)
_
.
Suppose
m
(t) is an ane function of V (t), i.e.,
m
(t) =
0
+
1
V (t).
O-U processes
CIR processes
dX(t) = (b X(t))dt +
_
X(t)dW(t)
12 / 23
Conditional and Unconditional Survival Probabilities
id
(t)
_
has explicit formula. As a result, we can obtain q
c
i
(t, z).
E
Q
_
e
id
(t)
_
.
Therefore,
q
c
i
(t, z) =
e
a
i
z
E
Q
[e
a
i
Z
]
q
i
(t).
This equality holds once
id
i
and
m
are independent. No
further assumption on
id
i
and
m
is needed.
13 / 23
Conditional and Unconditional Survival Probabilities
(Contd)
Possible solutions:
i
(t) =
i
(U
i
(t), V (t), Y
i
(t), Y (t)), where Y
i
is the vector of
rm-specic unobservable variables and Y is the vector of
market-wide unobservable variables. Those unobservable
variables are called frailty.
i
(t) = a
i
m
(t) +
id
i
(t)
where
m
is the market factor and
id
i
s are idiosyncratic
factors.
There are some works on contagion eect, but all are not rich
enough to price multi-name credit derivatives
18 / 23
Top-down Models
_
f
ij
(t)t +o(t), j > i,
1
i
(t)t +o(t), j = i,
0, j < i,
where
i
(t) :=
N
j=i+1
f
ij
(t) is the total jump intensity. This
model is actually a continuous-time Markov chain. See for
instance, [Schonbucher, 2005].
20 / 23
Comparison of Bottom-up and Top-down Models
Open problems: