Stochastic calculus for jump processes

Stochastic intensity based models

Stochastic recovery models

Credit Risk Models

5. Advanced credit models

Andrew Lesniewski

Baruch College
New York

Fall 2016

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Outline

1 Stochastic calculus for jump processes

2 Stochastic intensity based models

3 Stochastic recovery models

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Jump processes

Stochastic calculus, used in the study of diffusion processes driven by Brownian

motion, can be extended to stochastic processes driven by jump processes such
as counting processes (Poisson, compound Poisson, Cox).
We shall take a few slides to review basic facts about jump processes and their
stochastic calculus.
The processes we are considering in this course are known in mathematics as
CADLAG (which is the acronym of French for “right-continuous with left limits”).
We will be assuming that a process X can jump only at finitely many times. All
the jumps are assumed to be of finite magnitude.
The jump amount of X at time t is denoted by ∆X (t).

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Jump processes

For a jump process X , we denote:

X (t− ) = lim X (s) .

s%t

We also introduce the following notation:

(i) The jump size at time t is thus:

∆X (t) = X (t) − X (t− ).

(ii) The discontinuous part of the process:

X
X d (t) = ∆X (s) .
s≤t

(iii) The continuous part of the process:

X c (t) = X (t) − X d (t) .

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Ito’s lemma for jump processes

Ito’s lemma is a powerful tool allowing us to easily manipulate complex

stochastic processes.
Let X = (X1 , . . . , Xn ) be an n-dimensional process with a finite number of jumps.
For a smooth function f (t, X ) we have
Z t ∂f (s, X (s− )) X Z t ∂f (s, X (s− ))
f (t, X (t)) = f (0, X0 ) + ds + dXic (s)
0 ∂s 0 ∂Xi
1≤i≤n

1 X Z t ∂ 2 f (s, X (s− )) X
+ dXic (s) dXjc (s) + ∆f (sk , X (sk )),
2 0 ∂Xi ∂Xj s ≤t
1≤i,j≤n k

where the last (finite) sum extends over all jump times sk of the process.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Ito’s lemma for jump processes

Alternatively, we can formulate Ito’s lemma in the differential form:
∂f (t, X (t− )) X ∂f (t, X (t− ))
df (t, X (t)) = dt + dXic (t)
∂t ∂Xi
1≤i≤n

1 X ∂ 2 f (t, X (t− ))
+ dXic (t) dXjc (t) + ∆f (t, X (t))dN (t) ,
2 ∂Xi ∂Xj
1≤i,j≤n

where N (t) is the process counting the number of jumps up to time t:

X
N (t) = 1sk ≤t .
sk ≤t

The last term in the formula above can explicitly be written as

X
∆f (sk , X (sk ))δ(t − sk )dt,
sk ≤t

where δ (t) denotes Dirac’s delta.

We will see applications of this formula later in this lecture and in the following
lectures.
A. Lesniewski Credit Risk Models
 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Stochastic intensity based models

The methodology used to model CMDS contracts can easily be extended to any
European style payoff function u (x) of the par spread at expiration: we just set
g (x) = f (x) u (x) in the formulas derived in Lecture Notes #4.
Clearly, the swaption replication method continues to hold, and we can price an
option whose payoff is any function of the par CDS spread.
This does not include exotic options whose payoffs depend on the term structure
of spreads. Such options include barrier options and Bermudan options.
Modeling such options requires dynamic intensity models, very much like exotic
interest rate options require term structure models.
Let us consider a Cox process intensity with a single factor risk neutral diffusion
dynamics given by:

dλ (t) = µ(t, λ (t))dt + σ(t, λ (t))dW (t) .

This is a generalization of the Gaussian models considered in Lecture Notes #2.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Stochastic intensity based models

For simplicity, we assume that interest rates are deterministic, and consider a
credit risky derivative V (t, λ (t)).
By R(t) = R(t, λ (t)) we denote the recovery amount associated with the claim
V (t).
Applying Ito’s lemma with jumps yields

 ∂V ∂V 1 ∂2V  ∂V
dV = +µ + σ2 dt + σ dW + (R − V )dN (t) ,
∂t ∂λ 2 ∂λ2 ∂λ

where N (t) is the underlying counting process.

Notice the presence of the jump term (R − V )dN (t) in the equation above:
when N jumps, V becomes the recovery value R.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Stochastic intensity based models

Now, we take an expectation of this equation:

 ∂V ∂V 1 ∂2V 
E[dV (t)] = +µ + σ2 dt + λ(R − V )dt,
∂t ∂λ 2 ∂λ2

where we have used the fact that, for a counting process, E[dN (t)] = λ (t) dt.
In the risk neutral measure, this expectation is equal to r (t) V (t) dt.
As a consequence, we have obtain the following PDE for V = V (t, λ (t)):

∂V ∂V 1 ∂2V
+µ + σ2 + λR = (r + λ)V .
∂t ∂λ 2 ∂λ2

Note the following differences with the standard backward Kolmogorov equations
of option pricing:
(i) The right hand side contains the factor r + λ rather than r .
(ii) The left hand side contains the term λR(t, λ) without partial derivatives.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Stochastic intensity based models

This equation is tractable by means of the usual numerical methods such as

Crank-Nicolson, fully implicit, etc.
It is subject to a terminal boundary condition at option maturity, and, possibly, to
intermediate conditions if the option has barriers, early exercise, or other exotic
features.
For example:
(i) For a zero-recovery, risky zero coupon bond maturing at time T , R = 0,
and the terminal condition is V (T , λ) = 1.
(ii) For a risk-free zero-coupon bond, R(t, λ (t)) = P(t, T ), and
V (T , λ (T )) = 1.
(iii) For a contract paying $1 at the time of default, provided that the default
occurs before time T , R(t, λ (t)) = 1 and V (T , λ (T )) = 0.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Affine intensity models

Terminal boundary conditions in the PDE can be specified as function q(λ), i.e.
V (T , λ) = q(λ).
In practice, the boundary conditions arising in pricing applications are written in
terms of the survival probabilities S(t, s) depending on a range of s.
This raises the question, under what conditions can S(t, s) (where t ≤ s) be
expressed as a function of λ (t).
We say that the model admits a reconstitution formula, if the entire survival curve
S(t, s) depends on the intensity process only through λ (t).
An example of such a model is the Gaussian Hull-White model for λ (t) which we
discussed in Lecture Notes #2, namely

S(t, s) = eA(s,t)−hλ (s−t)λ(t) ,

with deterministic A(s, t).

The Hull-White model has some limitations. This is especially true for low credit
spreads, for which the probabilities of generating negative hazard rates may be
significant.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Affine intensity models

An alternative choice is a CIR process:

p
dλ (t) = κ (t) (θ (t) − λ (t))dt + σ (t) λ (t)dW (t) .

This is an affine model (as is the Hull-White model) in the sense that the survival
probability is of the form

S(t, T ) = eA(t,T )−C(t,T )λ(t) .

The coefficients A(t, T ) and C(t, T ) satisfy Riccati’s differential equations:

dA (t)
− κ (t) θ (t) C (t) = 0,
dt
dC (t) 1
− κ (t) C (t) − σ (t)2 C (t)2 + 1 = 0,
dt 2

subject to the terminal condition C(T , T ) = A(T , T ) = 0.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Affine intensity models

These ODEs can be solved for all t and T using standard numerical algorithms,
such as the Runge-Kutta method.
As a result, if we need to impose a terminal condition in terms of S(T , s), we set

V (T , λ) = eA(T ,s)−C(T ,s)λ .

The same is true for various intermediate boundary conditions required for exotic
options.
Very importantly, the existence of the reconstitution formula allows us, for given
σ (t) and κ (t), to calibrate the function θ (t) to match exactly the initial risky zero
coupon curve P(0, T ), for all T .
The same is, of course, true for the Hull-White model.
σ (t) is calibrated to match the market prices of CDS swaptions.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery modeling

So far, we have treated the recovery rate on a CDS as a distraction. The

recovery rate is assumed to be a constant determined by the market convention
(say, 40%).
In fact, modeling recovery rates is an inexact science, as available market
information is insufficient.
There is a (rather illiquid) market for recovery rate locks, which in principle
provides information about the implied recovery rates.
A recovery rate lock is an instrument structured similarly to a forward rate
agreement, on which the counterparties exchange, in case of a default of the
reference name, the net of an actual recovery rated and a fixed coupon.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery modeling

We will now show how to extend the reduced form modeling framework to
include the stochastic nature of the recovery rate.
We thus assume that the recovery process R (t) is adapted with respect to Gt .
The time 0 price P0rec (T ) of a nonzero recovery risky zero coupon bond is given
by
h RT Rτ i
P0rec (T ) = EQ e− 0 1τ ≥T + e−
r (s)ds 0 r (s)ds
1τ ≤T R(τ )
h RT i Z T h Rt i
= EQ e− 0 (r (s)+λ(s))ds + EQ e− 0 (r (s)+λ(s))ds R (t) λ (t) dt
0
Z T
− 0t (r (s)+λ(s))ds
h R i
Q
= P0 (T ) + E e R (t) λ (t) dt.
0

It straightforward to extend this formula to the case of a coupon bearing bond.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery modeling

Various approaches to modeling the recovery process have been proposed. We

focus on the following three approaches:
(i) recovery of treasury,
(ii) recovery of market value,
(iii) recovery of par.
These models do not attempt to model the outcomes of actual restructuring and
bankruptcy proceedings. Their purpose is to model the value of the settlement,
and they may be thought of as reflecting various different types of rules used to
distribute the recovery amount among the claimants.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery of treasury

In the recovery of treasury (RT) model, the recovery amount is expressed in

terms of the credit riskless zero coupon bond, namely

R (t) = xPt (T )
RT
= xe− t r (s)ds
,

where x is a constant factor (we shall relax this assumption shortly).

Consequently,
Z T h Rt RT i
P0rec (T ) = P0 (T ) + x EQ e− 0 (r (s)+λ(s))ds e− t r (s)ds
λ (t) dt
0
h RT Z T Rt i
= P0 (T ) + xE Q
e− 0 r (s)ds
e− 0 λ(s)ds
λ (t) dt
0
h RT  RT i
= P0 (T ) + xEQ e− 0 r (s)ds
1 − e− 0 λ(s)ds
.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery of treasury

Rearranging the terms we see that, under RT,

P0rec (T ) = xP0 (T ) + (1 − x)P0 (T ) .

In other words, the price of an RT zero coupon bond is the sum of x units of a
riskless zero coupon bond and 1 − x units of a zero recovery zero coupon bond.
The same result holds for a coupon bearing bond B:

B0rec = xB0 + (1 − x)B0 ,

where B is the price of the equivalent default free bond.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery of treasury

The RT model is viewed as weak.

Under RT coupon bonds can recover more than their par value. This occurs
when the risky bond has a high default risk, long maturity, and trades close to
par. Under these circumstances, the equivalent riskless bond will trade well
above par. A large recovery rate may results in a value larger than the par value
of the bond.
Under RT with constant fraction x, bonds of different maturities and different
coupons have different loss severities.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery of market value

Under the recovery of market value (RMV) model, the defaulted zero coupon
bond pays a fraction of its market value preceding the default:

R (t) = xPtrec
−
(T ) ,

where Ptrec
−
(T ) denotes the price of the bond just prior to the default.
We shall now show that, explicitly,
h RT i
P0rec (T ) = EQ e− 0 (r (s)+(1−x)λ(s))ds .

To this end, we notice first that

h i h i
EQ 1τ ≥T . . . | Ft = 1τ ≥t EQ 1τ ≥T . . . | Ft .

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery of market value

Therefore, denoting V (t) = Ptrec (T ), we find that

h RT Rτ i
1t≤τ V (t) = EQ 1τ ≥T e− t + 1t≤τ − 1τ ≥T xe− t r (s)ds V (τ ) | Ft
r (s)ds

h RT Z T Ru i
= 1t≤τ EQ e− t (r (s)+λ(s))ds + x e− t (r (s)+λ(s))ds V (u) λ (u) du | Ft .
t

Now, let us introduce the martingale

h RT Z T Ru i
M (t) = EQ e− 0 (r (s)+λ(s))ds + x e− 0 (r (s)+λ(s))ds
V (u) λ (u) du | Ft .
0

Then, we can express V (t) in terms of M (t) as follows:

Rt  Z t Ru 
V (t) = e 0 (r (s)+λ(s))ds M (t) − x e− 0 (r (s)+λ(s))ds V (u) λ (u) du .
0

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery of market value

Ito’s lemma implies that

 Rt  Rt
d e− 0 (r (s)+(1−x)λ(s))ds V (t) = ex 0 λ(s)ds dM (t) .

Taking the expectation value yields

h  Rt i
EQ d e− 0 (r (s)+(1−x)λ(s))ds V (t) = 0,

since M is a martingale and so EQ [dM (t)] = 0.

Integrating the equation above from 0 to T , yields
h RT i
EQ e− 0 (r (s)+(1−x)λ(s))ds V (T ) − V (0) = 0.

Keeping in mind that V (T ) = 1, we get the desired result.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Recovery of par

Under the recovery of par (RP) model, the defaulted bond pays a fraction of its
par value, R (t) = x, for all t ≥ 0.
The idea of the RP model is a liquidation under a bankruptcy court supervision.
Typically, the court distributes the proceeds in proportion to the claims filed.
For the zero coupon case, the RP model yields
Z T h Rt i
P0rec (T ) = P0 (T ) + x EQ e− 0 (r (s)+λ(s))ds λ (t) dt.
0

In general, the value of a claim under RP is the sum of the claim under zero
 Rt
recovery and x 0T EQ e− 0 (r (s)+λ(s))ds V (t) λ (t) dt, where V (t) is the value of
R 
the claim at time t.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Stochastic recovery rates

We have been assuming so far that the value of the recovery fraction x is a
constant. This simplifies the calculations but ignores the recovery risk.
Recovery risk may have a significant impact on the prices of credit sensitive
instruments.
In order to model this risk we assume that x is drawn from a probability
distribution, x ∼ F .
The prices of bonds under the recovery models discussed above are thus given
in terms of expected values under the joint distribution Q and F .

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Stochastic recovery rates

Assuming that Q and F are independent, we thus obtain:

(i) Under RT:

P0rec (T ) = EF [x]P0 (T ) + (1 − EF [x])P0 (T ) .

(ii) Under RMV:

h h RT ii
P0rec (T ) = EF EQ e− 0 (r (s)+(1−x)λ(s))ds .

(iii) Under RP:

Z T h Rt i
P0rec (T ) = P0 (T ) + EF [x] EQ e− 0 (r (s)+λ(s))ds λ (t) dt.
0

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Stochastic recovery rates

A possible choice of a parametric distribution for recovery fraction modeling is
the beta distribution B(p, q), where p, q > 0.
Its PDF f (x) is given by

1
f (x) = x p−1 (1 − x)q−1 ,
B(p, q)

for x ∈ [0, 1].

The beta function B(p, q) is defined by
Z 1
B(p, q) = t p−1 (1 − t)q−1 dt .
0

Its mean and variance are given by

p
E[X ] = ,
p+q
pq
Var[X ] = ,
(p + q)2 (p + q + 1)

respectively.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

Stochastic recovery rates

An alternative choice is given by a transformed Gaussian distribution: its domain

R should be mapped in a one-to-one manner on the interval [0, 1].
The logit transform:
ex
l (x) = ,
1 + ex
and its inverse
l
x = log
1−l
are a convenient choice.

A. Lesniewski Credit Risk Models

 Stochastic calculus for jump processes
Stochastic intensity based models
Stochastic recovery models

References

Andersen, L.: Credit models, lecture notes (2010).

Applebaum, D.: Levy Processes and Stochastic Calculus, Cambridge University

Press (2009).
Schönbucher, P. J.: Credit Derivatives Pricing Models, Wiley Finance (2003).

A. Lesniewski Credit Risk Models