Counting processes

Reduced form models

Credit Risk Models

2. Reduced Form Models

Andrew Lesniewski

Baruch College
New York

Fall 2016

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Outline

1 Counting processes

2 Reduced form models

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Reduced form models

In a structural model of credit, a default event is an endogenous feature of the

model and it results as a consequence of the (stochastic) value of the entity
falling below a certain threshold.
In order to take all factors properly into account, such a description requires a
detailed knowledge of the capital structure of the entity, and lead to complicated
computational problems.
In contrast, in the reduced form approach, a default event is an exogenous
occurrence, which is governed exclusively by the assumed probability
distribution of a random time τ , namely the time to default.
While structural models offer a better intuitive insight into the nature of the credit
of the entity, reduce form models are generally more transparent and numerically
tractable.
We begin by reviewing the basics of the theory of counting processes.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Reduced form models

Credit events are examples of financial processes with discrete outcomes

(unlike, say, stock prices which are usually treated as continuous processes).
The basic probabilistic set up to model event risk is based on the concept of a
random time τ .
The stochastic process 1τ ≤t defines a filtration (Gt )t≥0 , which is referred to as
the information set.
Of key importance to financial event modeling are counting processes.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Counting processes

An stochastic process N (t) is called a counting process if:

N (t) is defined on a probability space(Ω, (Gt )t≥0 , P).
N (t) is integer valued, non-negative, with N (0) = 0.
Sample paths of N (t) are right continuous, piecewise constant, may have jumps
of size 1.
Jump times τ1 < τ2 < . . . are stopping times at which the process N (t)
increases by 1.
The intensity λ (t) of N (t) is defined by

λ (t) dt = P(N(t + dt) − N (t) = 1 | Gt− )

Alternatively,
λ (t) dt = E[N(t + dt) − N (t) | Gt− .]

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Counting processes
The graph below shows a sample path of a counting process.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Counting processes
The simplest counting process is given by N (t) ∈ {0, 1} and constant intensity
λ. In this case,

P(N (t) = 0) = P(τ ≥ t)

= e−λt .

This probability distribution is called the exponential distribution.

Indeed,

n  j
Y j + 1 
P(τ ≥ t) = lim 1−P t ≤τ ≤ t
n→∞ n n
j=0
n 
Y t
= lim 1−λ
n→∞ n
j=0
 t n
= 1−λ
n
= e−λt .

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Poisson process

Another important example of a counting process is a Poisson process.

A Poisson process is defined by the two requirements:
The probability that given N (t) = n, N(t + dt) = n + 1 is λdt,
The probability that given N (t) = n, N(t + dt) > n + 1 is 0.
Here, λ > 0 is the intensity of the Poisson process.
Notice that the Poisson process has independent increments and is Markovian.
We shall show that
(λt)n −λt
P(N(t) = n) = e .
n!

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Poisson process

Indeed, using N (0) = 0,

N  t n  t N−n
P(N (t) = n) = lim λ 1−λ
N→∞ n N N
√
N N e−N 2πN  t n  t N−n
= lim p λ 1−λ
N→∞ n!(N − n)N−n e −N+n 2π(N − n) N N
(λt)n N N−n+1/2  t N−n
= lim 1−λ
n! N→∞ (N − n)N−n+1/2 en N
(λt)n 1  t N−n
= lim n 1 − λ
n! N→∞ (1 −
N
)N−n+1/2 en N
(λt)n 1  t N
= lim 1−λ
n! N→∞ (1 − n )N en N
N
(λt)n −λt
= e ,
n!
√
where we have used Stirling’s approximation n! ≈ (n/e)n 2πn.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Poisson process

Furthermore, we have
E[N (t)] = λt,
and
Var[N (t)] = λt.

An important property of the Poisson process is that it is memoryless, meaning

that
P(N(t + h) = n + m | N (t) = n) = P(N(h) = m).

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Poisson process

Let us consider the time of the first jump (or first arrival) τ1 in a Poisson process.
This is simply our first example of a counting process, and so its survival
probability is given by the exponential distribution:

P(τ1 > t) = e−λt ,

with the first moment

1
E[τ1 ] = .
λ
The density of τ1 can be found by differentiation of the cumulative distribution,

P(τ1 ∈ [t, t + dt]) = e−λt λdt.

The distributions of all interarrival times τ2 − τ1 , τ3 − τ2 , . . ., are all identical to

the distribution of τ1 (due to the memoryless property).

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Inhomogeneous Poisson process

Another important example of a counting process is an inhomogeneous Poisson

process.
It is a generalization of the Poisson procces, in which the intensity is a
deterministic function of time λ (t).
The calculations involving inhomogeneous Poisson processes follow the lines of
the homogeneous Poisson processes. It is easy to see that all the results stated
above for a homogeneous Poisson process hold also for an inhomogeneous
Poisson process with λ replaced by the average 1t 0t λ (s) ds.
R

In particular, the probability distribution is given by

Rt
n
λ (s) ds Rt
P(N(t) = n) = 0
e− 0 λ(s)ds
.
n!

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Inhomogeneous Poisson process

Furthermore,
Z t
E[N (t)] = λ (s) ds,
0
Z t
Var[N (t)] = λ (s) ds
0

For the properties of the first arrival time we have:

Rt
P(τ1 > t) = e− 0 λ(s)ds
,
Rt
P(τ1 ∈ [t, t + dt]) = e− 0 λ(s)ds λ (t) dt.

Rt
The integral A (t) = 0 λ (s) ds is called the compensator of the process N (t).

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Cox processes

A general class of counting processes are Cox processes.

A Cox process N (t) is associated with an information set of the form Gt ∨ Ft ,
t > 0, where Gt is the information set pertaining to jump times, and where Ft is
the “background” information set.
One can think of Ft as containing the background economic information.
A point process N (t) is a Cox process if, conditioned on Ft , N (t) is an
inhomogeneous Poisson process with a deterministic intensity λ (s), 0 ≤ s ≤ t.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Cox processes

Intuitively, a Cox process can be thought of as an inhomogeneous Poisson

process whose intensity is stochastic itself.
We can think about it in the following way: We draw an intensity path first, and
then - conditional on this path - we use the properties the inhomogeneous
Poisson process.
The compensator of a Cox process is the integrated intensity process
A (t) = 0t λ (s) ds.
R

In particular,

A (t)n −A(t)
P N (t) = n | {λ (s)}0≤s≤t = e ,
n!
and, taking expectation over all paths {λ (s)}, we get the absolute probability
distribution:
1 
E A (t)n e−A(t) .

P(N (t) = n) =
n!

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Cox processes

As another example, let us find P(τ1 > t). We have

P τ1 > t | {λ (s)}0≤s≤t = e−A(t) ,

and by taking expectation over all paths {λ (s)}, we get the absolute probability:

P(τ1 > t) = E e−A(t) .

 

Finally, the absolute expectation and variance of a Cox process are given by

E[N (t)] = E[A (t)],

Var[N (t)] = E[A (t)],

respectively.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Zero coupons and survival probabilities

Note that there is a close analogy between the intensity λ (t) in a Cox model and
the short rate r (t) in an interest rate model.
Let S(t, T ) denote the time t survival probability S(t, T ) = P(τ1 > T ), and let
P(t, T ) be the time t discount bond maturing at T . Then,

 RT
S(0, T ) = E e− 0 λ(s)ds

,
 RT
P(0, T ) = E e− 0 r (s)ds

.

More generally, at time t ≥ 0,

 RT
S(t, T ) = 1τ1 >t E e− t λ(s)ds ,


 RT
P(t, T ) = E e− t r (s)ds .


Note the presence of the process 1τ1 >t in the first formula: the time t survival
probability for T is zero, if the event has occurred prior to t.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Risky zero coupon

Consider now a risky zero coupon bond with a zero recovery rate.
Assuming that the credit intensity follows a Cox process, its price is given by:

 RT
P(0, T ) = E e− 0 r (s)ds 1τ1 >T


  RT
= E E e− 0 r (s)ds 1τ1 >T | {λ (s)}0≤s≤T


  RT RT
= E E e− 0 r (s)ds e− 0 λ(s)ds | {λ (s)}0≤s≤T

R T
= E e− 0 (r (s)+λ(s))ds .
 

Notice that the presence of event risk (i.e. credit) amounts to increasing the
discounting rate by the intensity of default. We have already noted this in Lecture
Notes 1.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Risky zero coupon

In particular, if the default probability is independent of the rates probability, then,

RT
P(0, T ) = E[e− 0 r (s)ds
]E[1τ1 >T ]
= P(0, T )S(0, T ).

We have thus recovered the formula for the risky coupon bond discussed in
Lecture Notes 1 as a special case of our general framework.
For any time t ≥ 0, we have the following expression for the risky zero coupon
bond:
 RT
P(t, T ) = 1τ1 >t E e− t (r (s)+λ(s))ds .


Note the time t value of the bond is zero, if the default happens prior to t.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Gaussian structural model

As an explicit example, we consider the model in which both the rate and
intensity are Gaussian processes:

dr (t) = µ (t) dt + σr dW (t) ,

dλ (t) = θ (t) dt + σλ dZ (t) ,

with correlated Brownian motions:

dW (t) dZ (t) = ρdt.

Our goal is to compute P(0, T ).

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Gaussian structural model

Note that
Z t
r (t) = r0 + µ (s) ds + σr W (t) ,
0
Z t
λ (t) = λ0 + θ (s) ds + σλ Z (t) ,
0

are Gaussian variables with means

Z t
E[r (t)] = r0 + µ (s) ds,
0
Z t
E[λ (t)] = λ0 + θ (s) ds,
0

and covariance matrix (for t ≤ u)

σr2 t
 
ρσr σλ t
Cov(r (t) , λ(u)) = 2u .
ρσr σλ t σλ

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Gaussian structural model

RT
Hence, the mean and variance of the Gaussian variable X , 0 (r (t) + λ (t))dt
are
Z T
E[X ] = (r0 + λ0 )T + (T − t)(µ (t) + θ (t))dt,
0
ZZ
Var[X ] = 2(σr2 + 2ρσr σλ + σλ
2
) t dt du
0≤t≤u≤T
3
2 T
= (σr2 + 2ρσr σλ + σλ ) .
3

Using E[exp(−X )] = exp(−E[X ] + 21 Var[X ]) we find

3
P(0, T ) = P(0, T )S(0, T )eρσr σλ T /3
.

The Gaussian model implies thus that positive correlation between rates and
credit implies a higher zero coupon bond price.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Mean reverting model

A more realistic model is obtained by assuming that the riskless rate and default
intensity follow mean reverting processes:
 dµ (t) 
dr (t) = + kr (µ (t) − r (t)) dt + σr dW (t) ,
dt
 dθ (t) 
dλ (t) = + kλ (θ (t) − λ (t)) dt + σλ dZ (t) ,
dt
with correlated Brownian motions:

dW (t) dZ (t) = ρdt.

Here µ (t) and θ (t) are the deterministic long term means of the rate process
and intensity process, respectively, and kr and kλ are the corresponding mean
reversion speeds.
We assume that
µ (0) = r0 ,
θ (0) = λ0 .

This specification is reminiscent of the one-factor Hull-White model used in

modeling interest rates.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Mean reverting model

The solutions to the equations above read
Z t
r (t) = µ (t) + σr e−kr (t−u) dW (u) ,
0
Z t
λ (t) = θ (t) + σλ e−kλ (t−u) dZ (u) .
0

These are Gaussian variables with means

E[r (t)] = µ (t) ,

E[λ (t)] = θ (t) ,

respectively, and covariance matrix (for t ≤ u)

Cov(r (t) , λ(u)) =

1−exp(−2kr t) exp(−kλ (u−t))−exp(−kr t−kλ u)
!
σr2 2kr
ρσr σλ kr +kλ
exp(−kλ (u−t))−exp(−kr t−kλ u) 2 1−exp(−2kλ u)
.
ρσr σλ kr +kλ
σλ 2kλ

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Mean reverting model

RT
Hence, the mean and variance of the Gaussian variable X , 0 (r (t) + λ (t))dt
are
Z T
E[X ] = (µ (t) + θ (t))dt,
0
Z T Z T Z T
Var[X ] = σr2 hr (s)2 ds + σλ
2
hλ (s)2 ds + 2ρσr σλ hr (s) hλ (s) ds.
0 0 0

The functions hr (t) and hλ (t) used in the expression above are defined by:

1 − e−kr t
hr (t) = ,
kr
1 − e−kλ t
hλ (t) = ,
kλ

respectively.

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

Mean reverting model

Using again E[exp(−X )] = exp(−E[X ] + 21 Var[X ]), we find that the riskless zero
coupon bond and survival probability are given by
RT
µ(s)ds+ 12 σr2 0T hr (s)2 ds
R
P(0, T ) = e− 0 ,
RT 2 RT 2
θ(s)ds+ 21 σλ
S(0, T ) = e− 0 0 hλ (s) ds ,

respectively.
As a consequence, we find the following relation between the risky zero coupon,
riskless zero coupon, and the survival probability in the mean reverting model:
RT
hr (s)hλ (s)ds
P(0, T ) = P(0, T )S(0, T )eρσr σλ 0 .

A. Lesniewski Credit Risk Models

 Counting processes
Reduced form models

References

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

Lesniewski, A.: Interest rate models, lecture notes (2015).

O’Kane, D.: Modelling Single-Name and Mult-Name Credit Derivatives, Wiley

(2008).

A. Lesniewski Credit Risk Models