NYU Lecture1
NYU Lecture1
NYU Lecture1
k=1
G
k
T
k
k!
,
where I is the identity matrix. Notice that
M(1) = e
G
and M(n) = e
Gn
= M(1)
n
, as
above
Fitting a generator matrix to historical data
allows us to make statements about default
probabilities at all horizons, not just annual
ones
The generator matrix used to produce the ta-
ble above is
4
G =
Grade Aaa Aa A Baa Ba B Default
Aaa
-0.0971753 0.0788 0.0087 0.0065 0.0026 0.0004 0.000175
Aa 0.0788 -0.16071 0.072 0.0051 0.0029 0.00143 0.000482
A 0.0182 0.13035 -0.22666 0.0718 0.0031 0.0025 0.000705
Baa 0.0025 0.0083 0.0433 -0.10179 0.0452 0.001 0.001491
Ba 0.0009 0.0079 0.0463 0.0846 -0.32923 0.1711 0.018426
B 0 0 0 0 0.1182 -0.24746 0.129255
Default 0 0 0 0 0 0 0
A generator matrix with risk-neutral probabil-
ities would imply much higher default proba-
bilites (risk aversion)
Structural Models
In the dynamic description of the (random)
default time, there are two major classes of
models: structural and reduced-form. The lat-
ter is nearly always the model of choice in a
derivatives trading setting; the former is more
fundamental and nds applications in prop
trading or econometrics.
Structural models are based on balance sheet
information and historically emerged rst (they
originate with Robert Merton in the 70s)
and will do so as well in this class
Simple Merton Model
A simple 1-period strutural model considers a
rm with assets V , debt B, and equity S. By
MM theorem
V (t) = B(t) +S(t).
Consider now a horizon T, and assume (sim-
plistically) that the company is left alone
until time T, at which point it is liquidated
and the proceeds handed out to debt and eq-
uity holders. Also assume that the rms debt
consists of T-maturity zero-coupon bonds with
notional D.
At time T, the debt holders evidently receive
B(T) = min(D, V (T)) = V (T) (V (T) D)
+
,
where we use the notaton x
+
= max(x, 0).
The equity holders receive
S(T) = (V (T) D)
+
= V (T) B(T).
5
In the Merton approach we assume that the
assets pay no dividends and satisfy a simple
risk-neutral diusion
dV (t)/V (t) = rdt +
V
dW(t).
Then we get, from Black-Scholes,
S(0) = V (0)(d
+
) De
rT
(d
);
B(0) = V (0) S(0),
where
d
=
ln(V (0)/D) +
_
r
1
2
2
V
_
T
T
We can use this expression in a number of
ways. First, if we know S(0) and B(0), we
can iterate to nd the asset volatility
V
. Sec-
ond, if we know
V
, we can work backward to
nd B(0) (and V (0)), assuming that only S(0)
is known.
The rst application is used in a number of
systems to create a functional link between
S(0) and B(0), used in risk-management ap-
plications (hedge bonds with equity, and vice-
versa). The second is the original idea of the
Merton model and gives us a way to price risky
debt B(0), even . Often we would estimate
V
from the volatility of the equity:
dS(t) = ..dt+
V
V (t)dW(t) ..dt+
S
(t)S(t)dW(t)
such that
V
V (t) =
S
(t)S(t), or
V
S
S(0)
V (0)
S
S(0)
S(0) +D
.
In the Merton model, the risk-neutral default
probability over the horizon [0, T] is
Pr (default) = Pr (V (T) < D) = (d
).
We can calibrate
V
to hit this number if we
know it or we can take it as a fundamental
structural output of the model.
Merton Extensions
One problem with the basic Merton model is
the fact that it is a one-period model and can-
not tell us anything other than default proba-
bilities to a single xed horizon
An obvious extension is to introduce a contin-
uous barrier H and have the rm default the
moment the assets hit this critical level. To
describe this, let be the default time of the
rm. Then
= inf {t > 0 : V (t) H} .
With the asset process following dV (t)/V (t) =
dt +
V
dW(t), it can be shown then that
Pr ( < t) =
_
h at
t
_
+e
2ah/
2
V
_
at +h
t
_
,
where a = 0.5
2
V
and h = ln(H/V (0)) .
6
While this extension allows for a full term struc-
ture of default probabilities, it is typically not
realistic. For instance, it can be demonstrated
(see Homework 1) that the credit spread term
structure generated by this model has collaps-
ing spreads for short maturities (contrary to
real markets).
This is essentially a consequence of the fact
that defaults generated by diusion processes
are predictable with slowly deteriorating rm
assets, rather than surprising (e.g, as in En-
ron or Parmalat).
More Extensions
To overcome the problem with collapsing credit
spreads, there has been a number of attempts
to infuse more uncertainty into the barrier model
above. One approach (CreditGrades) assumes
that the barrier H is unobservable at time 0
instead we can only see its mean and standard
deviation.
That means that we can be in default at time
0 (without knowing it). An instant later we
can then default.
CreditGrades makes no dynamic sense, how-
ever, and credit spreads would be innitely high
at t = 0 (and then collapse an instant later).
Better, but more convoluted ideas, exist in the
literature basically a matter of continuously
withholding information.
In practice, these models have few, if any, ap-
plications
7
A dierent extension involves replacing the idea
of a constant barrier with a curved one. There
is no closed-form expression for default prob-
abilities in this approach, so numerical work
(in a grid) must be employed. The resulting
barrier H(t) looks something like this:
t
V(0)
H(t)
Note that this model has poor stationarity prop-
erties: short-term default spreads will very likely
collapse once time passes
Finally, some authors have enrichened the ba-
sic model with a more realistic model for lever-
age ratios. In the simple model above, the
leverage ratio of the rm in question is dened
as
L(t) =
B(t)
V (t)
=
B(t)
B(t) +S(t)
.
In our simple models, this ratio can become
very high or very low, depending on how the
value of assets uctuate. In reality, however,
rms tend to keep their leverage ratios quite
sticky: when assets (and stocks) are very
high, they tend to issue more debt; and vice
versa.
See Goldstein and Colin-Dufresne (JOF, 2001)
for a mean-reverting model that tries to en-
compass this phenomenon. Basic idea is to
make the barrier H a random process, reect-
ing the fact that debt levels change over time.
Reduced-Form Models
We are done with our introduction to struc-
tural models,and shall only see them very briey
again in this class
We now turn to the type of models that are
used in credit derivatives trading environments:
reduced-form models. Understanding these mod-
els require us to have a good understanding of
Poisson processes and Cox processes, so we
proceed to this.
Poisson Processes
So far in this class, we have only considered
continuous stochastic processes, most notably
Brownian motion
We shall now look at a completely dierent
type of process, useful for processes with jumps
or as our primary interest shall be default
events
8
For this, we introduce the Poisson process N(t),
an increasing process that takes only integer
values 0, 1, 2, ...
The sample paths of the Poisson process is
a stair-case, with unit-sized jumps at random
times
1
,
2
, ...
N(t)
3
2
1
t
1
Given that N(t) = n, the probability that N(t+
dt) = n + 1 is dt (and the probability that
N(t +dt) = n is 1 dt)
Here, > 0 is the intensity of the Poisson
process
Notice that the Poisson process has indepen-
dent increments and is Markovian
It is not dicult to show that (with N(0) = 0)
Pr (N(t) = n) = e
t
(t)
n
n!
,
and that
E(N(t)) = t.
The Poisson process is memory-less, in the
sense that
Pr (N(t +) = n +m| N(t) = n) = Pr (N() = m) .
Let us consider the time of the rst jump
1
.
Its probability distribution is given by
Pr (
1
> t) = e
t
9
with rst moment
E(N(t)) = 1/.
The density of
1
can be found by dierentia-
tion of the cumulative distribution:
Pr (
1
[t, t +dt]) = e
t
dt.
The distributions of all interarrival time
2
1
,
2
, ... are all identical to the distribution of
1
(due to the memory-less property).
Non-homogenous and Doubly-Stochastic
Poisson processes
If the intensity of the Poisson process depends
on time (a so-called non-homogeous Poisson
process), we basically replace in previous re-
sults t with
_
t
0
(u)du. Key results are,
Pr (N(t) = n) = e
_
t
0
(u) du
_
_
t
0
(u) du
_
n
n!
,
E(N(t)) =
_
t
0
(u) du,
Pr (
1
> t) = e
_
t
0
(u) du
,
Pr (
1
[t, t +dt]) = (t)e
_
t
0
(u) du
dt
We can also let the intensity be stochastic.
The way this works is that a path of the in-
tensity is drawn rst, and then conditional
on this path we use the results for the non-
homogenous Poisson process. For instance,
Pr (
1
> t| {(u), 0 u t}) = e
_
t
0
(u) du
.
By forming the expectation over all paths, we
get the unconditional probability as
Pr (
1
> t) = E
_
e
_
t
0
(u) du
_
.
10
This doubly stochastic non-homogenous Pois-
son process is often known as a Cox process
Notice the close relation between the intensity
and the short rate r in an interest rate model.
Let X(0, T) be the time 0 survival probabil-
ity X(0, T) = Pr( > T) and let P(0, T) be
the time 0 discount bond maturing at time T.
Then
X(0, T) = E
_
e
_
T
0
(u) du
_
P(0, T) = E
_
e
_
T
0
r(u) du
_
More generally, at time t,
X(t, T) = 1
>t
E
t
_
e
_
T
t
(u) du
_
P(t, T) = E
t
_
e
_
T
t
r(u) du
_