MIP - Politecnico

Intensity Modelling
Massimo Morini

Head of Interest Rate and Credit Models and Coordinator of Model Research
Banca IMI-Intesa San Paolo
Copyright 2015 Massimo Morini

Intensity Approach: Poisson Processes

Default of an individual name happens when a jump process Nt jumps for the first time

Intensity Models

1.2

Default at 7y

0.8
N(t)

0.6

0.4 Intensity Models

No Default
until 10y

0.2

1
0
0 1 2 3 4 5 6 7 8 9 10
Time
Copyright 2015 Massimo Morini

Intensity Modelling
The fundamental assumptions of intensity models are

A the probability of 1 event in ∆t is λ proportional to ∆t.

B ∆t is so small that there cannot be 2 events in ∆t.
C events in disjoint intervals are independent.

These assumptions will lead us in most of the approximations developed by the market for
easily expressing the quotations of single name and multiname derivatives. They also lead
naturally to the continuous time definition of Poisson processes

2
Copyright 2015 Massimo Morini

Intensity Modelling
Between t and T consider m intervals ∆t
T −t
T − t = m × ∆t and ∆t =
m
We have:
) (
T −t
1) P r [Nt+∆t − Nt = 1] = λ∆t = λ ,
m ( )
T −t
2) P r [Nt+∆t − Nt = 0] = 1 − λ∆t = 1 − λ
m ( )m
λ(T − t)
3) P r [Nt+m×∆t − Nt = 0] = P r [NT − Nt = 0] = 1−
m
( )
x n
Let ∆t tend to 0 (m → ∞). Since 1 + n −→ ex as n → ∞,

−λ(T −t)
P r [NT − Nt = 0] = e .

3
Copyright 2015 Massimo Morini

Intensity Modelling
In single name modelling we will take τ 1, the first jump of the poisson process, as default
time τ . Based on the above analysis, we can give the distribution of default time. In fact
P r (NT = 0) = e−λT amounts to say that
−λT
P r (τ > T ) = e ,
−λT
P r (τ ≤ T ) = Fτ (T ) = 1 − e ,

and the density is

−λT
fτ (T ) = λe .
which is the density of a (negative) exponential distribution

−ϑx 1 1
f (x) = ϑe , E (x) = , V (x) = 2 .
ϑ ϑ

with parameter ϑ = λ. Notice P r [NT − Nt = 0] = e−λ(T −t) corresponds to

−λT
P r (τ > T |τ > t) = e −λt .
e

4
Copyright 2015 Massimo Morini

Fundamental results for simulation and multiname modelling

Given a random variable X with distribution function FX (x) = Pr (X ≤ x), what is
the distribution function FY (y) of Y = g (X), where g (·) is an increasing function ?
By definition
( ) ( )
−1
FY (y) = Pr (Y ≤ y) = Pr (g (X) ≤ y) = Pr X ≤ g (y) = FX g −1 (y)

What happens when I choose as a function g(·) = FX (·)? We want the distribution of
Y = FX (X), namely we use the distribution function of X as a function of X . It is
( )
−1
FY (y) = FX FX (y) = y

But this is the distribution of a uniform random variable, since FU nif orm (x) = x.
Since Y = FX (X) takes values in [0, 1], FX (X) is distributed as a Uniform[0, 1],

Y = FX (X) ∼ U nif orm [0, 1]

5
Copyright 2015 Massimo Morini

Fundamental results for simulation and multiname modelling

Reverse the perspective. Take a uniform U ∼ U nif orm [0, 1]. What
( −1is the) distribution
−1
of Y = FX (U ) ? Recall that when Y = g (X), FY (y) = FX g (y) .

−1
Y = FX (U ) ⇒ FY (y) = FU nif orm (FX (y)) = FX (y) ,
−1
so FX (U ) has the same distribution as X when U is U nif orm [0, 1]. This is
fundamental in simulation and for introducing copulas. If u is a draw from a uniform,
x = FX−1 (u) is a draw from the distribution of the random variable X . For the default
time τ ,

τ = Fτ−1 (U ) = − λ1 ln(1 − U ) since Fτ (T ) = 1 − e−λT ,

By the way, notice that, if we set τ λ = ε, we see ε is now a unit exponential rv, since
(z) −z ε
Fε (z) = Fτ λ = 1 − e . So I can also write τ =
λ

6
Copyright 2015 Massimo Morini

Intensity directly from CDS quotation

Since jumps in disjoint intervals are independent and

Pr (1 jump in (t, t + dt]) = λdt,

the probability of having default in [t, t + dt] is

Pr (τ ∈ (t, t + dt]) (1)

= Pr (1 jump in (t, t + dt]) × Pr (no jump in [0, t])

= λdt × Pr (τ > t) = λdt × e−λt

Instead λdt, the probability of one jump in [t, t + dt], is the probability of having default
in [t, t + dt] given that default has not happened yet at t,

Pr (τ ∈ [t, t + dt] |τ > t) = λdt

7
Copyright 2015 Massimo Morini

Intensity directly from CDS quotation

Now consider a CDS where the premium spread S is paid continuously. Every instant t in
the life of the contract, the protection buyer pays S dt 1{τ >t} so, with deterministic
interest rates, its discounted expected payment is

S dt Pr (τ > t) P (0, t) .

At the same time the protection seller pays LGD 1{τ ∈(t−dt,t]} so its discounted
expected payment is

LGD Pr (τ ∈ (t − dt, t]) P (0, t) = LGD λ dt Pr (τ > t) P (0, t)

The equilibrium spread S can be chosen to equal the value of the two payments at any
time:

S dt Pr (τ > t) P (0, t) = LGD λ dt Pr (τ > t) P (0, t) ,

∗ S∗
S = LGD λ, λ=
LGD

8
Copyright 2015 Massimo Morini

Intensity directly from CDS quotation

Here is a more formal proof. Consider the CDS payoff:

P ayof f
∑
b
CDSt (Ta, Tb, S) = Lgd D (t, τ ) 1{Ta≤τ <T }−S D(t, Ti)αi1{τ >T .
b i−1 }
i=a+1
(2)
Now suppose premium payment is paid continuously. Notice we can write the default leg
with a time integral collapsing to a single instant
∫ Tb
P ayof f 1
CDSt (Ta, Tb, S) = Lgd D (t, s) 1{s∈[τ −dt,τ ]}ds
dt Ta
∫ Tb
− SD(t, s)1{τ >s}ds
Ta

9
Copyright 2015 Massimo Morini

Intensity directly from CDS quotation

Now take expectation for pricing, assuming interest rates to be deterministic or in any case
independent of default,
Price
CDSt (Ta, Tb, S) =
∫ Tb ∫ Tb
1
= Lgd P (t, s) Pr {τ − dt ≤ s ≤ τ } ds − S P (t, s) Pr {τ > s} ds
dt Ta Ta
∫ Tb ∫ Tb
1
= Lgd P (t, s) Pr {s ≤ τ ≤ s + dt} ds − S P (t, s) Pr {τ > s} ds
dt Ta Ta
∫ Tb ∫ Tb
1
= Lgd P (t, s) λdt Pr {τ > s} ds − S P (t, s) Pr {τ > s} ds
dt Ta Ta

where last passage comes from (1).

10
Copyright 2015 Massimo Morini

Intensity directly from CDS quotation

In equilibrium, when CDStPrice is zero,
∫ Tb ∫ Tb
∗ 1
S · P (t, s) Pr {τ > s} ds = Lgd · λ · dt · · P (t, s) Pr {τ > s} ds
Ta dt Ta

thus
S∗
λ= . (3)
Lgd
With a flat intensity, clearly credit spreads for CDS would be flat, with no dependence
from the maturity. We could not calibrate the CDS of different maturities.

11
Copyright 2015 Massimo Morini

Time-dependent and Stochastic Intensity

12
Copyright 2015 Massimo Morini

A term-structure of credit spreads

For obtaining time dependent credit spreads, we need to build a different jump process,
called time-inhomogeneous Poisson Process. We need to start from a different assumption:

P r [Nt+∆t − Nt = 1] = λ (t) ∆t.

implying
P r [Nt+∆t − Nt = 0] = 1 − λ (t) ∆t.

13
Copyright 2015 Massimo Morini

A term-structure of credit spreads

T −t
With T − t = m × ∆t and ∆t = m ,

∏
m−1
P r [NT − Nt = 0] = (1 − λ (t + i∆t) ∆t)
i=0

∑
m−1
ln {P r [NT − Nt = 0]} = ln (1 − λ (t + i∆t) ∆t)
i=0

∑
m−1
= −λ (t + i∆t) ∆t + o (∆t)
i=0

since
1
ln (1 − x) = ln (1 − 0) − (x − 0) + o ((x − 0)) = −x + o (x) .
(1 − 0)

14
Copyright 2015 Massimo Morini

A term-structure of credit spreads

∑
m−1
ln P r [NT − Nt = 0] = −λ (t + i∆t) ∆t + o (∆t)
i=0
∫ T
−→ − λ (s) ds,
∆t−→0 t

∫
− tT λ(s)ds
P r [NT − Nt = 0] = e
For default time, ∫
− 0T λ(s)ds
Pr [NT = 0] = e = Pr (τ > T )
So ∫
− 0T λ(s)ds
Pr (τ ≤ T ) = Fτ (T ) = 1 − e .

15
Copyright 2015 Massimo Morini

Time-dependent credit spreads

Deriving we obtain the following density for the default time
( ∫T )
∂Fτ (T ) ∫ ∂ 0 λ (s) ds
− 0T λ(s)ds
fτ (T ) = = e
∂T ∂T
∫
− 0T λ(s)ds
= λ (T ) e

and at t > 0

Pr (τ > T ) ∫
− tT λ(s)ds
P r [τ > T |τ > t] = =e ,
Pr (τ > t)
∫
− tT λ(s)ds
fτ |τ >t (T ) = λ (T ) e .

16
Copyright 2015 Massimo Morini

Time-dependent credit spreads

With time inhomogeneous Poisson process we define
∫ T
Λ (T ) = λ (s) ds
0

For τ , we have that we can write it as

−1 −1
τ =Λ (ε) = Λ (− ln(1 − U ))

with ε a unit exponential rv and U a uniform rv. The generalized inverse, that we use in
simulation also with stochastic intensity (on every single path), is
{ ∫ t }
τ = inf t: λ (s) ds ≥ ε .
0

17
Copyright 2015 Massimo Morini

Stripping from CDS quotes: a simple example

Taking yearly quotations with yearly payments, it gives sufficient flexibility to set
λ(0, 1y] = λ1, λ(1y, 2y] = λ2, . . .   
 M KT − λ1  − λ1 
CDS(0, 1, R0,1y ; λ ) = P (0, 1) R0,1y × αie
M KT 1
− Lgd × 1 − e 

[ ( )]
−λ1 −λ1
M KT
CDS(0, 1, R0,1y ; λ1, λ2) = P (0, 1) R0,2y × αie
M KT
− Lgd × 1 − e +
  
 M KT −λ1 − λ2  −λ1 −λ1 − λ2 
+P (0, 2) R0,2y × αie − Lgd × e −e 

M KT
Solve CDS(0, 1, R0,1y ; λ1 ) = 0

M KT
and so find λ1. With this λ1 solve CDS(0, 1, R0,2y ; λ1, λ2 ) = 0
and so find λ2...

18
Copyright 2015 Massimo Morini

September 10, PW Constant

Maturity (dates) Maturity Tb (yr) R0,b
20-Sep-04 1 192.5
20-Sep-06 3 215
20-Sep-08 5 225
20-Sep-10 7 235
20-Sep-13 10 235

Table 1: CDS quotes in bps for September 10th, 2003. Recovery= 40%.

date intensity γ survival pr exp(−Γ)

0 - 100.000%
1 3.199% 96.714%
3 3.780% 89.578%
5 4.033% 82.516%
7 4.458% 75.402%
10 3.891% 66.978%

Table 2: Calibration with piecewise constant intensity on September 10th, 2003.

19
Copyright 2015 Massimo Morini

Figure 1: Piecewise constant intensity λ calibrated on CDS quotes on September 10th,

2003. 20
Copyright 2015 Massimo Morini

September 10, PW Linear

Maturity Tb (yr) Maturity (dates) R0,b
1 20-Sep-04 192.5
3 20-Sep-06 215
5 20-Sep-08 225
7 20-Sep-10 235
10 20-Sep-13 235

Table 3: CDS quotes in bps for September 10th, 2003. Recovery= 40%.

date intensity γ survival pr exp(−Γ)

0 - 100.000%
1 3.199% 96.714%
3 4.388% 89.552%
5 3.659% 82.508%
7 5.308% 75.357%
10 2.338% 67.078%

Table 4: Calibration with piecewise linear intensity on September 10th, 2003.

21
Copyright 2015 Massimo Morini

Figure 2: Piecewise linear intensity λ calibrated on CDS quotes on September 10th, 2003.

22
Copyright 2015 Massimo Morini

Figure 3: survival probability exp(−Λ) resulting from calibration on CDS quotes on

September 10th, 2003. 23
Copyright 2015 Massimo Morini

Stochastic Credit spread

After summer ’07, modelling the volatility of credit spreads can be a need for realistic single
name models. Modelling the volatility of credit spread amounts to assume that credit
spreads, and in turn default intensities, are stochastic. Poisson processes with stochastic
intensity are known as Cox Processes

λ −→ λ (t) −→ λ (t, ω)
Poisson Inhomogeneous
Cox Process
Process Poisson Process

24
Copyright 2015 Massimo Morini

Stochastic Credit spread

Definition 1 A process Nt is a Cox process with stochastic intensity λ (t) = λ (ω, t) if,
conditional on knowing the path of λ (t) from 0 to T , called λ [0, T ), then (Nt)0≤t≤T
is an inhomogeneous Poisson Process with intensity λ (t).

∫
− tT λ(s)ds
P r (NT − Nt = 0| λ [t, T )) = e
How to find standard, unconditional probability? We need to recall now the Law of
iterated expectation. When GBig represents a set of information larger than GSmall , we
have
E [E [X | GBig ] | GSmall ] = E [X | GSmall ] .

Following the law of iterated expectations, this implies

r (NT − Nt =] 0) =[ [
P[ ]]
E 1{N −N =0} = E E 1{N λ [t, T )
T t T −Nt =0}
E [Pr (NT − Nt = 0|λ [t, T ))]
[ ∫T ]
− t λ(s)ds
=E e .

25
Copyright 2015 Massimo Morini

Stochastic Credit spread

The above definition of Cox Processes guarantees high tractability but constraints our
freedom of modelling.

More formally, in credit modelling it can be meaningful to separate default free information
from information containing the default event:
τ
Ft = Ht ∨ Ft

• Ft = Ht ∨ Ftτ = all available information up to t

• Ftτ = σ({τ < u}, u ≤ t) = information up to t on the default event: if it has
already happened or not, and in the former case the exact time τ of default
• Ht =information up to t on economic quantities which affect default probability, but
no specific information on happening of default

Definition 2 A process Nt is a Cox process with intensity λ (t) = λ (ω, t) if, conditional
on HT , (Nt)0≤t≤T is an inhomogeneous Poisson Process with intensity λ (t) .

26
Copyright 2015 Massimo Morini

Credit Derivatives Modelling. Approaches

τ
Ft = Ht ∨ Ft

• Ft = all available information up to t

• Ftτ = σ({τ < u}, u ≤ t) = information up to t on the default event: if it has
already happened or not, and in the former case the exact time τ of default
• Ht =information up to t on economic quantities which affect default probability, but
no specific information on happening of default

In standard structural models the distinction is meaningless, we would have: Ht = Ft

since same economic quantities inform both on default probabilities and default time
Instead in standard reduced form models the distinction is fundamental: Ht ⊂ Ft
since default is exogenous to economic quantities, which only affect default probability.

27
Copyright 2015 Massimo Morini

Stochastic Credit spread: a practical example

[ ∫T ]
− t λ(s)ds
In a practical model, one would like Et e to have an analytic solution. Notice
it has got the same form as the bond price in short interest rate models. Thus spread
dynamics is often modelled analogously to the instantaneous spot rate in early interest rate
models. If we choose the CIR (Cox, Ingersoll, Ross) dynamics
√
dλ (t) = k[θ − λ (t)]dt + σ λ (t)dW (t),

one can guarantee positivity of the spreads through σ 2 < 2kθ , and the survival probability
can be explicitly computed as
[ ( ∫ )]
T
−B(t,T )λ(t)
Et exp − λ (u) du = A(t, T )e ,
t

28
Copyright 2015 Massimo Morini

Stochastic Credit Spreads in Practice: Intensity Models

[ ( ∫ )]
T
−B(t,T )λ(t)
Et exp − λ (u) du = A(t, T )e ,
t

[ ]2kθ/σ2
2h exp [(k + h) (T − t) /2]
A(t, T ) =
2h + (k + h) [exp ((T − t) h) − 1]
2 [exp ((T − t) h) − 1] (4)
B(t, T ) =
2h + (k + h) [exp ((T − t) h) − 1]
√
h = k2 + 2σ 2
This makes calibration to CDS very easy. Since most of the times there are no liquid
options on CDS, CDS calibration is the only calibration available. The model has four
parameters,
λ (0) , θ, k, σ
so a standard CDS term structure can be sufficient for stable calibration.

29
Copyright 2015 Massimo Morini

Stochastic Credit Spreads in Practice: Intensity Models

Calibration - CIR Model

70

60

50

40
Market
Model
30

20

10

0
0 2 4 6 8 10

Then, complex derivatives can be priced through montecarlo simulation, as we will see in
the practical example.

30
Copyright 2015 Massimo Morini

31