MScFE 622 Continuous-time Stochastic Processes - Module 6: Problem Set

Compiled Notes
Module 6
MScFE 622
Continuous-time Stochastic
Processes

Revised: 07/07/2020
MScFE 622 Continuous-time Stochastic Processes − Notes Module 6

Module 6: An Introduction to Levy Processes

Contents
Unit 1: The Poisson Process 3

Unit 2: Levy Processes and their Properties 5

Unit 3: Exponential Levy Models 9

Problem Set 11

2019 - WorldQuant University − All rights reserved. 1

MScFE 622 Continuous-time Stochastic Processes − Summary Module 6

Summary
Module 6 introduces Levy processes and demonstrates its application to modeling stock price returns.
The module begins with an introduction to the Poisson process and then continues by discussing Levy
processes and developing proofs of its properties. At the end of the module, applications of Levy
processes to financial modeling are discussed, with an emphasis on exponential Levy models.

2019 - WorldQuant University − All rights reserved. 2

MScFE 622 Continuous-time Stochastic Processes − Notes (1) Module 6: Unit 1

Unit 1: The Poisson Process

• Before introducing Levy processes in general, we will first talk about a very important dis-
continuous stochastic process: the Poisson process. Throughout this section, we fix a filtered
probability space (Ω, F, F, P) and assume that all stochastic processes defined on this space are
adapted to F.

• A stochastic process N = {Nt : t ≥ 0} is called a counting process if N is cadlag and the sample
paths of N are piecewise constant with jumps of size 1. We will also assume that N0 = 0.

• So we can think of a counting process N as a stochastic process such that Nt counts the
number of events that have occurred up to (and including) time t and the increment Nt − Ns
(for 0 ≤ s < t) counts the number of events that have occurred in the interval (s, t]. Every
sample path of a counting process N will move through the states N = {0, 1, 2, 3 . . .} in that
order.

• A counting process N is called a homogeneous Poisson process with rate λ > 0 if

1. N has independent increments

2. N has stationary increments
3. For s < t, the increment Nt − Ns has a Poisson distribution with parameter λ (t − s); that
is,
∞
X (λ(t − s))n e−λ(t−s)
PNt −Ns = δn .
n=0
n!

• For the remainder of this section, let N be a homogeneous Poisson process with rate λ > 0.
Define the random variables (stopping times) S0 , S1 , . . . as follows:

S0 := 0, Sn := inf {t ≥ 0 : Nt = n} n ≥ 1.

The Sn ’s are called the arrival times of N ; that is, Sn is the time of arrival of the nth event.
Note that S0 ≤ S1 ≤ S2 ≤ . . ..

• We also define the interarrival times T1 , T2 , . . . as

Tn := Sn − Sn−1 , n ≥ 1.

These random variables represent the time between successive events of N . For n ≥ 1, the
arrival times can be recovered as n
X
Sn = Ti .
i=1

• Let us now find the distribution of Tn , and consequently, that of Sn for n ≥ 1. First note the
equivalence of the following events:

{Nt = 0} = {T1 > t} = {S1 > t} .

The intuitive explanation of this relationship is that if no events have occurred by time t (i.e.,
Nt = 0), then the arrival of the first event is after time t (i.e., S1 > t or T1 > t since S1 = T1 )

2019 - WorldQuant University − All rights reserved. 3

MScFE 622 Continuous-time Stochastic Processes − Notes (1) Module 6: Unit 1

and vice versa. Since T1 is a non-negative random variable and Nt has a Poisson distribution
with parameter λt, we have (for t ≥ 0)
1 − FT1 (t) = P (T1 > t) = P (Nt = 0) = e−λt .
Hence T1 (and S1 ) has an exponential distribution with parameter λ > 0.
• Let us now find the joint distribution of S1 and S2 . For 0 ≤ w1 ≤ w2 we have
FS1 S2 (w1 , w2 ) = P (S1 ≤ w1 , S2 ≤ w2 ) = P (Nw1 ≥ 1, Nw2 ≥ 2)
= P (Nw1 = 1, Nw2 − Nw1 ≥ 1) + P (Nw1 ≥ 2)
−λw1
1 − e−λ(w2 −w1 ) + 1 − λw1 e−λw1 − e−λw1 = 1 − λw1 e−λw2 − e−λw1 .



= λw1 e
Hence the joint density of S1 and S2 is
(
λ2 e−λw2 0 ≤ w1 ≤ w2
fS1 S2 (w1 , w2 ) =
0 otherwise.

• To find the joint density of T1 and T2 we note that

T1 = S1 and T2 = S2 − S1 .
Hence, (
λ2 e−λ(t1 +t2 ) t1 , t2 ≥ 0
fT1 T2 (t1 , t2 ) = fS1 S2 (t1 , t1 + t2 ) × 1 =
0 otherwise.
Therefore, T1 and T2 are both independent exponential random variables with parameter λ.
• It can be shown that in general, the interarrival times T1 , T2 , . . . are i.i.d. exponential ran-
dom variables with parameter λ > 0. It then follows that the arrival time Sn has a gamma
distribution with parameters λ and n; that is,
λn wn−1 e−λw
fSn (w) = , w > 0.
(n − 1) !
This is easily shown by calculating the moment generating function of Sn :
n
!! n
! n
X Y Y
MSn (α) = E eαSn = E exp α eαTi = E eαTi



Ti =E
i=1 i=1 i=1
n    n
Y λ λ
= .
i=1
λ−α λ−α

• We can also define the Poisson process by starting with an i.i.d sequence of exponential random
variables T1 , T2 , . . . with parameter λ > 0. Then we define the arrival times as
S0 = 0, Sn = Sn−1 + Tn n ≥ 1.
Then the counting process N defined by
∞
X
Nt = I{Sn ≤t} = # {n : Sn ≤ t}
n=1

is a Poisson process.

2019 - WorldQuant University − All rights reserved. 4

MScFE 622 Continuous-time Stochastic Processes − Notes (2) Module 6: Unit 2

Unit 2: Levy Processes and their Properties

• We now introduce Levy processes.

• Let X = {Xt : t ≥ 0} be a stochastic processes (adapted to F). We say that X is a Levy process
if

1. X0 = 0
2. X has independent increments
3. X has stationary increments
4. X is stochastically continuous; i.e., for every t ≥ 0 and every  > 0,

lim P (|Xs − Xt | > ) = 0.

s→t

So, Xs converges to Xt in probability as s tends to t.

• It can be shown that if X is a Levy process, then X has a cadlag modification, and because of
that we will simply assume that X is cadlag.

• A trivial example of a Levy process is the deterministic process Xt = bt where b ∈ R. It is

trivial to show that all the above properties are satisfied.

• Another example of a Levy process is a Brownian motion process. Indeed, if X = W is a

Brownian motion, then properties 1,2 and 3 clearly hold and property 4 holds due to continuity
of the sample paths of X. Combining the two examples, it follows that the so-called Brownian
motion with drift Xt = bt + σWt is also a Levy process. It is the only Levy process with
continuous sample paths.

• The (homogeneous) Poisson process discussed above is a Levy process. Again, the first three
properties are clearly satisfied. For the 4th property, we use Markov’s inequality (or Cheby-
chev’s inequality) to obtain (for  > 0)

E (|Ns − Nt |) λ |s − t|
P (|Ns − Nt | > ) ≤ = → 0 as s → t.
 
The paths of N are of course discontinuous, with jumps of size 1 (i.e., ∆N ∈ {0, 1}).

• One drawback of using the Poisson process to model stock price returns is that the sizes of the
jumps are always equal to 1, which is not realistic. We now look at a generalization of this by
defining what is called a compound Poisson process.

• Let N be a homogeneous Poisson process with rate λ > 0 and Y1 , Y2 , . . . be a sequence of i.i.d.
random variables that are also independent of N . A compound Poisson process is a stochastic
process X defined by
∞
(
X 0 Nt = 0
Xt := Yn I{Nt ≥n} = PNt
n=1 k=1 Yk Nt ≥ 1.

• We can think of X as a generalization of the Poisson process, where the sizes of the jumps are
random variables, instead of just being all equal to 1.

2019 - WorldQuant University − All rights reserved. 5

MScFE 622 Continuous-time Stochastic Processes − Notes (2) Module 6: Unit 2

• The compound Poisson process is a Levy process. The full proof of this fact is left as an
exercise, but we will show the stationarity of increments. Consider the times 0 ≤ s < t. Then
the characteristic function of the increment is
 P Nt    P Nt 
ϕXt −Xs (u) = E eiu(Xt −Xs ) = E eiu( k=Ns +1 Yk ) = E E eiu( k=Ns +1 Yk ) |σ (Ns , Nt )

= E ϕY (u)Nt −Ns = exp (λ(t − s) (ϕY (u) − 1))

since Nt − Ns has a Poisson distribution with rate λ(t − s). So, clearly the distribution of the
increment only depends on t − s.

• Thus, so far our most general Levy process X is

Nt
X
Xt = bt + σWt + Yk .
k=1

The process consists of the following three terms:

– A deterministic term bt
– A diffusion term σWt
– A pure jump term N
P t
k=1 Yk

The characteristic function of Xt is

Nt
!!
X
iuXt iubt iuσWt



ϕXt (u) = E e =e E e E exp iu Yk
k=1
     Z 
1 2 2 1 2 2 iuy


exp t iub − σ u + λ (ϕY (u) − 1) = exp t iub − σ u + λ e − 1 dPY (y) .
2 2 R

• It turns out that the characteristic function of any Levy process is similar to the one above.
Let X be a Levy process and for a fixed u ∈ R, define gu : [0, ∞) → C to be the characteristic
function of Xt :
gu (t) := E eiuXt , t ≥ 0.

For s, t > 0, we have

gu (s + t) = E eiuXs+t = E eiu(Xt +Xs+t −Xt ) = gu (t)gu (s).

Together with gu (0) = 1, we get that

gu (t) = etψ(u)

for some function ψ : R → C.

• The function ψ above is given by the following theorem

Theorem 1 (Levy-Khintchine). Let X be a Levy process. Then the characteristic function of

Xt is given by
ϕXt (u) = etψ(u) ,

2019 - WorldQuant University − All rights reserved. 6

MScFE 622 Continuous-time Stochastic Processes − Notes (2) Module 6: Unit 2

where Z
1
ψ(u) = ibu − σ 2 u2 + eiuy − 1 − iuyI[−1,1] (y) dν(y)
2 R
2
for some b ∈ R, σ ≥ 0 and a measure on B(R) called the Levy measure of X, that satisfies
Z
ν({0}) = 0, and 1 ∧ y 2 dν(y) < ∞.
R

We call (b, σ 2 , ν) the Levy triplet of X.

• To understand how to find ν, we need to introduce the following Poisson random measure. Let
B ∈ B(R) be a Borel subset of R such that 0 ∈/ B̄, where B̄ is the closure of B. Define for each
t ≥ 0 define the random variable Nt (B) by
X
Nt (B)(ω) := IB (∆Xs (ω)) = # {s ≤ t : ∆Xs (ω) ∈ B} .
s≤t

That is, Nt (B)(ω) is the number of jumps of X that are of size B. This random variable is
well-defined since X is cadlag, and therefore, has finitely many jumps of a given size in a finite
interval. Now note that
– For fixed ω ∈ Ω, B 7→ Nt (B)(ω) is a positive measure (a counting measure)
– For fixed B ∈ B(R) with 0 ∈
/ B̄, ω 7→ Nt (B)(ω) is a random variable
– For fixed B ∈ B(R) with 0 ∈
/ B̄, (t, ω) 7→ Nt (B)(ω) is a counting process.
In fact, N· (B) is a Poisson process. Indeed, define the stopping times τ0 < τ1 < τ2 < . . . by
τ0 = 0, τn+1 := inf {t > τn : ∆Xt ∈ B} .
Then N (B) can be written as
∞
X
Nt (B) = I{τn ≤t} .
n=1
Hence all we need to show is that the interarrival times τn+1 − τn are i.i.d exponentially dis-
tributed. This is achieved by showing that the distribution of τn+1 − τn is memoryless, i.e.,
P (τn+1 − τn > s + t) = P (τn+1 − τn > t) P (τn+1 − τn > s) s, t > 0.
This is left as an exercise.
• So we get that N (B) is a Poisson process with rate E(N1 (B)) =: ν(B). This is how the Levy
measure is obtained.
• We now move on to the infinite divisibility property of Levy process. Let X be a Levy process
and t > 0 be fixed. Then for each n ≥ 1 we can write Xt as
    X n
Xt = Xt − X t(n−1) + . . . + X nt − X0 = Zi ,
n
i=1

where Zi := X ti − X t(i−1) for i = 1, . . . , n. Since X has stationary and independent increments,

n n
it follows that the Zi s are i.i.d random variables. So we can conclude that for each n ≥ 1, Xt is
equal (in distribution) to the sum of n i.i.d random variables. Such a random variable is said
to be infinitely divisible.

2019 - WorldQuant University − All rights reserved. 7

MScFE 622 Continuous-time Stochastic Processes − Notes (2) Module 6: Unit 2

• Here are some examples of distributions that are infinitely divisible:

2
– If Y ∼ N (µ, σ 2 ), then Y = ni=1 Zi where Zi ∼ N ( nµ , σn )
P

– If Y ∼ Gamma(α, λ), then Y = ni=1 Zi where Zi ∼ Gamma(α/n, λ)

P

– If Y ∼ Poisson(λ), then Y = ni=1 Zi where Zi ∼ Poisson(λ/n)

P

• Let Y be a random variable whose distribution is infinitely divisible. Then the characteristic
function of Y is given by
ϕY (u) = eψ(u)
where Z
1
ψ(u) = ibu − σ 2 u2 + eiuy − 1 − iuyI[−1,1] (y) dν(y)
2 R

for some b ∈ R, σ 2 ≥ 0 and a measure on B(R) called the Levy measure of X, that satisfies
Z
ν({0}) = 0, and 1 ∧ y 2 dν(y) < ∞.
R

• The next theorem relates infinitely divisible distributions to Levy processes.

Theorem 2. Let X be a Levy process. Then the distribution of Xt is infinitely divisible.

Conversely, if Y is a random variable whose distribution is infinitely divisible, then there exists
a Levy process X such that the law of X1 is the same as the law of Y .

• For example, the Levy process X such that X1 has a Gamma distribution is called a Gamma
process.

2019 - WorldQuant University − All rights reserved. 8

MScFE 622 Continuous-time Stochastic Processes − Notes (3) Module 6: Unit 3

Unit 3: Exponential Levy Models

• We now illustrate how to apply Levy processes to financial modelling. We will consider models
for the stock price S that are of the form St = S0 eXt or St = S0 E(X)t , where X is a Levy process.
Here E(X) is the stochastic exponential of X, which we will define below. An important result
that we will use here is that every Levy process is a semimartingale.

• We have seen such a model in the previous module (the Black-Scholes model), where S satisfies
the following SDE:
dSt = St (µ dt + σ dWt ) = St dXt ,
where Xt := µt + σWt is a Levy process. We can write S as
1 2
St = S0 eXt − 2 σ t = S0 E(X)t .

• To deal with general discontinuous Levy processes, we need to first introduce the stochastic
calculus for general semimartingales. We state only the one-dimensional version of Ito’s formula;
the multidimensional version can be found in any standard reference for stochastic calculus.

Theorem 3 (Ito’s Formula). Let X be a semimartingale and f : R → R be a twice continuously

differentiable function. Then f (X) is also a semimartingale and
Z t
1 t 00 X
Z 
0 0 1 00 2
f (Xt ) = f (X0 )+ f (Xs− ) dXs + f (Xs− ) d [X]s + ∆f (Xs ) − f (Xs− )∆Xs − f (Xs− )∆Xs
0 2 0 s≤t
2

Here [X] is the (optional) quadratic variation of X, rather than the predictable quadratic
variation of X denoted by hXi. The latter is not even defined for all semimartingales, but the
two are equal for continuous semimartingales.

• We now introduce the stochastic exponential of a general semimartingale X. Consider the

following stochastic differential equation for Y :

dY = Y− dX, Y0 = 1.

That is, we want to find a process Y that satisfies the equation

Z t
Yt = 1 + Ys− dXs .
0

Substituting Ut = ln Yt and applying Ito’s lemma gives

 Y  
1 1 2
Yt = exp Xt − X0 − [X]t (1 + ∆Xs ) exp −∆Xs + ∆Xs .
2 s≤t
2

We will denote this process by E(X).

• In general, exponential Levy models give rise to incomplete markets. Consider the following
popular jump-diffusion type model for S:

dSt = St− (µ dt + σ dWt + dJt )

2019 - WorldQuant University − All rights reserved. 9

MScFE 622 Continuous-time Stochastic Processes − Notes (3) Module 6: Unit 3

where
Nt
X
Jt = Yk
k=1

is a compound Poisson process with Yk > −1. The solution to this SDE is

St = S0 E(X)t

where Xt := µt + σWt + Jt . The solution is

   
1 2
Y 1 2
St = S0 exp Xt − σ t + [J]t (1 + ∆Xs ) exp −∆Xs + ∆Xs
2 s≤t
2

N N
1 2 Yt 1 2 Yt
 
= S0 exp Xt − σ t (1 + Yk ) exp (−Yk ) = S0 exp Xt − Jt − σ t (1 + Yk )
2 k=1
2 k=1
  Nt
!
1 2 X
= S0 exp µ − σ t + σWt + ln (1 + Yk )
2 k=1

since
X X Nt
X
[J]t = ∆Js2 = ∆Xs2 = Yk2 .
s≤t s≤t k=1

2019 - WorldQuant University − All rights reserved. 10

MScFE 622 Continuous-time Stochastic Processes − Problem Set Module 6

Problem Set

Problem 1. Let the number of accidents N occur according to a Poisson process with rate
λ = 5 per day. What is the expected number of accidents between the fifth day and the seventh
day?

Solution: Since N has a Poisson distribution, E(N ) = λt. That is, we expect λt accidents in
t times units. Thus, in our example, the expected number of accidents between the fifth day
and the seventh day is equat to 5 ∗ 2 = 10.

Problem 2. Let the number of accidents N occur according to a Poisson process with rate
λ = 2 per day. What is the probability the number of accidents between the third day and the
fourth day is 3?

Solution: Let start with the probability,

P (N4 − N3 = 3) = P (N1 = 3)
We also know from the lecture notes that,

e−λt (λt)k
P (Nt = k) =
k!
So, finally we get,

e−2 23
P (N4 − N3 = 3) = P (N1 = 3) =
3!
Which is the solution to the problem.

Problem 3. Let the number of accidents N occur according to a Poisson process with rate
λ = 2 per day. Given that no accidents have occurred in the last 3 days, what is the probability
that the next accident occurs within the next day?

Solution: The probability the the next accident occurs between day four and three is,

e−λt (λt)k
P (N4 − N3 > 0) = P (N1 > 0) = 1 − P (N1 ≤ 0) = 1 −
k!
with k = 0, λ = 2 and t = 1. The solution is equal to,

e−λt (λt)k
P (N4 − N3 > 0) = P (N1 > 0) = 1 − P (N1 ≤ 0) = 1 − = 1 − e−2
k!

2019 - WorldQuant University − All rights reserved. 11

MScFE 622 Continuous-time Stochastic Processes − Problem Set Module 6

Problem 4. Let N be a Poisson process with rate λ = 2 and Y1 , Y2 , . . . be i.i.d. normal random
variables with mean µ = 2 and variance σ 2 = 1. Define the compound Poisson process X by
Nt
X
Xt := Yk .
k=1

What is E(X2 )?

Solution: The expected value, considering that Y1 , Y2 , . . . are i.i.d. normal random variables
can be compute as follows,

Nt
X
E[Xt ] := E[ Yk ] = E[Nt ] ∗ E[Y ] = λt ∗ µ
k=1

For t = 2, we will have: E[X2 ] = 2 ∗ 2 ∗ 2 = 8.

Problem 5. Let N be a Poisson process with rate λ = 6 and Y1 , Y2 , . . . be i.i.d. uniform

random variables between 0 and 2. Define the compound Poisson process X by
Nt
X
Xt := Yk .
k=1

What is the variance of X1 ?

Solution: We can compute the variance as follows,

Var(X(t)) = E[Var(X(t)) | N (t)] + Var(E[X(t) | N (t)])

= E[N (t)Var(Y1 )] + Var(N (t)E[Y1 ])
= σ 2 E[N (t)] + m2 Var(N (t))
= σ 2 λt + m2 λt
= (σ 2 + m2 )λt.
Applying the above expression to our problem, we will get,

1
Var(X(t)) = (σ 2 + m2 )λt = ( + 1) ∗ 6 ∗ 1 = 8
3

Problem 6. Let X be a Levy process such that X2 ∼ Gamma(α = 20, λ = 2). If Xt + bt is a

martingale, then compute the value of b.

Solution: We need to compute b such that the expected value of Xt + bt, for t = 2, is equal to
zero. Thus,

20
E[X2 + 2b] = E[X2 ] + 2b = + 2b = 0 , then b = −5
2

2019 - WorldQuant University − All rights reserved. 12