© 2004 Kazuhisa Matsuda All rights reserved.

Introduction to Merton Jump Diffusion Model

Kazuhisa Matsuda

Department of Economics
The Graduate Center, The City University of New York,
365 Fifth Avenue, New York, NY 10016-4309
Email: [email protected]
http://www.maxmatsuda.com/

December 2004

Abstract

This paper presents everything you need to know about Merton jump diffusion (we call it
MJD) model. MJD model is one of the first beyond Black-Scholes model in the sense that
it tries to capture the negative skewness and excess kurtosis of the log stock price density
P ( ln( ST / S0 ) ) by a simple addition of a compound Possion jump process. Introduction of
this jump process adds three extra parameters λ , µ , and δ (to the original BS model)
which give the users to control skewness and excess kurtosis of the P ( ln( ST / S0 ) ) .
Merton’s original approach for pricing is to use the conditional normality of MJD model
and expresses the option price as conditional Black-Scholes type solution. But modern
approach of its pricing is to use the Fourier transform method by Carr and Madan (1999)
which is disccused in Matsuda (2004).

© 2004 Kazuhisa Matsuda All rights reserved.

1
© 2004 Kazuhisa Matsuda All rights reserved.

[1] Model Type

In this section the basic structure of MJD model is described without the derivation of the
model which will be done in the next section.

MJD model is an exponential Lévy model of the form:

St = S0 e Lt ,

where the stock price process {St ;0 ≤ t ≤ T } is modeled as an exponential of a Lévy

process {Lt ;0 ≤ t ≤ T } . Merton’s choice of the Lévy process is a Brownian motion with
drift (continuous diffusion process) plus a compound Poisson process (discontinuous
jump process) such that:

σ2 Nt
Lt = (α − − λ k )t + σ Bt + ∑ Yi ,
2 i =1

where {Bt ; 0 ≤ t ≤ T } is a standard Brownian motion process. The term

σ2
(α − − λ k )t + σ Bt is a Brownian motion with drift process and the term ∑
Nt
Y is ai =1 i
2
compound Poisson jump process. The only difference between the Black-Scholes and the
MJD is the addition of the term ∑ i =t1 Yi . A compound Poisson jump process ∑ i =t1 Yi
N N

contains two sources of randomness. The first is the Poisson process dN t with intensity
(i.e. average number of jumps per unit of time) λ which causes the asset price to jump
randomly (i.e. random timing). Once the asset price jumps, how much it jumps is also
modeled random (i.e. random jump size). Merton assumes that log stock price jump size
follows normal distribution, ( dxi ) ∼ i.i.d . Normal ( µ , δ 2 ) :

1 (dxi − µ ) 2
f (dxi ) = exp{− }.
2πδ 2 2δ 2

It is assumed that these two sources of randomness are independent of each other. By
introducing three extra parameters λ , µ , and δ to the original BS model, Merton JD
model tries to capture the (negative) skewness and excess kurtosis of the log return
density P ( ln ( St / S 0 ) ) which deviates from the BS normal log return density.

Lévy measure ( dx ) of a compound Poisson process is given by the multiplication of the

intensity and the jump size density f ( dx ) :

( dx ) = λ f ( dx ) .

2
© 2004 Kazuhisa Matsuda All rights reserved.

A compound Poisson process (i.e. a piecewise constant Lévy process) is called finite
activity Lévy process since its Lévy measure ( dx ) is finite (i.e. the average number of
jumps per unit time is finite):

∞
∫−∞
(dx) = λ < ∞ .

The fact that an asset price St is modeled as an exponential of Lévy process Lt means
S
that its log-return ln( t ) is modeled as a Lévy process such that:
S0

St σ2 Nt
ln( ) = Lt = (α − − λ k )t + σ Bt + ∑ Yi .
S0 2 i =1

Let’s derive the model.

[2] Model Derivation

In MJD model, changes in the asset price consist of normal (continuous diffusion)
component that is modeled by a Brownian motion with drift process and abnormal
(discontinuous, i.e. jump) component that is modeled by a compound Poisson process.
Asset price jumps are assumed to occur independently and identically. The probability
that an asset price jumps during a small time interval dt can be written using a Poisson
process dN t as:

Pr {an asset price jumps once in dt } = Pr{ dN t = 1 } ≅ λ dt ,

Pr {an asset price jumps more than once in dt } = Pr{ dNt ≥ 2 } ≅ 0 ,

Pr {an asset price does not jump in dt } = Pr{ dNt = 0 } ≅ 1 − λ dt ,

where the parameter λ ∈ R + is the intensity of the jump process (the mean number of
jumps per unit of time) which is independent of time t .

Suppose in the small time interval dt the asset price jumps from St to yt St (we call yt
as absolute price jump size). So the relative price jump size (i.e. percentage change in the
asset price caused by the jump) is:

dSt yt St − St
= = yt − 1 ,
St St

3
© 2004 Kazuhisa Matsuda All rights reserved.

where Merton assumes that the absolute price jump size yt is a nonnegative random
variables drawn from lognormal distribution, i.e. ln( yt ) ∼ i.i.d .N ( µ , δ 2 ) . This in turn
1
µ+ δ 2
implies that E[ yt ] = e and E[( yt − E[ yt ]) 2 ] = e 2 µ +δ (eδ − 1) . This is because if
2 2
2

1
a + b2
, e2 a +b (eb − 1)) .
2 2
ln x ∼ N ( a, b) , then x ∼ Lognormal (e 2

MJD dynamics of asset price which incorporates the above properties takes the SDE of
the form:

dSt
= (α − λ k )dt + σ dBt + ( yt − 1) dN t , (1)
St

where α is the instantaneous expected return on the asset, σ is the instantaneous

volatility of the asset return conditional on that jump does not occur, Bt is a standard
Brownian motion process, and N t is an Poisson process with intensity λ . Standard
assumption is that ( Bt ) , ( Nt ) , and ( yt ) are independent. The relative price jump size
1
µ+ δ 2
of St , yt − 1 , is lognormally distributed with the mean E[ yt − 1] = e 2
− 1 ≡ k and the
variance E[( yt − 1 − E[ yt − 1]) 2 ] = e 2 µ +δ (eδ − 1) .1 This may be confusing to some readers,
2 2

so we will repeat it again. Merton assumes that the absolute price jump size yt is a
lognormal randon variable such that:

1
µ+ δ 2
, e 2 µ +δ (eδ − 1)) .
2 2
( yt ) ∼ i.i.d .Lognormal (e 2
(2)

This is equivalent to saying that Merton assumes that the relative price jump size yt − 1 is
a lognormal random variable such that:

1
µ+ δ 2
( yt − 1) ∼ i.i.d .Lognormal (k ≡ e − 1, e 2 µ +δ (eδ − 1)) .
2 2
2
(3)

This is equivalent to saying that Merton assumes that the log price jump size ln yt ≡ Yt is
a normal random variable such that:

ln( yt ) ∼ i.i.d .Normal ( µ , δ 2 ) . (4)

yt St
This is equivalent to saying that Merton assumes that the log-return jump size ln( ) is
St
a normal random variable such that:

1
For random variable x , Variance[ x − 1] = Variance[ x] .

4
© 2004 Kazuhisa Matsuda All rights reserved.

yt St
ln( ) = ln( yt ) ≡ Yt ∼ i.i.d .Normal ( µ , δ 2 ) . (5)
St

It is extremely important to note:

1
µ+ δ 2
E[ yt − 1] = e 2
−1 ≡ k ≠ E[ln( yt )] = µ ,

because ln E[ yt − 1] ≠ E[ln( yt − 1)] = E[ln( yt )] .

The expected relative price change E[dSt / St ] from the jump part dN t in the time
interval dt is λ kdt since E[( yt − 1)dN t ] = E[ yt − 1]E[dN t ] = k λ dt . This is the predictable
part of the jump. This is why the instantaneous expected return on the asset α dt is
adjusted by −λ kdt in the drift term of the jump-diffusion process to make the jump part
an unpredictable innovation:

dSt
E[ ] = E[(α − λ k )dt ] + E[σ dBt ] + E[( yt − 1) dN t ]
St
dS
E[ t ] = (α − λ k )dt + 0 + λ kdt = α dt .
St

Some researchers include this adjustment term for predictable part of the jump −λ kdt in
the drift term of the Brownian motion process leading to the following simpler (?)
specification:

dSt
= α dt + σ dBt + ( yt − 1)dN t
St
λ kt
Bt ∼ Normal ( − ,t)
σ
dS λ kdt
E[ t ] = α dt + σ (− ) + λ kdt = α dt .
St σ

But we choose to explicitely subtract λ kdt from the instantsneous expected return α dt
because we prefer to keep Bt as a standard (zero-drift) Brownian motion process. Realize
that there are two sources of randomness in MJD process. The first source is the Poisson
Process dN t which causes the asset price to jump randomly. Once the asset price jumps,
how much it jumps (the jump size) is also random. It is assumed that these two sources of
randomness are independent of each other.

If the asset price does not jump in small time interval dt (i.e. dNt = 0 ), then the jump-
diffusion process is simply a Brownian motion motion with drift process:

5
© 2004 Kazuhisa Matsuda All rights reserved.

dSt
= (α − λ k ) dt + σ dBt .
St

If the asset price jumps in dt ( dN t = 1 ):

dSt
= (α − λ k ) dt + σ dBt + ( yt −1) ,
St

the relative price jump size is yt − 1 . Suppose that the lognormal random drawing yt is
0.8 , the asset price falls by 20%.

Let’s solve SDE of (1). From (1), MJD dynamics of an asset price is:

dSt = (α − λ k ) St dt + σ St dBt + ( yt − 1) St dN t .

Cont and Tankov (2004) give the Itô formula for the jump-diffusion process as:

∂f ( X t , t ) ∂f ( X t , t ) σ 2 ∂2 f ( X t , t)
df ( X t , t ) = dt + bt dt + t dt
∂t ∂x 2 ∂x 2
∂f ( X t , t )
+σ t dBt + [ f ( X t − + ∆X t ) − f ( X t − )] ,
∂x

where bt corresponds to the drift term and σ t corresponds to the volatility term of a
Nt
jump-diffusion process X t = X 0 + ∫ bs ds + ∫ σ s dBs + ∑ ∆X i . By applying this:
t t

0 0
i =1

∂ ln St ∂ ln St σ 2 St 2 ∂ 2 ln St
d ln St = dt + (α − λ k ) St dt + dt
∂t ∂St 2 ∂St 2
∂ ln St
+σ St dBt + [ln yt St − ln St ]
∂St
1 σ 2 St 2 ⎛ 1 ⎞ 1
d ln St = (α − λ k ) St dt + ⎜ − 2 ⎟ dt + σ St dBt + [ln yt + ln St − ln St ]
St 2 ⎝ St ⎠ St
σ2
d ln St = (α − λ k )dt − dt + σ t dBt + ln yt
2
σ2
ln St − ln S0 = (α − − λ k )(t − 0) + σ t ( Bt − B0 ) + ∑ i =t1 ln yi
N

2
σ2
ln St = ln S0 + (α − − λ k )t + σ t Bt + ∑ i =t1 ln yi
N

6
© 2004 Kazuhisa Matsuda All rights reserved.

⎧ σ2 ⎫
exp ( ln St ) = exp ⎨ln S0 + (α − − λ k )t + σ t Bt + ∑ i =t1 ln yi ⎬
N

⎩ 2 ⎭
⎧⎪⎛ ⎫⎪
St = S0 exp ⎨⎜ α −
⎩⎪⎝
σ2
2
⎞
⎠ ⎭⎪
N
(
− λ k ⎟ t + σ t Bt ⎬ exp ∑ i =t1 ln yi )
σ 2 Nt
St = S0 exp[(α − − λ k )t + σ Bt ]Π yi ,
2 i =1

or alternatively as:

σ2 Nt
St = S0 exp[(α − − λ k )t + σ Bt + ∑ ln yi ] .
2 i =1

Using the previous definition of the log price (return) jump size ln( yt ) ≡ Yt :

σ2 Nt
St = S0 exp[(α − − λ k )t + σ Bt + ∑ Yi ] . (6)
2 i =1

This means that the asset price process {St ;0 ≤ t ≤ T } is modeled as an exponential Lévy
model of the form:

St = S0 e Lt ,

where X t is a Lévy process which is categorized as a Brownian motion with drift

(continuous part) plus a compound Poisson process (jump part) such that:

σ2 Nt
Lt = (α − − λ k )t + σ Bt + ∑ Yi .
2 i =1
St
In other words, log-return ln( ) is modeled as a Lévy process such that:
S0

St σ2 Nt
ln( ) = Lt = (α − − λ k )t + σ Bt + ∑ Yi .
S0 2 i =1

Nt
Note that the compound Poisson jump process Πy
i =1
i = 1 (in absolute price scale) or
Nt Nt

∑ ln yi = ∑ Yi = 0 (in log price scale) if Nt = 0 (i.e. no jumps between time 0 and t ) or

i =1 i =1
positive and negative jumps cancel each other out.

7
© 2004 Kazuhisa Matsuda All rights reserved.

In the Black-Scholes case, log return ln( St / S0 ) is normally distributed:

σ2
St = S0 exp[(α − )t + σ Bt ]
2
St σ2
ln( ) ∼ Normal[(α − )t , σ 2t ] .
S0 2

Nt
But in MJD case, the existence of compound Poisson jump process ∑Y
i =1
i makes log

return non-normal. In Merton’s case the simple distributional assumption about the log
return jump size (Yi ) ∼ N ( µ , δ 2 ) enables the probability density of log return
xt = ln( St / S0 ) to be obtained as a quickly converging series of the following form:

∞
P ( xt ∈ A) = ∑ P ( N t = i )P ( xt ∈ A N t = i )
i =0
∞
e − λ t (λ t ) i σ2
P( xt ) = ∑ N ( xt ;(α − − λ k )t + i µ , σ 2t + iδ 2 ) (7)
i =0 i! 2

σ2
where N ( xt ;(α − − λ k )t + i µ , σ 2t + iδ 2 )
2
2
⎡ ⎧⎪⎛ σ2 ⎞ ⎫⎪⎤
⎢ xt − ⎨⎜ α − − λ k ⎟ t + iµ ⎬⎥
⎢ ⎩⎪⎝ 2 ⎠ ⎭⎪⎦⎥
exp[− ⎣
1
= ].
2π (σ 2t + iδ 2 ) 2(σ t + iδ )
2 2

e − λ t (λ t )i
The term P( N t = i ) = is the probability that the asset price jumps i times during
i!
σ2
the time interval of length t . And P ( xt ∈ A N t = i ) = N ( xt ;(α −− λ k )t + i µ , σ 2t + iδ 2 )
2
is the Black-Scholes normal density of log-return assuming that the asset price jumps i
times in the time interval of t . Therefore, the log-return density in the MJD model can be
interpreted as the weighted average of the Black-Scholes normal density by the
probability that the asset price jumps i times.

By Fourier transforming the Merton log-return density function with FT parameters

( a, b, ) = (1,1) , its characteristic function is calculated as:

∞
φ (ω ) = ∫ exp ( iω xt )P( xt )dxt
−∞

8
© 2004 Kazuhisa Matsuda All rights reserved.

⎡ ⎧1 ⎫ ⎤
= exp ⎢λ t exp ⎨ ω ( 2i µ − δ 2ω ) ⎬ − λ t (1 + iω k ) − tω {−2iα + σ 2 ( i + ω )}⎥ .
1
⎣ ⎩2 ⎭ 2 ⎦

After simplification:

φ (ω ) = exp [tψ (ω )]

with the characteristic exponent (cumulant generating function):

⎪⎧ ⎛ δ 2ω 2 ⎞ ⎪⎫ ⎛ σ2 ⎞ σ 2ω 2
ψ (ω ) = λ ⎨exp ⎜ iωµ − ⎟ − 1 ⎬ + iω ⎜ α − − λ k ⎟− , (8)
⎩⎪ ⎝ 2 ⎠ ⎭⎪ ⎝ 2 ⎠ 2

1
µ+ δ 2
where k ≡ e 2 − 1 . The characteristic exponent (8) can be alternatively obtained by
substituting the Lévy measure of the MJD model:

λ ⎧⎪ ( dx − µ )2 ⎫⎪
(dx) = exp ⎨− ⎬ = λ f (dx)
2πδ 2 ⎪⎩ 2δ 2 ⎪
⎭
f ( dx) ∼ N ( µ , δ )
2

into the Lévy-Khinchin representation of the finite variation type (read Matsuda (2004)):

σ 2ω 2 ∞
ψ (ω ) = ibω − +∫
−∞
{exp(iω x) − 1} (dx)
2
σ 2ω 2 ∞
ψ (ω ) = ibω − +∫
−∞
{exp(iω x) − 1} λ f (dx)
2
σ 2ω 2 ∞
ψ (ω ) = ibω − + λ∫
−∞
{exp(iω x) − 1} f (dx)
2
ψ (ω ) = ibω −
σ 2ω 2
2
+λ {∫ ∞

−∞
eiω x f (dx) − ∫
−∞
∞
}
f (dx)

∞
Note that ∫−∞
eiω x f (dx) is the characteristic function of f ( dx ) :

∞ ⎛ δ 2ω 2 ⎞
∫−∞
iω x
e f ( dx ) = exp ⎜ i µω − ⎟.
⎝ 2 ⎠

Therefore:

9
© 2004 Kazuhisa Matsuda All rights reserved.

σ 2ω 2 ⎪⎧ ⎛ δ 2ω 2 ⎞ ⎪⎫
ψ (ω ) = ibω − + λ ⎨exp ⎜ i µω − ⎟ − 1⎬ ,
2 ⎩⎪ ⎝ 2 ⎠ ⎭⎪

σ2
where b = α − − λ k . This corresponds to (8). Characteristic exponent (8) generates
2
cumulants as follows:

σ2
cumulant1 = α − − λ k + λµ ,
2
cumulant2 = σ 2 + λδ 2 + λµ 2 ,
cumulant3 = λ (3δ 2 µ + µ 3 ) ,
cumulant4 = λ (3δ 4 + 6 µ 2δ 2 + µ 4 ) .

Annualized (per unit of time) mean, variance, skewness, and excess kurtosis of the log-
return density P ( xt ) are computed from above cumulants as follows:

σ2⎛ µ + 12 δ 2 ⎞
E [ xt ] = cumulant1 = α − −λ ⎜e − 1⎟ + λµ
2 ⎝ ⎠
Variance [ xt ] = cumulant2 = σ + λδ + λµ 2
2 2

cumulant3 λ (3δ 2 µ + µ 3 )
Skewness [ xt ] = =
( cumulant2 ) (σ + λδ 2 + λµ 2 )
3/ 2 3/ 2
2

cumulant4 λ (3δ 4 + 6µ 2δ 2 + µ 4 )
Excess Kurtosis [ xt ] = = . (9)
( cumulant2 ) (σ + λδ 2 + λµ 2 )
4/ 2 2
2

We can observe several interesting properties of Merton’s log-return density P ( xt ) .

Firstly, the sign of µ which is the expected log-return jump size, E[Yt ] = µ , determines
the sign of skewness. The log-return density P ( xt ) is negatively skewed if µ < 0 and it is
symmetric if µ = 0 as illustrated in Figure 1.

10
© 2004 Kazuhisa Matsuda All rights reserved.

3.5
Density 3
2.5 µ=−0.5
2
1.5 µ=0
1
0.5 µ=0.5
0
-0.75 -0.25 0 0.25 0.75
log −return xt
Figure 1: Merton’s Log-Return Density for Different Values of µ . µ = −0.5 in blue,
µ = 0 in red, and µ = 0.5 in green. Parameters fixed are τ = 0.25 , α = 0.03 , σ = 0.2 ,
λ = 1 , and δ = 0.1 .

Table 1
Annualized Moments of Merton’s Log-Return Density in Figure 1

Model Mean Standard Deviation Skewness Excess Kurtosis

µ = −0.5 -0.0996 0.548 -0.852 0.864

µ =0 0.005 0.3742 0 0.12
µ = 0.5 -0.147 0.5477 0.852 0.864

Secondly, larger value of intensity λ (which means that jumps are expected to occur
more frequently) makes the density fatter-tailed as illustrated in Figure 2. Note that the
excess kurtosis in the case λ = 100 is much smaller than in the case λ = 1 or λ = 10 .
This is because excess kurtosis is a standardized measure (by standard deviation).

11
© 2004 Kazuhisa Matsuda All rights reserved.

3.5
Density 3
2.5 λ=1
2
1.5 λ=10
1
0.5 λ=100
0
-0.75 -0.25 0 0.25 0.75
log −return xt
Figure 2: Merton’s Log-Return Density for Different Values of Intensity λ . λ = 1
in blue, λ = 10 in red, and λ = 100 in green. Parameters fixed are τ = 0.25 , α = 0.03 ,
σ = 0.2 , µ = 0 , and δ = 0.1 .
Table 2
Annualized Moments of Merton’s Log-Return Density in Figure 2

Model Mean Standard Deviation Skewness Excess Kurtosis

λ =1 0.00499 0.2236 0 0.12

λ = 10 -0.04012 0.3742 0 0.1531
λ = 100 -0.49125 1.0198 0 0.0277

Also note that Merton’s log-return density has higher peak and fatter tails (more
leptokurtic) when matched to the Black-Scholes normal counterpart as illustrated in
Figure 9.3.

3
2.5
Density

2
1.5 µ=−0.5
1
0.5 BS
0
-1 -0.5 0 0.5 1
log −return xt

12
© 2004 Kazuhisa Matsuda All rights reserved.

Figure 3: Merton Log-Return Density vs. Black-Scholes Log-Return Density

(Normal). Parameters fixed for the Merton (in blue) are τ = 0.25 , α = 0.03 , σ = 0.2 ,
λ = 1 , µ = −0.5 , and δ = 0.1 . Black-Scholes normal log-return density is plotted (in red)
by matching the mean and variance to the Merton.

Table 3
Annualized Moments of Merton vs. Black-Scholes Log-Return Density in Figure 3

Model Mean Standard Deviation Skewness Excess Kurtosis

Merton with
µ = −0.5 -0.0996 0.548 -0.852 0.864

Black-Scholes -0.0996 0.548 0 0

[3] Log Stock Price Process for Merton Jump-Diffusion Model

Log stock price dynamics can be obtained from the equation (6) as:

⎛ σ2 ⎞ Nt
ln St = ln S0 + ⎜ α − − λ k ⎟ t + σ Bt + ∑ Yi . (10)
⎝ 2 ⎠ i =1

Probability density of log stock price ln St is obtained as a quickly converging series of

the following form (i.e. conditionally normal):

∞
P (ln St ∈ A) = ∑ P ( N t = i ) P (ln St ∈ A N t = i )
i =0
∞
e − λ t (λ t ) i ⎛ ⎛ σ2 ⎞ ⎞
P (ln St ) = ∑ N ⎜⎜ ln St ;ln S0 + ⎜ α − − λ k ⎟ t + i µ , σ 2t + iδ 2 ⎟⎟ , (11)
i =0 i! ⎝ ⎝ 2 ⎠ ⎠

where:

⎛ ⎛ σ2 ⎞ ⎞
N ⎜⎜ ln St ;ln S0 + ⎜ α − − λ k ⎟ t + i µ , σ 2t + iδ 2 ⎟⎟
⎝ ⎝ 2 ⎠ ⎠
⎡ ⎧ ⎛ ⎞ ⎫⎪ ⎤
2
⎪
⎢ ⎨ln St − ⎜ ln S0 + ⎜ α −
⎛ σ 2
⎞
− λ k ⎟ t + iµ ⎟ ⎬ ⎥
1 ⎢ ⎩⎪ ⎝ ⎝ 2 ⎠ ⎠ ⎭⎪ ⎥
= exp ⎢ − ⎥.
(
2π σ 2t + iδ 2 ) ⎢ (
2 σ t + iδ
2 2
) ⎥
⎢ ⎥
⎣⎢ ⎦⎥

13
© 2004 Kazuhisa Matsuda All rights reserved.

By Fourier transforming (11) with FT parameters ( a, b, ) = (1,1) , its characteristic function

is calculated as:

∞
φ (ω ) = ∫ exp ( iω ln St )P(ln St )d ln St
−∞

⎡ ⎛ δ 2ω 2 ⎞ ⎛ σ2 ⎞ σ 2ω 2 ⎤
= exp ⎢λ t ⎜ exp(i µω − ) − 1⎟ + iω ⎜ ln S0 + (α − − λ k )t ⎟ − t⎥ , (12)
⎣ ⎝ 2 ⎠ ⎝ 2 ⎠ 2 ⎦

δ2
µ+
where k = e 2
−1.

[4] Lévy Measure for Merton Jump-Diffusion Model

Lévy measure ( dx ) of a compound Poisson process is given by the multiplication of the

intensity and the jump size density f ( dx ) :

( dx ) = λ f ( dx ) .

The Lévy measure ( dx ) represents the arrival rate (i.e. total intensity) of jumps of sizes
[ x, x + dx ] . In other words, we can interpret the Lévy measure ( dx ) of a compound
Poisson process as the measure of the average number of jumps per unit of time. Lévy
measure is a positive measure on R , but it is not a probability measure since its total mass
λ (in the compound Poisson case) does not have to equal 1:

∫ (dx) = λ ∈ R + .

A Poisson process and a compound Poisson process (i.e. a piecewise constant Lévy
process) are called finite activity Lévy processes since their Lévy measures ( dx ) are
finite (i.e. the average number of jumps per unit time is finite):

∞
∫−∞
(dx) < ∞ .

In Merton jump-diffusion case, the log-return jump size is ( dxi ) ∼ i.i.d . Normal ( µ , δ 2 ) :

1 (dxi − µ ) 2
f (dxi ) = exp{− }.
2πδ 2 2δ 2

Therefore, the Lévy measure ( dx ) for Merton case can be expressed as:

λ (dx − µ ) 2
(dx) = λ f (dx) = exp{− }. (13)
2πδ 2 2δ 2

14
© 2004 Kazuhisa Matsuda All rights reserved.

An example of Lévy measure ( dx ) for the log-return xt = ln( St / S0 ) in MJD model is

plotted in Figure 5. Each Lévy measure is symmetric (i.e. µ = 0 is used) with total mass
1, 2, and 4 respectively.

15
12.5
Density

10 λ=1
7.5
λ=2
5
2.5 λ=4
0
-0.4 -0.2 0 0.2 0.4
log −return xt
Figure 5: Lévy Measures ( dx ) for the Log-Return xt = ln( St / S0 ) in MJD Model for
Different Values of Intensity λ . Parameters used are µ = 0 and δ = 0.1 .

[5] Option Pricing: PDE Approach by Hedging

Consider a portfolio P of the one long option position V ( S , t ) on the underlying asset S
written at time t and a short position of the underlying asset in quantity ∆ to derive
option pricing functions in the presence of jumps:

Pt = V ( St , t ) − ∆St . (14)

Portfolio value changes by in a very short period of time:

dPt = dV ( St , t ) − ∆dSt . (15)

MJD dynamics of an asset price is given by equation (1) in the differential form as:

dSt
= (α − λ k )dt + σ dBt + ( yt − 1) dN t ,
St
dSt = (α − λ k ) St dt + σ St dBt + ( yt − 1) St dN t . (16)

Itô formula for the jump-diffusion process is given as (Cont and Tankov (2004)):

15
© 2004 Kazuhisa Matsuda All rights reserved.

∂f ( X t , t ) ∂f ( X t , t ) σ t2 ∂2 f ( X t , t)
df ( X t , t ) = dt + bt dt + dt
∂t ∂x 2 ∂x 2
∂f ( X t , t )
+σ t dBt + [ f ( X t − + ∆X t ) − f ( X t − )] ,
∂x

where bt corresponds to the drift term and σ t corresponds to the volatility term of a
Nt
jump-diffusion process X t = X 0 + ∫ bs ds + ∫ σ s dBs + ∑ ∆X i . Apply this to our case of
t t

0 0
i =1
option price function V ( S , t ) :

∂V ∂V σ 2 St 2 ∂ 2V ∂V
dV ( St , t ) = dt + (α − λ k ) St dt + dt + σ St dBt
∂t ∂St 2 ∂St 2
∂St
+[V ( yt St , t ) − V ( St , t )]dN t . (17)

The term [V ( yt St , t ) − V ( St , t )]dN t describes the difference in the option value when a
jump occurs. Now the change in the portfolio value can be expressed as by substituting
(16) and (17) into (15):

dPt = dV ( St , t ) − ∆dSt
∂V ∂V σ 2 St 2 ∂ 2V ∂V
dPt = dt + (α − λ k ) St dt + dt + σ St dBt + [V ( yt St , t ) − V ( St , t )]dN t
∂t ∂St 2 ∂St 2
∂St
−∆{(α − λ k ) St dt + σ St dBt + ( yt − 1) St dN t }
⎧ ∂V ∂V σ 2 St 2 ∂ 2V ⎫ ⎛ ∂V ⎞
dPt = ⎨ + (α − λ k ) St + − ∆(α − λ k ) St ⎬ dt + ⎜ σ St − ∆σ St ⎟ dBt
⎩ ∂t ∂St 2 ∂St ∂St
2
⎭ ⎝ ⎠
+ {V ( yt St , t ) − V ( St , t ) − ∆( yt − 1) St } dNt . (18)

If there is no jump between time 0 and t (i.e. dNt = 0 ), the problem reduces to Black-
Scholes case in which setting ∆ = ∂V / ∂St makes the portfolio risk-free leading to the
following (i.e. the randomness dBt has been eliminated):

⎧ ∂V ∂V σ 2 St 2 ∂ 2V ∂V ⎫ ⎛ ∂V ∂V ⎞
dPt = ⎨ + (α − λ k ) St + − (α − λ k ) St ⎬ dt + ⎜ σ St − σ St ⎟ dBt
⎩ ∂t ∂St 2 ∂St ∂St ∂St ∂St
2
⎭ ⎝ ⎠
⎧ ∂V σ St ∂ V ⎫
2 2 2
dPt = ⎨ + ⎬ dt .
⎩ ∂t 2 ∂St 2 ⎭

This in turn means that if there is a jump between time 0 and t (i.e. dNt ≠ 0 ), setting
∆ = ∂V / ∂St does not eliminate the risk. Suppose we decided to hedge the randomness
caused by diffusion part dBt in the underlying asset price (which are always present) and

16
© 2004 Kazuhisa Matsuda All rights reserved.

not to hedge the randomness caused by jumps dN t (which occur infrequently) by setting
∆ = ∂V / ∂St . Then, the change in the value of the portfolio is given by from equation
(9.18):

⎧ ∂V ∂V σ 2 St 2 ∂ 2V ∂V ⎫ ⎛ ∂V ∂V ⎞
dPt = ⎨ + (α − λ k ) St + − (α − λ k ) St ⎬ dt + ⎜ σ St − σ St ⎟ dBt
⎩ ∂t ∂St 2 ∂St ∂St ∂St ∂St
2
⎭ ⎝ ⎠
⎧ ∂V ⎫
+ ⎨V ( yt St , t ) − V ( St , t ) − ( yt − 1) St ⎬ dNt ,
⎩ ∂St ⎭
⎧ ∂V σ St ∂ V ⎫
2 2 2
⎧ ∂V ⎫
dPt = ⎨ + 2⎬
dt + ⎨V ( yt St , t ) − V ( St , t ) − ( yt − 1) St ⎬ dNt . (19)
⎩ ∂t 2 ∂St ⎭ ⎩ ∂St ⎭

Merton argues that the jump component ( dN t ) of the asset price process St is
uncorrelated with the market as a whole. Then, the risk of jump is diversifiable (non-
systematic) and it should earn no risk premium. Therefore, the portfolio is expected to
grow at the risk-free interest rate r :

E[dPt ] = rPdt
t . (20)

After substitution of (14) and (19) into (20) by setting ∆ = ∂V / ∂St :

⎧ ∂V σ 2 St 2 ∂ 2V ⎫ ⎧ ∂V ⎫
E[⎨ + 2⎬
dt + ⎨V ( yt St , t ) − V ( St , t ) − ( yt − 1) St ⎬ dNt ] = r{V ( St , t ) − ∆St }dt
⎩ ∂t 2 ∂St ⎭ ⎩ ∂St ⎭
∂V σ St ∂ V
2 2 2
∂V ∂V
( + )dt + E[V ( yt St , t ) − V ( St , t ) − ( yt − 1) St ]E[dN t ] = r{V ( St , t ) − St }dt
∂t 2 ∂St 2
∂St ∂St
∂V σ 2 St 2 ∂ 2V ∂V ∂V
( + )dt + E[V ( yt St , t ) − V ( St , t ) − ( yt − 1) St ]λ dt = r{V ( St , t ) − St }dt
∂t 2 ∂St 2
∂St ∂St
∂V σ 2 St 2 ∂ 2V ∂V ∂V
+ + λ E[V ( yt St , t ) − V ( St , t ) − ( yt − 1) St ] = r{V ( St , t ) − St }
∂t 2 ∂St 2
∂St ∂St

Thus, the MJD counterpart of Black-Scholes PDE is:

∂V σ 2 St 2 ∂ 2V ∂V ∂V
+ + rSt − rV + λ E[V ( yt St , t ) − V ( St , t )] − λ St E[ yt − 1] = 0 . (21)
∂t 2 ∂St 2
∂St ∂St

17
© 2004 Kazuhisa Matsuda All rights reserved.

where the term E[V ( yt St , t ) − V ( St , t )] involves the expectation operator and

1
µ+ δ 2
E[ yt − 1] = e 2 − 1 ≡ k (which is the mean of relative asset price jump size). Obviously,
if jump is not expected to occur (i.e. λ = 0 ), this reduces to Black-Scholes PDE:2

∂V σ 2 St 2 ∂ 2V ∂V
+ + rSt − rV = 0 .
∂t 2 ∂St 2
∂St

Merton’s simple assumption that the absolute price jump size is lognormally distributed
(i.e. the log-return jump size is normally distributed, Yt ≡ ln( yt ) ∼ N ( µ , δ 2 ) ) makes it
possible to solve the jump-diffusion PDE to obtain the following price function of
European vanilla options as a quickly converging series of the form:

∞
e − λτ (λτ )i
∑
i =0 i!
VBS ( St ,τ = T − t , σ i , ri ) , (22)

1
µ+ δ 2
where λ = λ (1 + k ) = λ e 2
,
iδ 2
σ i2 = σ 2 + ,
τ
1
i(µ + δ 2 )
i ln(1 + k ) 1
µ+ δ 2
ri = r − λ k + = r − λ (e 2
− 1) + 2 ,
τ τ
and VBS is the Black-Scholes price without jumps.

Thus, MJD option price can be interpreted as the weighted average of the Black-Scholes
price conditional on that the underlying asset price jumps i times to the expiry with
weights being the probability that the underlying jumps i times to the expiry.

[6] Option Pricing: Martingale Approach

Let {Bt ;0 ≤ t ≤ T } be a standard Brownian motion process on a space (Ω, F , P ) . Under

actual probability measure P , the dynamics of MJD asset price process is given by
equation (6) in the integral form:

σ2 Nt
St = S0 exp[(α − − λ k )t + σ Bt + ∑ Yk ] .
2 k =1

2
This equation not only contains local derivatives but also links together option values at discontinuous
values in S. This is called non-local nature.

18
© 2004 Kazuhisa Matsuda All rights reserved.

Nt Nt
We changed the index from ∑Y i =1
i to ∑Y k =1
k . This is trivial but readers will find the reason

soon. MJD model is an example of an incomplete model because there are many
equivalent martingale risk-neutral measures Q ∼ P under which the discounted asset
price process {e − rt St ; 0 ≤ t ≤ T } becomes a martingale. Merton finds his equivalent
martingale risk-neutral measure Q M ∼ P by changing the drift of the Brownian motion
process while keeping the other parts (most important is the jump measure, i.e. the
distribution of jump times and jump sizes) unchanged:

σ2 Nt
St = S0 exp[(r − − λ k )t + σ B t
QM
+ ∑ Yk ] under Q M . (23)
2 k =1

Note that BtQM is a standard Brownian motion process on (Ω, F , Q M ) and the process
{e − rt St ; 0 ≤ t ≤ T } is a martingale under Q M . Then, a European option price V Merton (t , St )
with payoff function H ( ST ) is calculated as:

V Merton (t , St ) = e− r (T −t ) E QM [ H ( ST ) Ft ] . (24)

Standard assumption is Ft = St , thus:

σ2 NT −t
V Merton
(t , St ) = e − r (T − t )
E QM
[ H ( St exp[(r − − λ k )(T − t ) + σ B QM
T −t + ∑ Yk ]) St ]
2 k =1

σ 2 NT −t
V Merton (t , St ) = e− r (T −t ) E QM [ H ( St exp[(r − − λ k )(T − t ) + σ BTQ−Mt + ∑ Yk ])] . (25)
2 k =1

Poisson counter is (we would like to use index i for the number of jumps):

NT −t = 0,1, 2,... ≡ i .

And the compound Poisson process is distributed as:

NT −t

∑Y
k =1
k ∼ Normal (i µ , iδ 2 ) .

Thus, V Merton (t , St ) can be expressed as from equation (25) (i.e. by conditioning on i ):

V Merton (t , St )
σ2 i
=e − r (T − t )
∑Q
i ≥0
M ( NT −t = i)E QM
[ H ( St exp[(r −
2
− λ k )(T − t ) + σ BTQ−Mt + ∑ Yk ])] .
k =1

19
© 2004 Kazuhisa Matsuda All rights reserved.

Use τ = T − t :

V Merton (t , St )
e − λτ (λτ )i QM ⎡ ⎛ σ2 i
⎞⎤
= e − rτ ∑ E ⎢ ⎜ t
H S exp[{r − − λ ( e µ +δ 2 / 2
− 1)}τ + σ Bτ
QM
+ ∑ Yk ] ⎟ ⎥ . (26)
i≥0 i! ⎣ ⎝ 2 k =1 ⎠⎦

Inside the exponential function is normally distributed:

σ2 i
{r − − λ ( e µ +δ − 1)}τ + σ BτQM + ∑ Yk
2
/2

2 k =1

⎛ σ 2
⎞
∼ Normal ⎜ {r − − λ (e µ +δ / 2 − 1)}τ + i µ , σ 2τ + iδ 2 ⎟ .
2

⎝ 2 ⎠

Rewrite it so that its distribution remains the same:

σ2 σ 2τ + iδ 2 Q
{r − − λ (e µ +δ − 1)}τ + i µ +
2
/2
Bτ M

2 τ
⎛ σ2 ⎞
∼ Normal ⎜ {r − − λ (e µ +δ / 2 − 1)}τ + i µ , σ 2τ + iδ 2 ⎟ .
2

⎝ 2 ⎠

Now we can rewrite equation (24) as (we can do this operation because a normal density
is uniquely determined by only two parameters: its mean and variance):

V Merton (t , St )
e− λτ (λτ )i QM ⎡ ⎛ σ2 σ 2τ + iδ 2 QM ⎞ ⎤
= e− rτ ∑ E ⎢ H ⎜ St exp[{r − − λ (e µ +δ / 2 − 1)}τ + i µ + Bτ ] ⎟ ⎥
2

i ≥0 i! ⎢⎣ ⎜⎝ 2 τ ⎟⎥
⎠⎦ .

iδ 2 iδ 2
We can always add ( − ) = 0 inside the exponential function:
2τ 2τ

e− λτ (λτ )i
V Merton (t , St ) = e− rτ ∑ ×
i ≥0 i!
⎡ ⎛ σ 2 iδ 2 iδ 2 iδ 2 QM ⎞ ⎤
µ +δ 2 / 2
QM
E ⎢ H ⎜ St exp[{r − +( − ) − λ (e − 1)}τ + iµ + σ +
2
B ] ⎟⎥
⎢⎣ ⎜⎝ 2 2τ 2τ τ τ ⎟⎠ ⎥
⎦
− λτ
e (λτ ) i
V Merton (t , St ) = e− rτ ∑ ×
i ≥0 i!

20
© 2004 Kazuhisa Matsuda All rights reserved.

⎡ ⎛ 1 iδ 2 iδ 2 iδ 2 QM ⎞ ⎤
E QM ⎢ H ⎜ St exp[{r − (σ 2 + )+ − λ (e µ +δ / 2 − 1)}τ + i µ + σ 2 + B ] ⎟⎥
2

⎢⎣ ⎜⎝ 2 τ 2τ τ τ ⎟⎠ ⎥
⎦

iδ 2
Set σ i2 = σ 2 + and rearrange:
τ

V Merton (t , St )
e − λτ (λτ )i QM ⎡ ⎛ 1 2 iδ 2 ⎞⎤
= e − rτ ∑ − σi + − λ (e µ +δ / 2 − 1)}τ + i µ + σ i BτQM ] ⎟ ⎥
2
E ⎢ ⎜ t
H S exp[{r
i≥0 i! ⎣ ⎝ 2 2τ ⎠⎦
e − λτ (λτ )i QM ⎡ ⎛ iδ 2 1 ⎞⎤
= e − rτ ∑ µ + − λ (e µ +δ / 2 − 1)τ }exp{(r − σ i2 )τ + σ i BτQM } ⎟ ⎥ .
2
E ⎢ ⎜ t
H S exp{i
i ≥0 i! ⎣ ⎝ 2 2 ⎠⎦

Black-Scholes price can be expressed as:

1
V BS (τ = T − t , St ; σ ) = e − rτ E QBS [ H {St exp(r − σ 2 )τ + σ BτQ BS }] .
2

Finally, MJD pricing formula can be obtained as a weighted average of Black-Scholes

price conditioned on the number of jumps i :

V Merton (t , St )
e − λτ (λτ )i BS iδ 2 iδ 2
=∑ V (τ , Si ≡ St exp{i µ + − λ (e µ +δ 2 / 2
− 1)τ }; σ i ≡ σ +
2
). (27)
i ≥0 i! 2 τ

Alternatively:

V Merton (t , St )
e − λτ (λτ )i QM ⎡ ⎪⎧ ⎛ i µ + iδ 2 / 2 1 2 ⎞ ⎪⎫⎤
= e − rτ ∑ E ⎢ H ⎨ St exp{⎜ r − λ (e
µ +δ 2 / 2
− 1) + − σ i ⎟τ + σ i BτQM }⎬⎥
i ≥0 i! ⎢⎣ ⎩⎪ ⎝ τ 2 ⎠ ⎭⎪⎥⎦
e − λτ (λτ )i BS iδ 2 i µ + iδ 2 / 2
=∑ V (τ , St ; σ i ≡ σ 2 + , ri ≡ r − λ (e µ +δ / 2 − 1) +
2
). (28)
i≥0 i! τ τ

1
µ+ δ 2
where λ = λ (1 + k ) = λ e 2
. As you might notice, this is the same result as the option
pricing formula derived from solving a PDE by forming a risk-free portfolio in equation
(22). PDE approach and Martingale approach are different approaches but they are
related and give the same result.

[7] Option Pricing Example of Merton Jump-Diffusion Model

21
© 2004 Kazuhisa Matsuda All rights reserved.

In this section the equation (22) is used to price hypothetical plain vanilla options: current
stock price St = 50, risk-free interest rate r = 0.05, continuously compounded dividend
yield q = 0.02, time to maturity τ = 0.25 years.

We need to be careful about volatility σ . In the Black-Scholes case, the t -period standard
deviation of log-return xt is:

Standard Deviation BS ( xt ) = σ BS t . (29)

Equation (10) tells that the t -period standard deviation of log-return xt in the Merton
model is given as:

Standard Deviation Merton ( xt ) = (σ Merton 2 + λδ 2 + λµ 2 )t . (30)

This means that if we set σ BS = σ Merton , MJD prices are always greater (or equal to) than
Black-Scholes prices because of the extra source of volatility λ (intensity), µ (mean
log-return jump size), δ (standard deviation of log-return jump) (i.e. larger volatility is
translated to larger option price):

Standard Deviation BS ( xt ) ≤ Standard Deviation Merton ( xt )

σ BS t ≤ (σ Merton 2 + λδ 2 + λµ 2 )t .

This very obvious point is illustrated in Figure 6 where diffusion volatility is

set σ BS = σ Merton = 0.2 . Note the followings: (1) In all four panels MJD price is always
greater (or equal to) than BS price. (2) When Merton parameters ( λ , µ , and δ )
are small in Panel A, the difference between these two prices is small. (3) As intensity λ
increases (i.e. increased expected number of jumps per unit of time), the t -period Merton
standard deviation of log-return xt increases (equation (30)) leading to the larger
difference between Merton price and BS price as illustrated in Panel B. (4) As Merton
mean log-return jump size µ increases, the t -period Merton standard deviation of log-
return xt increases (equation (30)) leading to the larger difference between Merton price
and BS price as illustrated in Panel C. (5) As Merton standard deviation of log-return
jump size δ increases, the t -period Merton standard deviation of log-return xt increases
(equation (30)) leading to the larger difference between Merton price and BS price as
illustrated in Panel D.

22
© 2004 Kazuhisa Matsuda All rights reserved.

20

Call Price
15

10
Merton
5
BS
0
30 40 50 60 70
Strike K
A) Merton parameters: λ = 1, µ = -0.1, and δ = 0.1.

20
Call Price

15

10
Merton
5
BS
0
30 40 50 60 70
Strike K
B) Merton parameters: λ = 5, µ = -0.1, and δ = 0.1.

20
Call Price

15

10 Merton
5
BS
0
30 40 50 60 70
Strike K
C) Merton parameters: λ = 1, µ = -0.5, and δ = 0.1.

20
Call Price

15

10 Merton
5
BS
0
30 40 50 60 70
Strike K
D) Merton parameters: λ = 1, µ = -0.1, and δ = 0.5.

23
© 2004 Kazuhisa Matsuda All rights reserved.

Figure 6: MJD Call Price vs. BS Call Price When Diffusion Volatility σ is same.
Parameters and variables used are St = 50, r = 0.05, q = 0.02, τ = 0.25, and
σ BS = σ Merton = 0.2 .

Next we consider a more delicate case where we restrict diffusion volatilities σ BS and
σ Merton such that standard deviations of MJD and BS log-return densities are the same:

Standard Deviation BS ( xt ) = Standard Deviation Merton ( xt ) ,

σ BS t = (σ Merton 2 + λδ 2 + λµ 2 )t .

Using the Merton parameters λ = 1, µ = -0.1, and δ = 0.1 and BS volatility σ BS = 0.2 ,
Merton diffusion volatility is calculated as σ Merton = 0.141421. In this same standard
deviation case, call price function is plotted in Figure 7 and put price function is plotted
in Figure 8. It seems that MJD model overestimates in-the-money call and underestimates
out-of-money call when compared to BS model. And MJD model overestimates out-of-
money put and underestimates in-the-money put when compared to BS model.

20
Call Price

15

10
Merton
5
BS
0
30 40 50 60 70
Strike K
A) Range 30 to 70.

8
Call Price

7
6
5
4 Merton
3
2 BS
1
42 44 46 48 50 52
Strike K
B) Range 42 to 52.

24
© 2004 Kazuhisa Matsuda All rights reserved.

Figure 7: MJD Call Price vs. BS Call Price When Restricting Merton Diffusion
Volatility σ Merton . We set σ BS = 0.2 and σ Merton = 0.141421. Parameters and variables used
are St = 50, r = 0.05, q = 0.02, τ = 0.25.

Call Price
15

10
Merton
5
BS
0
30 40 50 60 70
Strike K
A) Range 30 to 70.

2.5
Call Price

2
1.5
Merton
1
0.5 BS
0
42 44 46 48 50 52
Strike K
B) Range 42 to 52.

Figure 8: MJD Put Price vs. BS Put Price When Restricting Merton Diffusion
Volatility σ Merton . We set σ BS = 0.2 and σ Merton = 0.141421. Parameters and variables used
are St = 50, r = 0.05, q = 0.02, τ = 0.25.

25
© 2004 Kazuhisa Matsuda All rights reserved.

References

Black, F. and Scholes, M., 1973, “The Pricing of Options and Corporate Liabilities,”
Journal of Political Economy 3.

Carr, P., Madan, D. B., 1998, “Option Valuation Using the Fast Fourier Transform,”
Journal of Computational Finance 2, 61-73.

Cont, R. and Tankov, P., 2004, Financial Modelling with Jump Processes, Chapman &
Hall/CRC Financial Mathematics Series.

Matsuda, K., 2004, Introduction to Option Pricing with Fourier Transform: Option
Pricing with Exponential Lévy Models, Working Paper, Graduate School and University
Center of the City University of New York.

Merton, R. C., 1971, “Optimum Consumption and Portfolio Rules in a Continuous-Time

Model,” Journal of Economic Theory, 3, 373-413.

Merton, R. C., 1976, “Option Pricing When Underlying Stock Returns Are Continuous,”
Journal of Financial Economics, 3, 125-144.

Wilmott, P., 1998, Derivatives, John Wiley & Sons

26