Journal of Applied Mathematics & Bioinformatics, Vol.10, No.

1, 2020, 11-30
ISSN: 1792-6602 (print), 1792-6939 (online)
Scientific Press International Limited, 2020

Modeling Underlying Assets Log-return

in Merton Jump-Diffusion Framework

Megang Nkamga Junile Staures1 , Jane Akinyi Aduda2 and Romuald Momeya3

Abstract

In this present paper we analyze two exponential Lévy models, the

Black-Scholes model and the Merton Jump-Diffusion model from the
perspective of the investigation of the skewness and excess kurtosis
present in underlying assets log-returns distribution. Calibrating both
models on real-world financial data and investigating their various mo-
ments and mean square error, we obtain results which show how the
Merton jump-diffusion model performs better than the Black-Scholes
model for modeling log-returns. This conclusion was also confirmed by
using the Diebold-Mariano test to compare the forecast accuracy of the
two models.

Keywords: Black-Scholes (BS) model; Merton jump-diffusion (MJD) model;

log-returns; skewness; excess kurtosis

1
Department of Mathematics, Pan African University, Institute of Basic Sciences,
Technology, and Innovation, Kenya. E-mail: [email protected]
2
Department of Actuarial Science and Statistics, Jomo Kenyatta University of
Agriculture, Technology and Innovation. E-mail: [email protected]
3
Quantitative Analyst, CIBC Asset Management Inc. Montreal, QC.
E-mail: [email protected]

Article Info: Received : August 17, 2019. Revised : September 15, 2019.
Published online : January 5, 2020
12 Modeling Underlying Assets Log-return in Merton...

1 Introduction

The prices of financial instruments such as stocks vary constantly and cause
significant risk on businesses or organizations connected to such fluctuating
prices. To mitigate this risk, modern finance establishes suitable models to
assist investors or traders analyze real-world data to predict the future behavior
of asset prices in the financial market. In this paper, we will focus on two
exponential Lévy models (BS and MJD) to analyze the trend of the underlying
asset prices log-return.
The BS model is a pioneer valuation model, proposed by [1] and a bench-
mark against which other models can be compared. BS model assumes that
the log-return of the underlying assets is Gaussian, i.e., normally distributed.
Unfortunately, assets in real markets rarely exhibit this behavior (see [2, 3]).
Real asset prices and related market indexes show far more extreme fluctua-
tions than predicted by Gaussian statistics. According to [2, 4], these extreme
fluctuations leads to two empirical phenomena: leptokurtic feature and volatil-
ity smile. Therefore the need to propose another valuation model to handle
this issue since the BS model experiences numerous inadequacies and is unable
to explain empirical stylized facts in the financial markets.
In 1976 Robert C. Merton received the Nobel Prize award in Economics
for developing a JDM called Merton jump-diffusion (MJD) model, a general-
ization of the BS model. Merton’s main idea was to implant discontinuous
jump processes in the classical BS model to help describe discontinuous price
behavior in stock. The formula is the geometric Brownian motion (random
walk) plus a compound Poisson process to account for “jumps” in the stock
prices. The random walk together with the jump component are assumed to be
sources of randomness in the stock prices. By adding three more parameters to
the classical BS model, the MJD model captures the skewness and leptokurtic
feature in the underlyings log-return distribution which differs from the BS
normal log-return distribution. To compare the forecast accuracy of the two
models, both models are calibrated and the results are used to compare the
plot of their log-density functions, the values of their corresponding moments
and mean square error. To boost our comparison, the Diebold-Mariano test
[5] is applied and indicates the predictive performance of the MJD model over
the BS model.
Megang Nkamga Junile Staures, Jane Akinyi Aduda and Romuald Momeya 13

To investigate both models, some basic tools in probability and stochastic

process theory will be introduced in section 2. Also in section 2, we will intro-
duce the two exponential Lévy models and estimation of the model parameters
by Multinomial Maximum Likelihood function. We conclude this section with
a brief introduction to the Diebold-Mariano test. In section 3 and 4 we present
the corresponding results and conclusion respectively.

2 Model Specification
2.1 The BS model
In this model the asset price is described as follows
dXt
= µdt + σdWt , (1)
Xt
where µ is the expected return of the asset, σ is the volatility of the asset and
Wt is a Wiener process. The solution to equation 1 is given by
1 2 )t+σW
Xt = X0 e(µ− 2 σ t
(2)

where X0 is the stock price at time t=0, µ, σ and Wt are define as in equation
1. The assets are modeled by 2. From 2, the log-return of the asset price is:
Xt 1
Lt = ln( ) = (µ − σ 2 )t + σWt
X0 2
For the given parameters µ = 0.17, σ = 0.20 and X0 = 87 and simula-
tion of the standard Brownian motion, we easily generate BS modeled asset
prices from equation 2 as shown in Figure 1. Under the BS model, for a time
increment 4t the log-return of the asset price is given by:
Xt+4t
L4t = ln( )
Xt
1
= (µ − σ 2 )4t + σ(Wt+4t − Wt ) (3)
 2
σ2

2
∼ N (µ − )4t, σ 4t
2
which means
σ2
L4t − (µ − )4t = σ4Wt ∼ N (0, σ 2 4t)
2
14 Modeling Underlying Assets Log-return in Merton...

Figure 1: BS-modeled asset prices

The above equation shows that the log-returns follow the Gaussian distribu-
tion. It is no news that the BS model is complete, and in this framework there
exists a unique risk-neutral measure P∗ under which the discounted asset price
is a martingale, see for example [6].
By the Girzanov theorem allouba1996different, the asset prices are de-
scribed by the risk-neutral dynamic as follows

dXt
= rdt + σdWt∗ ,
Xt

where
µ−r
Wt∗ = t + Wt , t ∈ [0, T ]
σ
is a P∗ -Brownian motion and µ−r σ
the market price of risk measured per unit
of volatility. Under BS model, the characteristic function associated with the
log-return asset price L4t is given by:

σ2
  
1 2 2
Φ(θ) = exp 4t i(µ − )θ − σ θ (4)
2 2

Regarding the leptokurtic feature which is a measure of how heavy or fat

the tail of the log-returns distribution is, we’ll exploit the approach suggested
in [8].The theorem below is a consequence of equation 3.
Megang Nkamga Junile Staures, Jane Akinyi Aduda and Romuald Momeya 15

Theorem 2.1.
Let E∗BS , MiBS , i = 2 : 4, SBS and KBS denote the mean, the central moments,
the skewness and excess kurtosis for the BS-modeled log-returns respectively.
We’ve 

 E∗BS = (µ − 12 σ 2 )4t

M2BS = σ 2 4t




 M BS = 0

3
BS


 M 4 = 3σ 4 (4t)2



 SBS = 0

KBS = 3


Proof.
From equation 3, we clearly see that
E∗BS = E∗ [L4t ]
1
= (µ − σ 2 )4t since 4Wt = Wt+4t − Wt ∼ N (0, 4t).
2
By the central moment definition
MiBS = E∗ (Lt − E∗ [Lt ])i
 

= E∗ (σ4Wt )i
 
(5)
= σ i E∗ (4Wt )i
 

Let Φ4Wt (θ) denote the characteristic function of 4Wt . Upon calculation we
obtain
Φ4Wt (θ) = E∗ (eiθ4Wt )
Z ∞
1 x2
= eiθx . √ e− 24t dx
−∞ 2π4t
Z ∞
1 (x−iθ4t)2 +(θ4t)2
=√ e− 24t dx
2π4t −∞
Z ∞
2
− θ 24t 1 (x−iθ4t)2
=e since √ e− 24t dx=1
2π4t −∞
= exp {Ψ(θ)}
θ 2 4t
where Ψ(θ) = e− 2 .
Assuming that the characteristic function is sufficiently differentiable,
 
∗ n 1 dΨ(0)
=⇒ E [(4Wt ) ] = n
i dθ
=⇒ E∗ [(4Wt )2 ] = 4t
16 Modeling Underlying Assets Log-return in Merton...

From equation 5 we get

M2BS = σ 2 4t
Also we have that
E∗ [(4Wt )3 ] = 0
leading to

M3BS = 0 and
M4BS = 3σ 4 (4t)2

which yields

M3BS
S BS = =0
(M2BS )1 .5
M4BS
KBS = =3
(M2BS )2

2.1.1 Estimation of the parameters µ and σ for the BS model

Figure 2: Empirical log-returns rdt of NASDAQ index.

To find estimates of µ and σ based on empirical data, let us consider the

empirical log-returns rdt of NASDAQ index data from 2014-02-06 to 2019-05-
23, with 1332 trading days in Figure 2 above.
Megang Nkamga Junile Staures, Jane Akinyi Aduda and Romuald Momeya 17

Estimation of model parameters (µ, σ) is required for us to fit empirical

data to the BS model. From theorem 2.1:
(
E∗BS = (µ − 21 σ 2 )4t
M2BS = σ 2 4t
so that  ∗ +Var[L
 µ̂ = 2Ed
BS
d 4t ]4t
q 24t
 σ̂ = Var[r
d dt ]
4t

∗
where Ed
BS and Var[L4t ] are the sample mean and sample variance of the
d
empirical log-returns respectively.

2.2 The MJD model

The model propose that the underlying asset price evolves according
Nt
!
dXt X
= (µ − λk)dt + σdWt + d (Ji − 1) (6)
Xt− | {z }
i=1
Continuous part | {z }
Discontinuous or jump part

where: µ is the instantaneous expected return on the stock; k = E[Jt − 1]

is the expected percentage change in the stock price if a Poisson event oc-
cur; σ is the instantaneous variance of the return, condition on no arrival of
important new information (i.e. a Poisson process does not occur) that one
assumes constant. Xt− the asset price before a jump occurs at time t; dWt is
a standard Brownian motion; Nt is a Poisson process represents the arrival of
new information (events) which has a significant effect on the stock price with
parameter λ which stand for the average number of jump arrivals per unit of
time; Jt the jump sizes are i.i.d. Because of the jump component provided
by the compound Poisson process, the model considered in 6 is a process not
purely Gaussian representing a particular case of Lévy process. The solution
to 6 is given by:
1 2 PNt
Xt = X0 e(µ− 2 σ −λk)t+σWt + i=1 Ui (7)
where N
P t
i=1 Ui is a normally distributed compound Poisson process with mean
Nt µj , variance Nt σj2 and intensity λ. In other-words, for n = Nt , ni=1 Ui ∼
P

N (nµj , nσj2 ). The index j represents the jump part of MJD model. For the
18 Modeling Underlying Assets Log-return in Merton...

Figure 3: MJD-modeled asset prices.

assumed parameters µ = 0.21, σ = 0.056, λ = 4, µj = 0.051 and σj = 0.097, we

present simulated sample path of MJD-modeled asset prices in Figure 3.
Equation 7 describes stock price behavior under MJD model. ¿From equa-
tion 7, the log stock prices are given by

t N
1 X
ln Xt = ln X0 + (µ − σ 2 − λk)t + σWt + Ui
2 i=1

Then For a time increment 4t, we get the log-returns L4t of the asset
prices modeled by MJD model as

Xt+dt
L4t = ln( )
Xt
4Nt
X (8)
= γ4t + σ4Wt + Ui
i=1

0
where γ = µ − 21 σ 2 − λk, Ui s are i.i.d and Ui ∼ N (µj , σj2 ); 4Wt = Wt+4t − Wt
is the standard Brownian motion increment; and lastly 4Nt = Nt+4t − Nt
a Poisson process with mean λ4t. [8] proof that the assumption about the
log-return jump size distribution being normal facilitates the derivation of the
probability density function of the log-returns L4t which is a converging series
Megang Nkamga Junile Staures, Jane Akinyi Aduda and Romuald Momeya 19

of the following form:

∞
X
gt = P(L4t ∈ B) = P(4Nt = a)P(L4t |4Nt = a) for B ⊂ R
h=0
∞
X e−λ4t (λ4t)a
N L4t ; γ4t + µj a, σ 2 4t + σj2 a


=
a=0
a!

where
2
(L4t −[γ4t+µj a])
1 −
2(σ 2 4t+σ 2 a)
N L4t ; γ4t + µj a, σ 2 4t + σj2 a = q


e j

2π(σ 2 4t + σj2 a)

is the normal density function of L4t assuming that the asset price jumps a
times in the time interval 4t and P(4Nt = a) is the probability that the
asset price jumps a times in the time interval 4t. gt is expressed as the
weighted sum of normal densities. In such a framework, the market model is
incomplete, hence the existence of more than one risk-neutral measure P∗ . To
have a risk-neutral measure, we need to replace the compound Poisson process
in equation 6 by a compensated compound Poisson process.The latter can be
done in several ways. Let us consider the discounted asset price e−rt Xt such
that

d(e−rt Xt ) = −re−rt Xt dt + e−rt dXt

Nt
!
X
= −re−rt Xt dt + e−rt Xt (µ − λk)dt + e−rt σdWt Xt + e−rt Xt d (Ji − 1)
i=1
−rt ∗ ∗ −rt
= e Xt (µ − r − λk + λ k )dt + e σd(Wt∗ − ut)Xt +
Nt
!
X
+ e−rt Xt d (Ji − 1) − λ∗ k ∗
i=1
−rt
= e Xt (µ − r − λk + λ∗ k ∗ − σu)dt + e−rt σdWt∗ Xt +
Nt
!
X
+ e−rt Xt d (Ji − 1) − λ∗ k ∗
i=1

For e−rt Xt to be a martingale we suppose

µ − r − λk + λ∗ k ∗ − σu = 0

where u is such that

Wt∗ = ut + Wt t ∈ [0, T ]
20 Modeling Underlying Assets Log-return in Merton...

is a P∗ -Brownian motion thanks to the Girzanov theorem. λ∗ > 0 is the new

intensity and k ∗ = E∗ [J − 1]. Merton proposes the following for the change of
measure?
λ∗ = λ

So that
fU∗ (j) = fU (j) =⇒ k = k ∗
µ−r
=⇒ u =
σ
For simplicity, if we let u = 0, then µ = r. Since the jump risk is diversify and
no risk premium is attached to it, we can leave the jump part unchanged. So
the new asset price dynamic under P∗ is

Nt
!
dXt X
= (r − λk)dt + σdWt∗ + d (Ji − 1) (9)
Xt− i=1

For the MJD model, the central moment is given by:

MiM JD = E∗ (Lt − E∗ [Lt ])i

 

= E∗ (σ4Wt + U 4Nt − λµj 4t)i

 

with Wt , U, 4Nt being i.i.d. The characteristic function of L4t is obtained by

applying Fourier transform to gt as

Φ(w) = e4tΨ(w)

where
   
1 2 2 1 2 1
Ψ(w) = λ exp (iwµj − σj w ) − 1 + iw µ − σ − λk − σ 2 w2
2 2 2

is the characteristic exponent (cumulant generating function) and

k = E[eU − 1]
1 2
= eµj + 2 σj − 1.

Theorem 2.2.
Considering the central moment and the characteristic exponent defined above,
Megang Nkamga Junile Staures, Jane Akinyi Aduda and Romuald Momeya 21

we’ve

E∗M JD = (γ + λµj )4t

2 2 2 2


MM JD = σ + λ(σj + µ j ) 4t
3 2 3 2 2
MM JD = λ(3σj + µj )4t + 6µj σj (λ4t)
4 4 4 2 2 2 2 4 4 2 2 2
MM JD = λ(3σj + µj + 6σj µj )4t + 3(σ 4t) + (3µj + 21σj + 30µj σj )(λ4t) +

+ 6σ 2 4t(σj2 + µ2j )λ4t + (6µ2j σj2 + 18σj4 )(λ4t)3 + 6σ 2 σj2 4t(λ4t)2 + 3σj4 (λ4t)4
3
MM JD
SM JD = 2 1.5
[MM JD ]
4
MM JD
KM JD = 2 2
[MM JD ]

For the proof of the above theorem, the reader is referred to [8, 9].

2.2.1 Some Important Properties of the PDF gt of Merton log-

returns

1. The sign of skewness is determined by the size of E[J] = µj . Figure 4

shows that gt is asymmetric if µj 6= 0 and symmetric if µj = 0.

2. The value of λ makes the density fatter-tailed as illustrated in Figure 4.

Note that the excess kurtosis in the case λ = 50 is much smaller than in
the case λ = 2 or λ = 4. This is because excess kurtosis is a standardized
measure (by standard deviation)

Figure 4: Pdf of Merton log-returns with different values of µj and λ

22 Modeling Underlying Assets Log-return in Merton...

2.2.2 Estimation of model parameter Θ = (µ, σ, λ, µj , σj )

Estimation of model parameter Θ is required for us to fit empirical data to

the MJD model. From theorem 2.2, we see that the five parameters µ, σ, λ, µj ,
and σj do not have explicit expressions. To address this issue, we will consider
the situation when a single jump occurs i.e. Nt = 1. Let us consider the
empirical log-returns L4t of NASDAQ index data in Figure 2.
We propose a decision rule by [10] about the occurrence of jumps whereby
a threshold  > 0 is chosen by observing the plot of the empirical log-returns,
such that a jump occurs only if the absolute value of the log-returns is greater
than .
Splitting the empirical log-return data into two categories D and J for
a given , the category D includes log-returns with no jump. The category
J includes log-returns with jumps. We will refer to the model parameters
estimated using the decision rule as primary estimates.
Case 1: No occurrence of jumps, i.e. 4Nt = 0.
Here the first and second moments of the log-return are
E∗ [LD4t ] = E∗ [L4t |4Nt = 0]
σ2 (10)
= (µ − )4t
2
and
V ar[LD4t ] = V ar[L4t |4Nt = 0]
(11)
= σ 2 4t
respectively. From equation 10 and 11, we estimate µ and σ as follows:
 b ∗ [LD ]+Var(L
d D )4t
2E 4t 4t
 µ̂ =

24t
r
d D
 σ̂ = Var(L4t )

4t

b ∗ [LD ] and Var[L

where E d D ] are the sample mean and the sample variance of
4t 4t
the empirical log-returns in category D.
Case 2: Occurrence of 4Nt = 1 jump
Here the first and second moments of the log-return are
E∗ [LJ4t ] = E∗ [L4t |4Nt = 1]
σ2 (12)
 
= µ− 4t + µj
2
Megang Nkamga Junile Staures, Jane Akinyi Aduda and Romuald Momeya 23

and

V ar[LJ4t ] = V ar[L4t |4Nt = 1]

(13)
= σj2 + σ 2 4t

respectively. From equation 12 and 13, we estimate µj and σj as follows

b ∗ [LJ ] − (µ̂ − σ̂2 )4t

(
µˆj = E
q 4t 2
J 2
σˆj = Var(L4t ) − σ̂ 4t
d

b ∗ [LJ ] and Var[L

where E d J ] are the sample mean and the sample variance of
4t 4t
the empirical log-returns in category J.
Supposing time is measured in years, the parameter λ is estimated as follows

λ̂ = Average number of jumps per year

Total number of jumps
=
Total length in years

From Figure 2, if we take for example  = 0.01, we obtain the values for
our primary estimators as follows: λ̂ = 60.74, µ̂ = 0.1899, σ̂ = 0.0741, µˆj =
−0.0012 and σˆj = 0.0182.
To get optimal estimators let us consider the case below:
Case 3: Occurrence of 4Nt > 1 jumps
For the estimates of Θ = (µ, σ, λ, µj , σj ) when Nt > 1, we will use the
Multinomial Maximum Likelihood Approach proposed by [8]. To be implement
this approach, we will do the folowing:

1. Divide the empirical data set into categories of length n < length (empirical data),
which has already been achieved above. i.e. D and J.

2. Minimize the objective function by finding an optimal Θ

b that minimizes
the likelihood function
n
Y
L(Θ; x) = − log(P(Li4t ) ∈ B) B ⊂ R
i=0

where the Li4t represent the empirical log-returns.

24 Modeling Underlying Assets Log-return in Merton...

The primary estimators obtained in the previous section is used to numerically

minimize the MJD model objective function. The presence of the sum of
accumulated jumps in the log-return evolution makes it non-normal. In section
2.2, we saw that the probability density function of L4t is given by
2
∞ (L4t −[γ4t+µj a])
X e−λ4t (λ4t)a 1 −
2(σ 2 4t+σ 2 a)
P(L4t ∈ B) = q e j

a=0
a! 2π(σ 2 4t + σj2 a)

Some authors like [11] and [12, 13] highlight that minimization problem is
easily obtained through the regime switching technique. In this paper, the
MATLAB code fminsearch will ease the process of finding an optimal Θb that
minimizes L(Θ; x).

2.3 Diebold-Mariano Test

This test is often used to compare time series models. Suppose we denote
the empirical data set by {yt , t = 1, . . . , N }. Let the samples from BS model
and MJD model be denoted by {b y1t , t = 1, . . . , N } and {b
y2t , t = 1, . . . , N } re-
spectively. We refer to the BS model as model one and the MJD model as
model two. The question we ask ourselves is which forecasting model is actu-
ally good to better represent the empirical data.
Define the forecast errors as

eit = ybit − yt , i=1:2

The loss associated with forecast i is assumed to be a function of the forecast

error, eit , and is denoted by g(eit ). The function g(.) is a loss function, that is
a function such that:

. it takes the value zero when no error is made;

. is never negative;

. is increasing in size as the errors become larger in magnitude.

Typically, g(eit ) is the square (squared-error loss) or the absolute value (absolute-
error loss) of eit . We define the loss differential between the two forecasts by

dt = g(e1t ) − g(e2t )
Megang Nkamga Junile Staures, Jane Akinyi Aduda and Romuald Momeya 25

and say that the two forecasting models have equal accuracy if and only if
the loss differential has zero expectation ∀t. So, we would like to test the null
hypothesis
H0 : E[dt ] = 0, ∀t
versus the alternative hypothesis

H1 : E[dt ] 6= 0, ∀t

The null hypothesis is that the two models have the same accuracy. The
alternative hypothesis is that the two models have different levels of accuracy.
In the case where both models have different accuracy, we perform another
test to detect which model is more accurate than the other using the hypothesis
below:
H0 : Model one and two have the same accuracy
versus the alternative hypothesis

H1 : Model two is more accurate than model one

For the level of significance α = 0.05, we fail to reject the null hypothesis if
the P-value > α

3 Main Results
The two option pricing models (BS and MJD) introduced in section 2 holds
the potential to reproduce their real world counter parts. Here we discuss the
modeling results obtained as describe in section 2.
The model parameters are estimated from NASDAQ index. Table 1 presents
results under the BS model. Under the MJD model, we use the primary pa-
rameters estimated in section 2.2.2 and apply the MATLAB fminsearch code
to obtain optimal parameters for different thresholds . See results in table 2.
From table 2, P.E and Emle represent the model’s primary and MLE es-
timates respectively. We observe that the MLE estimates of the parameters
µ̂, σ̂, λ̂, µˆj and σˆj are the same for  ∈ {0.01, 0.02, 0.03}. For  = 0.06 we obtain
different values of Θ because the maximum number of function evaluations has
been exceeded. The values of  are chosen by taking a close look at Figure 2.
26 Modeling Underlying Assets Log-return in Merton...

Table 1: Parameter estimates for BS model.

Parameters Estimates
µ 0.0005
σ 0.0101

Table 2: MLE parameters with respect to 

 = 0.01  = 0.02  = 0.03  = 0.06
Parameter
P.E Emle P.E Emle P.E Emle P.E Emle
µ 0.1899 0.3615 0.2424 0.3615 0.2170 0.3615 0.1291 0.12914
σ 0.0741 0.0655 0.1176 0.0655 0.1411 0.0655 0.1602 0.1601
λ 60.74 210.8795 15.6372 210.8795 4.2100 210.8795 0 0
µj -0.0012 -0.0028 -0.0076 -0.0028 -0.0216 -0.0028 NaN NaN
σj 0.0182 0.0097 0.0267 0.0097 0.0310 0.0097 NaN NaN

It is logical to think that if there are no jumps like the case of  = 0.06, the
jump parameters would not have estimated values. ¿From this same ”Table”
2, we conclude that the MLE method is independent of the threshold .

Table 3: Merton MLE parameter estimates for different index.

Index  µ σ λ µj σj
NASDAQ [0.01, 0.03] 0.3615 0.0656 210.8795 -0.0028 0.0097
AAPL [0.01, 0.04] 0.4252 0.1522 109.7153 -0.0038 0.0213
AOI [0.01, 0.08] -0.9321 0.1716 305.5716 0.0092 0.0209
BSESN [0.005, 0.02] 0.1607 0.1040 54.7931 -0.0016 0.0084

From table 3, we see that for a given range of , MLE gives us the value
for the model parameters and as such it gives a realiable estimate of λ which
represents the the number of jumps per unit of time (2 years). On average for
example the AAPL stock index will have 54.8576 jumps in a year.
From section 2.1.1 and 2.2.2 and considering the MLE estimates for Nasdaq
index, the values for the moments and mean square error under the BS and
MJD model are given in table 4 and 5 respectively.
From table 4, estimated means and variances of log-returns for the BS
model coincides with the real log-returns. The estimated variances of log-
returns for both the BS model and the real log-returns coincides with MJD-
modeled log-returns.The MJD model captures the negative skewness of the
empirical log-returns since its skewness coefficient is closer to that of real data.
Megang Nkamga Junile Staures, Jane Akinyi Aduda and Romuald Momeya 27

Table 4: Comparing moments of the BS model, MJD model with that of real
data.
Moments Empirical BS model MJD model
mean 0.0005 0.0005 0.0004
Variance 0.0001 0.0001 0.0001
Skewness -0.4658 0 -0.6494
Kurtosis 9.1592 3 5.4869

Also, the kurtosis coefficient for the MJD model seems to be closer to the
empirical kurtosis than for the BS model indicating more pronounce fatter
tails for the log-returns distribution. The mathematical features of the MJD
model makes it possible to handle the large spikes in empirical log-returns, see
Figure 2.

Table 5: Comparing mean square error of BS model and MJD model.

Mean Square Error
BS model 1.8734 × 105
MJD model 1.6821 × 103

Investigating the performance of both models in terms of mean square error,

table 5 clearly shows that the MJD model is better than the BS model since
its mean square error is smaller. Hence, we are tempted to conclude at least
for our empirical NASDAQ data that the MJD model is significantly more
suitable than the BS model for modeling log-returns.
For NASDAQ data considered in this paper, Figure 5 clearly shows that
the distribution of the empirical log-returns is not Gaussian, and the MJD
model seems to reflect reality on the market since its density happens to be
above the kernel density of the log returns. On the other hand, the BS model
happens not to be very suitable in modeling log-returns as it assumes that the
log-returns of the underlying asset is Gaussian and our empirical log-returns
show far more extreme fluctuations than predicted by Gaussian statistics.
The Diebold-Mariano test carried out numerically, we obtained a P-value<
0.01 for the first test which is less than 0.05 showing that both models have dif-
ferent accuracy. For the second test, we still obtain a P-value< 0.01. According
28 Modeling Underlying Assets Log-return in Merton...

Figure 5: Density for IXIC1 time series. Solid blue line: Kernel density es-
timator applied directly on data. Dashed line: MJD model simulation with
estimated parameters. Solid yellow line: BS model simulation with estimated
parameters

to section 2.3, this simply indicates that the MJD model is more accurate than
the BS model.

4 Conclusion
In this paper we have compared the fitness to the data performances char-
acterizing the BS model and the MJD model, which is obtained from the BS
by adding a compensated compound Poisson process to the main stochastic
differential equation. The financial assumptions behind these market mod-
els makes them similar in many ways. Nevertheless, the BS market model is
complete, while the Merton model is incomplete. The latter fact is due to
the impossibility to completely mitigate the risk carried by the introduction
of sudden and unpredictable moves in the stock price. Hence, even if one can
consider the latter as an advantage carried by the BS-approach, at least in
terms of mathematical simplicity and numerical tractability, the MJD model
turns out to outperform the BS model, when one takes into account the per-
formances of the two with respect to real financial data. In particular, moving
Megang Nkamga Junile Staures, Jane Akinyi Aduda and Romuald Momeya 29

from a theoretical comparison to an empirical one, the addition of the jump

parameters results in a great improvement in simulation of log-returns distri-
bution. Also, the log-returns leptokurtic feature is much more evident using
the MJD model approach instead of the BS model when we compare their
density functions with the kernel density estimation for the NASDAQ empir-
ical data. To boost our conclusion, we perform two tasks: first we compare
the moments and density functions of both models with empirical moments
and kernel density respectively and secondly we use the mean square error and
the Diebold-Mariano test to compare both models. Despite the advantages of
MJD model, it does not incorporate the volatility clustering effect. In future
works, we plan to apply the Vasicek model approach similar to the one pro-
posed by Merton, namely taking random jumps into consideration, to what
concerns the interest rates financial frameworks.

ACKNOWLEDGEMENTS. First of all I want to thank God almighty for

his support in my life. I also want to express my huge and sincere gratitude
to my supervisors Dr.Jane Akinyi and Dr.Romuald Momeya for all their wise
suggestions. I must thank them for their time, it was great pleasure to write
this paper under their direction. I also appreciate very much all my teachers
for the precious knowledge which I got from them and my friends for their sup-
port. Special thanks to the Pan African University Institute for Basic Sciences,
Technology and Innovation for the good conditions and support.

References
[1] F. Black and M. Scholes, The pricing of options and corporate liabilities,
Journal of political economy, 81, (1973), 637–654.

[2] R. Cont, Empirical properties of asset returns: stylized facts and statistical
issues, 2001.

[3] N. Gugole, Merton jump-diffusion model versus the black and scholes ap-
proach for the log-returns and volatility smile fitting, International Jour-
nal of Pure and Applied Mathematics, 109, (2016), 719–736.
30 Modeling Underlying Assets Log-return in Merton...

[4] V.G. Kulkarni, Modeling and analysis of stochastic systems, Chapman

and Hall/CRC, 2016.

[5] F. X. Diebold, Comparing predictive accuracy, twenty years later: A per-

sonal perspective on the use and abuse of diebold–mariano tests, Journal
of Business & Economic Statistics, 33, (2015), 1–1.

[6] M. Kijima, Stochastic processes with applications to finance, Chapman

and Hall/CRC, 2016.

[7] H. Allouba, Different types of SPDEs: existence, uniqueness, and Gir-

sanov’s theorem, Cornell University, August, 1996.

[8] F. B. Hanson and Z. Zhu, Comparison of market parameters for jump-

diffusion distributions using multinomial maximum likelihood estimation,
in: 43rd IEEE Conference on Decision and Control (CDC)(IEEE Cat. No.
04CH37601), 4, (2004), 3919–3924.

[9] K. Matsuda, Introduction to merton jump diffusion model, Department

of Economics. The Graduate Center, The City University of New York,
2004.

[10] F. Tang, Merton jump-diffusion modeling of stock price data, 2018.

[11] C. Kim and C. Nelson, State-space models with regime switching: Clas-
sical and gibbs-sampling approaches with applications, 1(0262112388),
(1999).

[12] L. Di Persio and M. Frigo, Maximum likelihood approach to markov

switching models, WSEAS Transactions on Business and Economics, 12,
(2015), 239–242.

[13] L. Di Persio and M. Frigo, Gibbs sampling approach to regime switching

analysis of financial time series, Journal of Computational and Applied
Mathematics, 300, (2016), 43–55.