Russ. J. Numer. Anal. Math. Modelling, Vol. 16, No. 5, pp.

363–384 (2001)
c VSP 2001

Mathematical models of the prices of various

financial instruments in the problem of constructing
the investment portfolio in futures and option trading

S. S. ARTEMIEV and M. A. YAKUNIN

Abstract — We consider various stochastic models of stock prices, financial futures, and rates of
exchange of currencies both in the form of stochastic differential equations and in recurrent form.
Price models are used to construct the stock portfolio, organize futures and option trading. We study
in detail the problem of estimating unknown coefficients in the models and give examples of statistical
modelling.

1. INTRODUCTION
Presently under growing instability conditions for the price development of a vari-
ety of financial instruments such as stocks, currencies, financial futures, etc. spec-
ulations in the stock, exchange, and terminal markets are superrisky financial op-
erations. Small speculators run extra risks, in particular, those trading through the
Internet because small pledged deposits do not allow them to sustain the losses
yielded by the strong unfavourable price development of sold assets. In view of this,
financial analysts have to develop special trading low-risk strategies, make trading
computer programs for generating purchase and sale signals, and use option strate-
gies with fixed losses for the hedging of risks [7].
Stock prices, the rates of exchange of currencies, and espesially the spread val-
ues of one-type financial instruments are affected by a great number of random
factors so that their trajectories in some situations can be regarded as stochastic
ones. This suggests the use of stochastic differential equations (SDE) as mathemat-
ical price models. These mathematical models can be used for different purposes,
for example, for modelling hypothetical market situations when checking the effec-
tiveness of planned financial operations, for developing trading computer programs,
a variety of futures and option strategies [1].
A mathematical model must, on the one hand, be adequate to the historical
prices and on the other hand be sufficiently simple and flexible when used prop-
erly. The stochastic mathematical model is of major importance when checking the
effectiveness of trading strategies and trading computer programs because it allows
 Institute of Computational Mathematics and Mathematical Geophysics, Siberian Branch of the
Russian Academy of Sciences, Novosibirsk 630090, Russia
The work was supported by the Foundation ‘Basic Research-Russian Universities’ (360782).

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
364 S. S. Artemiev and M. A. Yakunin

one to model an ensemble of trajectories, which is impossible in practice where

the only realization of price is possible. Trading computer programs have recently
become vitally important to financiers at any point on the globe because it is pos-
sible to get current quotations from most exchanges in real time and give orders
to a broker through the Internet. Adequate trading computer programs not only en-
sure profits but also help traders avoid nervous breakdowns when taking purchase
and sale decisions because they allow them to considerably reduce the number of
erroneous breaks into a market and withdrawals from it.
The choice of the structure for the mathematical price model of a financial in-
strument is based on the statistical analysis of time series of the historical price
values. The statistical analysis allows one to reveal the specific regularities of the
behaviour of various probabilistic price characteristics, estimate the distribution law
of price values, and choose financial instruments suitable for real profitable trad-
ing. For example, the price develpment of the stocks of most issuers in a separate
stock market at different times is close to a synchronous one, which can be checked
statistically by calculating correlations between specially chosen samples. This is
used for the construction of the optimal investment stock portfolio in the Markowitz
theory [10]. The Markowitz stock portfolio of N issuers is specified by the average
profit
N
r = ∑ xi ri
i=1

and the average risk

v
uN
u N
σr = t ∑ ∑ xi x j cov(ri ; r j)
i=1 j=1

where ri is the annual profitability of the i-th stock, ri is its annual average profitabil-
ity, xi  0 is the nonrandom share of the i-th stock cost in the portfolio, ∑Ni=1 xi = 1.
We assume here that all the random values ri have the Gaussian distribution, which,
however, is inconsistent with the long-term observations of the profitability of stocks
sold in high liquid markets. The use of multidimensional SDE systems in the sense
of Ito allows one to construct the investment portfolio with non-Gaussian distribu-
tion of stock profitability.
The structure of the stochastic price model of a financial instrument is generally
chosen depending on a problem solved by this model. Thus, in order for the cost
of stock options in the Black-Scholes theory to be calculated, the familiar model of
stock price development in the form of the Ito SDE is used [6]:

dSt = µSt dt + σSt dw(t ) (1.1)

where St is the stock price, w(
) is the standard Wiener process, µ and σ are con-
stant parameters called growth and volatility coefficients, respectively. However,

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 365

this model does not take into account the wave character of the stock price devel-
opment at all and cannot be used, for example, for modelling hypothetical market
situations and testing trading computer programs. The wave price development is
observed not only for stocks but also for the rates of exchange of sold currencies
and for the prices of various financial futures. The models of this kind can even be
less suitable for reflecting the price development of spreads constructed by com-
bining purchases and sales of contracts for one-type financial instruments. These
spreads are generally constructed to reduce the risk of trading operations in cur-
rency and terminal markets and characterized by the explicit wave development
of spread prices. The wave developments of the solutions of ordinary differential
equations can be introduced in different ways: natural oscillations, parametric os-
cillations, self-oscillations, forced oscillations. The same methods can be used in
models of random processes, which are solutions of the SDE.
Since the mathematical price models of financial instruments are used in the
form of SDE systems, there arises a problem of estimating their parameters by price
observations. The present work generalizes the results of [2, 12] to the systems of
nonlinear SDEs with linearly dependent unknown parameters in them. The models
of this type include most of the below-studied models of stock prices, currencies,
financial futures, which take into account the wave character of their dynamics, and
the linear multidimensional price model which is a generalization of the Black-
Scholes model. Estimates are derived by the maximum likelihood method. In this
case the observation model is given as the simplest discrete analogue of the SDE
solution, viz. the generalized Euler scheme. We emphasize that in real exchange
trading the price observations are in essence discrete (e.g. five minutes’, daily, etc.)
ones. Therefore a time discrete scheme is suitable as a price model and the use of
continuous models in the form of SDEs is somewhat of an auxiliary character. The
maximum likelihood estimates of unknown parameters are calculated by a stable
iterative scheme, and in most special cases it is possible to obtain explicit formulae
for estimates and their variances [4].
Using the multidimensional model of stock price development in the form of
SDEs, in the present work we study a new approach to the solution of the problem
of constructing the optimal investment stock portfolio. Its idea is to specify a risk
characteristic which is more general than the standard deviation of the portfolio
profitability in the Markowitz theory, viz. the possibility that profitability turns out
to be lower than the permissible value. In this case, using the Monte Carlo method,
we can construct the set of permissible portfolios in the plane (profitability, risk) and
solve a problem for the non-Gaussian distribution of stock profitabilities [5]. Below
we give an example of the construction of the concrete set of permissible portfolios
and examples of the calculation of the parameter estimates in the concrete models
of stock price development and the spread in prices of currency futures, using the
model and real price observations.

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
366 S. S. Artemiev and M. A. Yakunin

2. ESTIMATES OF THE PARAMETERS LINEARLY DEPENDENT

ON THE SDE SYSTEM OF GENERAL FORM

Suppose the financial instrument price develpment model is given as the SDE sys-
tem in the sense of Ito
dY (t ) = (V (t ; Y ) + F (t ; Y )Θ)dt + G(t ; Y )CdW (t ); t0
(2.1)
Y (0) = Y0
where Y (
) is an N-dimensional random process observed, W (
) is the M-dimen-
sional standard Wiener process, Θ is the K-dimensional vector of constant param-
eters, C is the constant parameter matrix of dimension N  M, V (t ; Y ) is an N-
dimensional vector function, F (t ; Y ), G(t ; Y ) are function matrices of dimensions
N  K and N  N, respectively.
The method of obtaining the maximum likelihood estimates of the unknown pa-
rameters Θ, C is based on the Euler scheme for the SDE (2.1) with the use of the
exact observations of the single trajectory Yi = Y (ti ) at discrete instants ti+1 = ti + hi ;
i = 0; : : : ; N p 1; t0 = 0, where the number of observations is N p  max(N ; K ; M ).
Restrictions imposed on the SDE parameters and observations become obvious in
the following. We write the approximate solution of the SDE (2.1), using the gener-
alized Euler method:
p
Yi+1 = Yi + Vi hi + Fi Θhi + GiC hi ξi ; i = 0; : : : ; N p 1 (2.2)
where Vi = V (ti ; Yi ), Fi = F (ti ; Yi ), Gi = G(ti ; Yi ), fξi g is a sequence of independent
standard Gaussian vectors with independent components. The maximum likelihood
estimates of the parameters Θ, C are their values Θ = Θ, b C =C b such that the mul-
tidimensional probability distribution density of the Markov sequence (2.2) p(Y0 ;
Y1 ; : : : ; YNp ) reaches a maximum. Following [11] this density can be represented as
Np 1
p(Y0 ; Y1 ; : : : ; YNp ) = p(Y0 ) ∏ p(Yi+1 jYi ) (2.3)
i=0

where p(Y0 ) is the vector distribution density of the initial conditions, which is as-
sumed to be independent of Θ and C, p(Yi+1 jYi ) is the transition probability density
of the Markov sequence (2.2). Suppose that M  N, the matrices D = CCT and Gi ,
i = 0; : : : ; N p 1; are not singular. Then

p(Yi+1 jYi ) = (2π)

N =2
[det(hi GiCC
T
GTi )℄ 1=2
n
 exp 1
2hi
(Yi+1 Yi Vi hi FiΘhi )T
o
(GTi ) 1
(CC )
T 1
Gi 1
(Yi+1 Yi Vi hi FiΘhi ) (2.4)
i = 0; : : : ; N p 1:

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 367

The problem of maximizing the function (2.3) is equivalent to the problem of

maximizing the logarithmic likelihood function ∑i=p0 ln p(Yi+1 jYi ). On substituting
N 1

p(Yi+1 jYi ) from (2.4) this function reduces to the minimization problem:
Np 1
Ri Θ) ! min
1
Φ(Θ; D) = ln det D +
Np ∑ (Zi RiΘ)T D 1
(Zi
D>0; Θ
(2.5)
i=0

where D > 0 implies the positive definiteness of the matrix D and

p
Zi = p1 Gi 1
(Yi+1 Yi Vi hi ); Ri = hi Gi 1 Fi ; i = 0; : : : ; N p 1: (2.6)
hi

b ; D = D,
If the function Φ(Θ; D) reaches a unique minimum for the values Θ = Θ b
these values are the maximum likelihood estimates of the given parameters. The
maximum likelihood estimate Cb satisfies the equation CCT = D.
b

Theorem 2.1. Suppose the observations Yi and the values Vi , Fi , Gi ; i = 0; : : : ;

Np 1 are such that the function matrix

Np 1
Ψ(Θ) = ∑ (Z i RiΘ)(Zi RiΘ)T
i=0

is positive definite for all Θ 2 RK . Then the function Φ(Θ; D) in the domain Θ 2 RK ,
D > 0; reaches an absolute minimum.
The proof of Theorem 2.1, which is given in [3], is based on reducing the prob-
lem of minimizing the function Φ(Θ; D) to the problem of minimizing the function
det Ψ(Θ) with respect to the single vector parameter Θ. The latter, being a positive
polynomial function, has an absolute minimum. Given the necessary condition for
a minimum of the function (2.5) ∂Φ=∂Θ = 0, ∂Φ=∂D = 0, we obtain the system of
equations for the unknown parameters Θ, D:

(2:7)

Np 1 Np 1
∑ RTi D 1 Ri Θ ∑ RTi D 1 Zi = 0 (2.7)
i=0 i=0

(2:8)

Np 1
1
D=
Np ∑ (Z i Ri Θ)(Zi Ri Θ)T : (2.8)
i=0

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
368 S. S. Artemiev and M. A. Yakunin

Theorem 2.2. Let the hypotheses of Theorem 2.1 be true. Then the necessary
uniqueness condition for a minimum of the function Φ(Θ; D) is the nonsingularity
N 1
of the matrix R = ∑i=p0 RTi Ri .

Proof. Let the matrix R be singular. By Theorem 2.1 the function Φ(Θ; D)
reaches an absolute minimum at least at one point. The coordinates of this point
Θ = Θ1 , D = D1 satisfy the system of equations (2.7), (2.8). The singularity of
N 1
the matrix R is equivalent to that of the matrix ∑i=p0 RTi QRi , where Q is any sym-
metrical positive definite matrix of dimension N  N. This implies that the vector
equation (2.7) for Θ with fixed D = D1 has a continuum of solutions along with the
solution Θ = Θ1 . The equality holds for each solution Θ :
Np 1
(Θ1 Θ )T ∑ RTi D1 1 Ri (Θ1 Θ ) = 0
i=0

from which it follows that

Ri Θ = Ri Θ1 ; i = 0; : : : ; N p 1:

Therefore the function Φ(Θ; D) according to (2.5) has the same minimum value
ln det D1 + N for D = D1 and each Θ = Θ , i.e. it has a nonunique minimum. This
implies that the minimum uniqueness entails the nonsingularity of the matrix R.
This completes the proof of Theorem 2.2.
Note that the condition R > 0 is not sufficient for the minimum of the function
Φ(Θ; D) to be unique because the polynomial det Ψ(Θ) may have more than one
minimum in this case as well. We write the system of equations (2.7), (2.8) as
 Np 1  1 Np 1
Θ= ∑ RTi D 1 Ri ∑ RTi D 1 Zi (2.9)
i=0 i=0

Np 1
1
D=
Np ∑ (Z i Ri Θ)(Zi Ri Θ)T : (2.10)
i=0

By Theorem 2.1 the solution of this system exists. It is calculated by the iteration
method and in this case we can use, for example, an identity matrix as the initial
guess D0 in (2.9). The value of the function (2.5) at each iteration does not increase.
If after the first iteration we obtain the approximations of the parameters Θ1 , D1
and after the second iteration those of Θ2 , D2 , then Φ(Θ1 ; D1 )  Φ(Θ1 ; D0 ) due to
the attainment of the unique minimum of the function Φ with respect to D for fixed
Θ = Θ1 , and Φ(Θ2 ; D1 )  Φ(Θ1 ; D1 ) due to the attainment of the unique minimum
of the quadratic form with respect to Θ in (2.5) for fixed D = D1 . In practice, the
solution of the system of equations (2.9), (2.10) is generally unique and only a few
iterations are needed to calculate its sufficiently exact approximation.

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 369

Since the transition probability density of the Markov process Y (t ) depends on

CCT rather than on C, it is possible to find the estimate Cb only for such types of ma-
trices that are uniquely reconstructed on the basis of D. b If C is given by a quadratic
b
triangular matrix, we can calculate the estimate C, using the Cholesky factorization
[9]. We consider special cases of the SDE (2.1).

Case 1. N = 1, M = 1. The scalar SDE has the form

dy(t ) = (v(t ; y) + f T (t ; y)Θ)dt + g(t ; y)c dw(t )
where f T (t ; y) is a row vector of the functions of dimension K and v(t ; y), g(t ; y) are
scalar functions. From equations (2.9), (2.10) we derive the estimates
 Np 1  1 Np 1
hi yi+1 yi hi vi
b=
Θ ∑ fi f T
g2i i
∑ g2i
fi (2.11)
i=0 i=0

1
Np 1
(yi+1 yi hi vi b )2
hi fiT Θ
cb2 =
Np ∑ hi g2i
: (2.12)
i=0

Case 2. M = N and the matrix C in the right-hand side of the SDE (2.1) is a
diagonal one with elements c j ; j = 1; : : : ; N. In this case the maximum likelihood
method leads to the system of equations: the vector Θ is defined by (2.9), D 1 is a
diagonal matrix with elements 1=c2j , instead of (2.10) we have the equations
Np 1
1 K
2
c2j =
Np ∑ zi j ∑ ri jq θq ; j = 1; : : : ; N
i=0 q=1

where zi j ; ri jq , j = 1; : : : ; N, q = 1; : : : ; K ; are the elements of the vector Zi and the

matrix Ri , respectively. If in this case the general formulae (2.9), (2.10) are used to
calculate the estimates by the iteration method, according to (2.10) only the diagonal
elements of the matrix D are defined and the off-diagonal elements at each iteration
are taken to be zero.

Case 3. The SDE (2.1) is given as the system with additive noise:
dY (t ) = (V (t ; Y ) + AU (t ; Y )) dt + C dW (t ) (2.13)
where V (t ; Y ) and U (t ; Y ) are N-dimensional and L-dimensional vector functions,
respectively, A and C are the matrices of unknown constant parameters of dimen-
sions N  L; N  M, respectively. The maximum likelihood estimates of the param-
eters A, D = CCT can be derived by finding the unique solution of the system (2.9),
(2.10). We represent AU (t ; Y ) as
AU = FΘ

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
370 S. S. Artemiev and M. A. Yakunin

where the vector Θ of dimension NL is obtained by ordering the matrix A with

respect to columns, the matrix F of dimension N  NL has the structure

. u2 I ..


.. uL I ):
.. . .
F = (u1 I

For this matrix we have

FT D 1
= QF
T
(2.14)

where Q is a block diagonal matrix of dimension NL  NL whose blocks are ma-

trices D 1 . Putting Gi = I in (2.6) for all i and taking into account (2.14), (2.6), we
represent (2.9) as the equation
Np 1 Np 1
Q ∑ hi FiT Fi Θ = Q ∑ FiT (Yi+1 Yi hiVi ): (2.15)
i=0 i=0

If b is an arbitrary N-dimensional vector, we can transform the NL-dimensional vec-

tor FiT b to the matrix bUiT of dimension N  L, where Ui = U (ti ; Yi ). If we transform
the vector equation (2.15) in this way, cancel Q in it, and substitute AUi for Fi Θ, we
find from (2.15) the estimate
Np 1  Np 1  1
b=
A ∑ (Yi+1 Yi hiVi )UiT ∑ hiUiUiT (2.16)
i=0 i=0

and from (2.10) the estimate

Np 1
1 1
b=
D
Np ∑ hi
(Yi+1 Yi hiVi b i )(Yi+1
hi AU Yi hiVi b i )T :
hi AU (2.17)
i=0

3. STOCK PRICE DEVELOPMENT MODEL WITH VARIABLE

VOLATILITY

One of the ways of improving the stock price development model (1.1) in order to
bring it into line with real prices is to transform the parameter σ in the right-hand
side of equation (1.1) to a time function or to a random process. The model of this
type given in [8] is the system of two SDEs in the sense of Ito:
p
dSt = µSt dt + νt St dw1 (t )
p (3.1)
dνt = γ νt dt + δνt (ρdw1 (t ) + 1 ρ2 dw2 (t ))

where St is the stock price, νt is the unobserved varying volatility, µ, γ, δ, ρ are

unknown real parameters, and δ > 0, 1  ρ  1. The initial stock price S0 > 0 is

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 371

known, the initial value ν0 = σ2 is not known. We represent the SDE system (3.1)
as

dy1 (t ) =
p
θ1 y1 dt + c11 y2 y1 dw1 (t )
(3.2)
dy2 (t ) = θ2 y2 dt + c21 y2 dw1 (t ) + c22 y2 dw2 (t )

= St , y2 (t ) = νt =σ2 , the unknown parameters are θ1 = µ, θ2 = γ, c21 =

where y1 (t )p
δρ, c22 = δ 1 ρ2 , c11 = σ. Then this SDE system is a special case of the system
(2.1) in which

T
T
Y (t ) = y1 (t ); y2 (t ) ; V (t ; Y )  0; Θ = θ1 ; θ2


y1 0
  py y1 0
 
c11 0

F = ; G= 2 ; C=
0 y2 0 y2 c21 c22

with initial conditions y1 (0) = S0 , y2 (0) = 1. We now pass from Y (t ) to the random

T
process Z (t ) = y1 (t ); ln y2 (t ) . Using the formula for changing the variables in
the stochastic Ito integral, we obtain the SDE system which is also a special case of
(2.1):

dz1 (t ) = p1 z1 dt + c11 ez2 2 z1 dw1 (t )

=

(3.3)
dz2 (t ) = p2 dt + c21 dw1 (t ) + c22 dw2 (t )

where p1 = θ1 , p2 = θ2 (c221 + c222 )=2, c11 , c21 , c22 are unknown parameters. This
system is preferable to (3.2) when calculating the parameter estimates because the
difference Euler scheme yields an exact solution for the second equation of the
system (3.3) rather than an approximate one.
It is impossible to estimate all the unknown parameters in (3.3) when the real
observations of only the first component of the solution (stock price) are available.
Table 1 gives an example of the maximum likelihood estimates (MLE) of the param-
eters, which are calculated by the modelled observations of two components fz1 (ti ),
z2 (ti )g. The observation interval is T = 500 days, the initial stock price is S0 = 10,
the number of uniform observations is N p = 500, the number of steps of modelling
the solution between two adjacent observations is 300. The maximum likelihood
estimates of the parameters in the initial system (3.1) are calculated, using the max-
imum likelihood estimates pb1 , pb2 , cb11 , cb21 , cb22 , by the formulae

p cb21 b
δ2
b
µ = pb1 ; σ
b = cb11 ; b
δ= cb21 2 + cb22 2 ; ρ=
b ; γ = pb2 +
b :
b
δ 2

Eight correct significant digits of the exact values of the estimates are calculated
at three iterations. The last column in Table 1 gives the values of the estimates µ̃, σ̃

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
372 S. S. Artemiev and M. A. Yakunin

Table 1.

Parameter Exact value MLE for SDE system (3.3) Separated MLE for each equation

µ 0.0002 0.00016957 0.00017011

γ 0.0005 0.00052243 0.00052083
σ 0.0004 0.00039043 0.00039043
δ 0.002 0.0019320 0.0019320
ρ 0.6 0.61211

calculated by the same observations, using only the first equation of the SDE system
(3.3), and the values of γ̃, δ̃ calculated by using only the second equation of the same
system. These values, as seen from the table, differ little from the estimates of the
corresponding parameters for the complete system (3.3). For the estimates δ̃2 , γ̃ the
mean values of the squares of their errors ε2δ̃2 , ε2γ̃ are determined by the formulae

2δ4 δ4 δ2 δ4 δ4
ε2δ̃2 = ; ε2γ̃ = + :
Np N p2 T 2N p 4N p2

In the example considered we have εδ̃2 = 2:5286  10 7 , εγ̃ = 8:9443  10 5 . The

same values can be taken for the errors of the estimates b δ2 , bγ. Note that the parameter
ρ can be estimated, using only both equations of the system (3.3).
When used in practice the model (3.1) has a drawback, as mentioned above: the
second component of the solution of the SDE system is not observed and hence it
is impossible to identify the parameters in the second equation of the system. If we
assume in the above example that the observations of the stock rate fz1 (ti )g only
are known and put z2 (t )  0 in the first equation of the SDE system (3.3), we obtain
the estimate value 0.00017007 for the parameter µ and a somewhat underestimated
value 0.00036493 for the parameter σ.

4. MULTIDIMENSIONAL MODEL OF THE STOCK PRICE

DEVELOPMENT

We take the model of the stock price development of a single issuer in the form of
the Ito SDE (1.1) as a base one. If we pass from St to the random process y(t ) =
ln St , use the formula for changing the variables in the stochastic Ito integral, we can
rewrite the SDE (1.1) as

dy(t ) = p dt + σ dw(t ); 0t T

y(0) = y0

where T is the observation interval, p = µ σ2 =2, y0 = ln S0 , S0 > 0. The natural and

simplest generalization of this model to the case of the stock portfolio of N issuers

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 373

fSti g is the SDE system

dY (t ) = P dt + C dW (t ); 0t T
(4.1)
Y (0) = Y0
where
Y (t ) = (ln St1 ; ln St2 ; : : : ; ln StN )

Y0 = (ln S01 ; ln S02 ; : : : ; ln S0N )

and W (
) is the N-dimensional standard Wiener process, P is the N-dimensional vec-
tor of constant parameters, C is the N  N nonsingular matrix of constant parameters
which is further given by a lower triangular matrix for definiteness. The proposed
multidimensional model (4.1) is more suitable than the generalization of the model
(1.1) in the form of the linear system of Ito SDEs with multiplicative noise because
an exact solution of (4.1) can be written. According to the model (4.1) each of the
components Sti , given the initial price S0i , has a logarithmically normal distribution
and the average stock price development is exponential:
 1 i 2
hSti i = S0i exp pi +
2 k∑
cik t ; i = 1; : : : ; N :
=1

The SDE system (4.1) is special case 3 of the system of the general form (2.1), i.e.
the system (2.13) in which L = 1, V (t ; Y )  0, U (t ; Y )  1, A = P. The maximum
likelihood estimates (2.16), (2.17) in this case have the form
1
Pb = (YNp Y0 ) (4.2)
T

Np 1
1 1
b=
D
Np ∑ hi
(Yi+1 Yi b i )(Yi+1
Ph Yi b i )T :
Ph (4.3)
i=0

The estimate Cb satisfying the equation CCT = D b can be found, using the Cholesky
factorization. All the obtained maximum likelihood estimates are exact in that the
Euler scheme yields an exact solution of the SDE system (4.1).
The estimate Pb is unbiased. The bias of the estimate D
b can be eliminated, multi-
plying it by the value N p =(N p 1). The correlation matrix of the estimate errors for
the elements of the vector P is defined as

h(Pb P)(Pb P)T i =

1 T
T
CC

i.e. the estimates pbi have the variances

σ2pbi  h( pbi pi )2 i =
dii
T
; i = 1; : : : ; N : (4.4)

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
374 S. S. Artemiev and M. A. Yakunin

Using (4.3), for the quadratic mean errors of the estimates for the elements of the
matrix D we can obtain the formulae
dii d j j + di2j
ε2db
ij
h(dbi j di j ) 2
i= Np
dii d j j
N p2
; i; j = 1; : : : ; N (4.5)

where di j = d ji = ∑Nk=1 cik c jk . It is obvious from relations (4.4), (4.5) that the ac-
curacy of the estimates dbi j increases as the number of observations N p increases
regardless of T and the accuracy of pbi increases as the observation interval T in-
creases regardless of N p . If the number of real observations is large, the estimates
dbi j turn out to be sufficiently exact as it follows from (4.5). In particular, for the
diagonal elements of the matrix D the quadratic mean errors of their estimates are
p p
εdbii  2dii = Np:

Therefore the errors of the estimates pbi , dbi j can be calculated, replacing the un-
knowns di j in formulae (4.4), (4.5) by the estimates dbi j . In practice, as is often the
case, the accuracy of the estimates pbi is not satisfactory
q because the values p bi differ
insignificantly from zero due to the large values of dbii =T comparable with j pbi j.
As mentioned above, the stock price model in the form of the SDE with constant
parameters is not always adequate to the real historical prices. This can be confirmed
by a large variation of the estimates of the model parameters when the volume of the
historical prices used for estimation changes. Thus, weekly historical volatility can
differ significantly from monthly volatility which can in turn differ from quarterly
one. It should be noted that the average stock price in some cases is approximated
by an exponential function especially inaccurately. In order for the model adequate
to the historical prices to be constructed, it is sufficient for the parameters P and C
in the model (4.1) to be time dependent:
dY (t ) = P(t )dt + C(t )dW (t ): (4.6)
If the price changes in the stock market are almost of a periodic character, the func-
tions P(t ), C(t ) are also of an analogous periodic character. The variable parameters
can be estimated by formulae (4.2), (4.3), using a ‘sliding window’ of size N p . Mov-
ing it along the sequence of the historical price observations, we can obtain a table
of the values of the parameters fPbi , Cbi g, which corresponds to the last time points
fti g. As the size of the ‘window’ increases, the parameter estimates become increas-
ingly smoother and change more slowly. Given the large volumes of the historical
prices observed for more than a year, the estimates of the parameters fPbi , Cbi g can be
derived for the whole year. When modelling a hypothetical market situation for the
year ahead it is possible to use both the parameter estimates derived by the historical
prices and those given by other means depending on the extent to which the future
behaviour of the prices is assumed to be close to their past behaviour. We empha-
size that stock prices unlike bond prices do not depend on the presence of days off

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 375

and holidays in the trading process. This allows us to use the uniform time grid fti g
with constant step h when modelling the prices and estimating the parameters, only
trading days taken into account. Using the model with variable parameters, we can
give any average behaviour of the stock price, and not just the exponential one as
is the case with constant parameters. The distribution law of stock prices remains
unchanged, i.e. it is logarithmically normal.

5. CHARACTERISTICS OF THE RISK AND PROFITABILITY

OF THE STOCK PORTFOLIO

The prospective annual stock profitability (without dividend payment) ri ,

i = 1; : : : ; N, of each of the i-th issuer is specified by the model (4.6) as

eyi (T ) ey0i
ri =
STi S0i
S0i
 100% =
ey0i
 100% = (eψ i (T )
1)  100% (5.1)

where S0i is the current price of the i-th stock, STi is its prospective price in a year
(T = 1), ψi (T ) are the elements of the vector

ZT ZT
Ψ(T ) = Y (T ) Y0 = P(t )dt + C(t )dW (t ):
0 0

The random vector Ψ(T ) is a normal one with the mathematical expectation
ZT Np 1
M (T )  hΨ(T )i = P(t )dt  ∑ P(ti )hi
i=0
0

and with the correlation matrix

ZT
D(T )  h(Ψ(T ) M (T ))(Ψ(T ) M (T )) T
i= C(t )CT (t )dt
0
Np 1
 ∑ C(ti )CT (ti )hi
i=0

where the values of P(ti ), C(ti ) are either obtained by the price observations over the
previous year or given hypothetically to reproduce the concrete future behaviour of
stock prices. According to (5.1) the stock profitabilities ri are dependent logarithmi-
cally normal random values with mathematical expectations

ri  hri i = (emi (T )+dii (T )

= 2
1)  100%; i = 1; : : : ; N
where mi (T ) are the elements of the vector M (T ) and dii (T ) are the diagonal el-
ements of the matrix D(T ). If xi is the cost share of the i-th issuer in the stock

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
376 S. S. Artemiev and M. A. Yakunin

portfolio, the prospective annual profitability r of the stock portfolio of N issuers is

specified by the formula
N
r = ∑ xi ri
i=1

and the average profitability r by the formula

N  N 
r = ∑ xi ri = ∑ xi emi (T )+dii(T ) = 2
1  100% : (5.2)
i=1 i=1

Since, given ri , the portfolio is fully specified by the set of weights fxi gNi=1 , in the
following the concrete portfolio implies the concrete set of values fxi g. As a mea-
sure of the portfolio risk, we use the probability p f of the fact that the portfolio
profitability r turns out to be lower than the permissible value r f :
n N  o
pf =P fr  r f g = P ∑ xi eψ i (T )
1  100%  r f : (5.3)
i=1

Recall that in the classical theory of the optimal Markowitz portfolio the standard
deviation of the random value r is used as a measure of the portfolio risk. In the
case of the non-Gaussian distribution of the portfolio profitability the standard devi-
ation incorrectly reflects the investor’s risk associated with losses or narrow margin
which, for example, may be narrower than the bank rate. The suggested criterion
for risk estimation is close to the criterion VaR, viz. losses, given the risk level. The
problem of constructing the optimal portfolio is to find the cost shares fxi g of all the
issuers in the portfolio for which the average profitability (5.2) reaches a maximum
value, given the risk level (5.3), or the risk level is minimal, given the average prof-
itability (a dual problem). If the stock prices are independent, i.e. the matrix D(T )
is diagonal, the distribution law of the random value r is defined by the convolution
of laws of its separate terms distribution. In the special case of the identity of sta-
tistical characteristics and the independence of the stock prices of all N issuers the
minimum risk level p f is attained for x1 = x2 = : : : = xN = 1=N. In the general case
the distribution law of the portfolio profitability can be deduced statistically in the
same way as the optimal portfolio can be found.

6. CONSTRUCTING THE SET OF PERMISSIBLE PORTFOLIOS

AND FINDING THE OPTIMAL PORTFOLIO

We construct the set of permissible portfolios in the Markowitz theory, using de-
terminate calculations because we suppose that the distribution law of the portfolio
profitability is the Gaussian one and the characteristic of the portfolio risk is the
standard deviation of the portfolio profitability. In order for the set of permissible
portfolios to be constructed it is necessary only to give the vector fri g and the matrix

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 377

pf
0.42
0.4
0.38
0.36
0.34
0.32
0.3

15 20 25 30 35 40 r; %

Figure 1. Domain of permissible portfolios of three stocks.

cov(ri r j ). In the case of the non-Gaussian law the calculation of the risk-profitability
characteristics (r; p f ) for the given portfolio fxi gNi=1 involves the use of the Monte
Carlo method for modelling the stock profitabilities (5.1) and the calculation of p f
according to (5.3). In order to reduce the computational cost of the algorithm for
statistical modelling we should use each trajectory of the solution of the SDE (4.6)
from the modelled ensemble for all the portfolios fxi g at once. If each portfolio fxi g
in the plane (r; p f ) is associated with a single point, the filled domain is the set of
permissible portfolios. In principle, this set is continuous, however, the computa-
tions on the finite set of values of fxi g yield its discrete analogue. The closed set
of permissible portfolios is inside the rectangle [min1iN ri , max1iN ri ℄  [0; 1℄.
It is logical that the portfolios with optimal risk-profitability characteristics should
be sought at the boundary of the set of permissible portfolios. The boundary of any
set of permissible portfolios includes at least a single point with coordinates (r ; pf )
such that r = max r for fixed p f = pf and pf = min p f for fixed r = r are attained
at it simultaneously. The totality of all the boundary points with this property forms
the effective boundary of this set of permissible portfolios. The effective boundary
is generally the increasing function p f (r), i.e. the portfolio risk increases as prof-
itability increases.
Given in Fig. 1 is the example of the set of permissible portfolios, which is con-
structed by the statistical modelling algorithm. The portfolio consists of three Rus-
sian stocks. To describe the stock price development, for ease of calculations we
use the model (4.1) with constant parameters, which is not a restriction in this ex-
periment. The values of the model parameters are p1 = 0:22, p2 = 0:18, p3 = 0:14,
c11 = 0:45, c21 = 0:15, c22 = 0:30, c31 = c32 = 0, c33 = 0:15. The variation step xi
is 0.025, the number of tests is 105 , r f = 10%. The calculated annual average stock

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
378 S. S. Artemiev and M. A. Yakunin

profitabilities are r1 = 37:89%, r2 = 26:65%, r3 = 16:33%. The effective boundary

of the domain of permissible portfolios in Fig. 1 is between the points with coordi-
nates (21.37, 0.2910) and (37.89, 0.3928). Thus, for the portfolio with profitability
25% the minimum risk is 0.31. This portfolio is associated with the set x1 = 0:3,
x2 = 0:2, x3 = 0:5. Note that the apparent three-dimensionality of the shown do-
main is due to the algorithm iterations and occurs after the neighbouring points
corresponding to permissible portfolios are joined by lines (a total of 861 versions
of permissible portfolios is calculated).
It is not quite proper to compare the sets of permissible Markowitz portfolios
with those constructed in this work because they are represented in different planes
(r ; σr ), (r; p f ) and have an absolutely different form as the calculations show. Be-
sides being used for the construction of the set of permissible portfolios, the Monte
Carlo method can also be used to calculate the cost development of the concrete
stock portfolio fxi g in a given hypothetical market situation, i.e. when the stock
price behaviour due to the corresponding choice of the values of P(ti ) and C(ti ) is
given. The main advantage of the proposed method of constructing the stock portfo-
lio is the possibility to construct a portfolio suitable for the predicted future market
situation rather than a portfolio based on the statistical data obtained in the past
years.

7. SPREAD IN PRICES OF TWO CURRENCY FUTURES AND

ITS STOCHASTIC WAVE MODELS

Nowadays the most popular and liquid currencies in the World trading are Euro
(EC), British Pound Sterling (GBP), Swiss Franc (SFR), Japanese Yen (JPY) quoted
in US dollars. The observations of rates of exchange of the above currencies for a
long period allow one to make a conclusion about their significant statistical interde-
pendence at separate periods except, perhaps, Japanese Yen whose rate of exchange
varies quite independently of the other main World currencies.
This interdependence can be used to restrict the trading risk, for example, by
constructing the spreads of two currencies. Currency spread implies a trading op-
eration of the simultaneous purchase of some amount of one currency and the sale
of the corresponding amount of the other currency. Besides, the term spread is used
below to denote the current cost of open positions in this trading operation. Risk in
spread trading is reduced by compensating for losses in one currency by profits in
the other. The use of spreads to reduce trading risk is common practice in futures
and option markets. Not every spread can be used for limited risk trading. The main
requirement for spread is the absence of long-term trends in one direction and the
presence of fluctuations about an average level. The ideal spread is a sine curve with
a sufficiently small period and a sufficiently large amplitude. The period and the am-
plitude of spread fluctuations define the risk and profitability of a trading operation.
The appropriate trading spreads can be based on pairs of one-type financial instru-
ments which have a strong long-term correlation or, in other words, on assets whose
price development is sufficiently syncronous. The moments of the narrow and wide

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 379

margin of fluctuation in prices become those of purchase and sale.

Currency contracts are sold both in the spot-market (FOREX) and in the futures
market (CME–Chicago Merchantile Exchange). Since the exchange sale is better
organized and the spot-rates of exchange of currencies and the prices of the corre-
sponding currency futures are rigidly connected through the bank interest rate, in
this work currency futures contracts are used as base financial instruments.
With organized exchange sale the quite definite values, viz. 62500 GBP or
125000 EC are used as a unit of trading or contract volume. Therefore organiz-
ing the spread of two currencies, it is necessary to take into account the volume and
the number of traded contracts. The formula for profits/losses has the form
ProfT = N1 (S10 S1T ) + N2 ( S20 + S2T ) (7.1)
where N1 is the quantity of the first currency, S10 is the sale price of the first currency
in US dollars, S1T is the purchase price of the first currency within the interval T .
The same is true for the second currency, with sale replaced by purchase. Using
formula (7.1), we can write
ProfT = δ0 δT (7.2)
where δt = N1 S1t N2 S2t , t = 0; T . If the parameter t runs from 0 to T , then δt spec-
ifies the dynamics of profits/losses of a financial operation in [0; T ℄. In the following
it is the value δt that is called the spread of two currencies.
When constructing spread it is always supposed in the following that δ0 > 0
(the cost of the first currency is higher than that of the second one). Then the above
operation of the simultaneous trading of two currencies is profitable for δ0 > δT .
If at the initial instant the first currency is purchased and the second one is sold,
formula (7.2) is written as
ProfT = δ0 + δT (7.3)
and the operation is profitable for δ0 < δT . Thus, if the behaviour of the trajectory
δt is similar to that of a stationary process, for profitable trading to be organized one
should purchase spread (the purchase of the first currency and the sale of the second
one) when the spread value is small and sell spread (the sale of the first currency
and the purchase of the second one) when the spread value is large.
If we use the model (1.1), the simplest way for assessing the wave price de-
velopment of a financial instrument is to introduce in this model the parametric
oscillations:
dP(t ) = (µ + A cos ωt )P(t )dt + σP(t )dw(t ); P(0) = P0 (7.4)
where P(t ) is the random process of the price alternation of the financial instrument
studied, for example, a stock, currency, spread in prices and so on. The SDE (7.4)
has the exact solution
A

P(t ) = P0 exp (µ 0:5σ2 )t + sin ωt + σw(t ) : (7.5)
ω

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
380 S. S. Artemiev and M. A. Yakunin

In random processes of the form (7.5), it is possible to attain the necessary wave
trend of the probabilistic characteristics of the process P(
) by varying the parame-
ters µ, σ, A, ω.
If the spread dynamics δt is regarded as random, as its general mathematical
model we can use the Ito SDE in the form

dδt = a(t ; δt )dt + σ(t ; δt )dw(t ) + c(t ; δt )dν(t ) (7.6)

where w(
) is the standard Wiener process, ν(
) is the standard Poisson point pro-
cess. The functions of drift a(t ; x), diffusion σ(t ; x), and jump c(t ; x) are defined for
a concrete pair of traded currencies. Much attention is given to the average trend
of spread, the dynamics of its variance, the law of spread value distribution. The
Poisson component allows us to take into account the abrupt spread developments
in the model at separate random instants for the chosen distribution law. In the gen-
eral case when constructing a mathematical price model it is necessary to take into
account such characteristics as the durations and the values of trends, time intervals
between transactions, time intervals between jumps and the values of jumps, the
number of corrections and reversals, the increment value of neighbouring transac-
tions, the maximum deviation of price from the initial one within the given time
interval.
The model (7.6) implies implicitly that the market is superliquid. This allows us
to replace the discrete model by a continuous one and disregard the differences in
the prices of supply and demand at any instant t. It is sufficient to use the generalized
Euler method for passing to the discrete model.
If spread is regarded as a stochastic oscillator, then as an alternative to the SDE
(7.6) we can use, as a spread model, the Van der Pol SDE which describes the
behaviour of the stochastic self-sustained oscillation system [11]:

dδt = ϕt dt
(7.7)
dϕt = (ε(t )ϕt (1 p(t )δt2 ) + q(t )δt )dt + σ(t ; δt )dw(t )

where ε(t ), p(t ), q(t ), σ(t ; x) are given regular (nonrandom) functions whose choice
specifies a certain behaviour of probabilistic characteristics of self-oscillations such
as a period, an amplitude, volatility. Varying the parameter ε, we can change the
form of the oscillations of the solution (7.7) from a harmonic to a telegraphic one.
Applying the generalized Euler method to (7.7), we obtain the discrete mathe-
matical spread model in the form

δi+1 = δi + hi ϕi
(7.8)
ϕi+1 = ϕi + (εi ϕi (1 pi δ2i ) + qi δi )hi + σi∆wi

where hi is a grid time step, δi , ϕi , εi , pi , qi , σi are grid functions, ∆wi is the incre-
ment of the Wiener process.

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 381

We can construct a ‘double spread’, puchasing and selling two spreads simulta-
neously, for example, (GBP, SFR) and (GBP, EC). The mathematical double spread
model is the SDE system of the form
dy1 = y2 dt
dy2 = (ε1 y2 (1 p1 y21 ) + q1 y1 + q2 y3 )dt + σ1 (t ; y1 )dw1 (t )
(7.9)
dy3 = y4 dt
dy4 = (ε2 y4 (1 p2 y23 ) + q3 y3 + q4 y1 )dt + σ2 (t ; y3 )dw2 (t )
where w1 (
), w2 (
) are independent standard Wiener processes, y1 = δt1 , y3 = δt2 .
The extent to which the spreads are correlated among themselves is specified by the
values of the parameters q2 , q4 .

8. SPREAD DYNAMICS MODEL WITH HARMONIC OSCILLATIONS

As a model of spread dynamics, which is rather simplified but convenient for prac-
tical application, we can use the model of the form
 2 t  
δt = exp θ1 + θ2t + θ3t 2 + ∑ r j sin 2π + ψj + σξ(t ) (8.1)
j=1 τj

where ξ(t ) is a stationary random process with zero mathematical expectation and
unit variance, θ1 , θ2 , θ3 , r1 , r2 , τ1 , τ2 , ψ1 , ψ2 , σ are constant coefficients. The
random process (8.1) can readily be represented as the SDE solution similar to (7.5).
We assume for definiteness that r1 > 0, r2 > 0, oscillation periods τ1 > τ2 > 0,
the initial oscillation phases ψ1 , ψ2 are in the interval [0; 1℄. The model (8.1) can be
represented in the form convenient for calculation:
ln δt = f T (t ; τ1 ; τ2 )Θ + σξ(t ) (8.2)

where Θ = (θ1 ; θ2 ; : : : ; θ7 )T ,
θ2+2 j = r j cos 2πψ j ; θ3+2 j = r j sin 2πψ j ; j = 1; 2

 t t t t T
f (t ; τ1 ; τ2 ) = 1; t ; t 2 ; sin 2π ; cos 2π ; sin 2π ; cos 2π :
τ1 τ1 τ2 τ2
Suppose there are observations of the spread dynamics δi = δti at instants ti , i =
1; : : : ; N p . Then using the method of least squares, the estimates of the unknown
parameters Θ, b b
τ1 , b
τ2 are derived from the condition for the function to attain a min-
imum:
Np
Q(Θ; τ1 ; τ2 ) = ∑ (ln δi f T (ti ; τ1 ; τ2 )Θ)2 : (8.3)
i=1

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
382 S. S. Artemiev and M. A. Yakunin

δ; $
1300
1200
1100
1000
900
800
700
600
500

0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 t

Figure 2. Spread dynamics δt ( ) and the estimates of its mathematical expectation m̂δ(t )( ).

Solving the equation ∂Q=∂Θ = 0, we get

 Np  1 Np
Θ= ∑ f (ti ; τ1 ; τ2 ) f T (ti ; τ1 ; τ2 ) ∑ f (ti ; τ1 ; τ2 ) ln δi : (8.4)
i=1 i=1

Given τ1 , τ2 , the estimate Θ b is immediately calculated by this formula. If τ1 , τ2

are not known, then replacing in (8.3) the vector of the variables, Θ, by the vec-
tor function Θ̃(τ1 ; τ2 ) according to (8.4), we obtain the function Q(Θ̃(τ1 ; τ2 ); τ1 ; τ2 )
rather than (8.3). It is sufficient to minimize this function by two variables only,
for example, by simple sorting on a two-dimensional grid. When using minimiza-
tion algorithms we should take into account that even in the domain with restriction
τ1 > τ2 > 0 the function Q has the set of minima among which an absolute mini-
mum must be chosen. The estimate of the root-mean-square deviation σ b is derived
by the formula

q
σ
b= b ;b
Q (Θ τ1 ; b
τ2 )=N p :

r j and the initial phases ψ

The estimates of the amplitudes b b j of oscillations, j = 1; 2,
are derived by the formulae

q
b
θ2 b2
b
rj = 2+2 j + θ3+2 j

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
 Mathematical models of the prices 383

8
> b
θ2+2 j
>
>
<
1
2π
arccos
b
r
; b
θ3+2 j 0 
ψ
bj = j
b
>
> 1 θ2+2 j b
>
: 1 arccos ; θ3+2 j < 0:
2π b
rj

If all the random values ξ(ti ) at observation instants ti are normal and independent,
all the above estimates of the least squares of unknown parameters are simultane-
ously their maximum likelihood estimates.
Figure 2 shows the graph of currency spread dynamics (GBP, SFR) calculated
on the basis of changes in the exchange rates of these currencies during 1999 and
the graph of the estimation of the mathematical expectation of spread

τ1 ; b
b δ (t ) = exp( f T (t ; b
m b +σ
τ2 )Θ b2 =2):

The number of the observations of the spread values is N p = 4703. The observation
interval equal to 5 minutes is used as a unit of measurement of time in the graph.
An absolute minimum of the function Q is attained for the parameter values used
in the graph construction and given in Table 2. The calculation of the root-mean-
square deviations of the estimates for the components of the vector Θ shows that
all the values b
θ j , j = 1; : : : ; 7, significantly differ from zero. As seen from Fig. 2,
the model development as a whole corresponds to that of the real currency spread.
In addition, if only one sine curve is available in the model (8.1), we estimate the
component of spread fluctuations only with the large period: b τ1 = 2690.

Table 2.

Parameter Estimate

θ1 7.13084
θ2 4:6058  10 4
θ3 8:1132  10 8
r1 0.15819
r2 0.09205
ψ1 0.41064
ψ2 0.67184
τ1 2927
τ2 828
σ 0.10273

Note that in attempting to analogously estimate the constant parameters ε, p,

q, σ in the model (7.8) the estimates p and q turn out to insignificantly differ from
zero and the fluctuations are identified as random ones and make spread variance
increase due to the increasing estimate of the input noise intensity σ in the SDE.
Because of this problem the wave nonlinear models call for further investigation and
the development of new numerical algorithms for their parameter identification.

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM
384 S. S. Artemiev and M. A. Yakunin

9. CONCLUSION
The construction of the mathematical price model of any financial instrument is pre-
ceded by the thorough statistical analysis of its historical price and by revealing the
specific features of the price behaviour. Using the statistical analysis, we choose the
appropriate structure of the mathematical model and develop the algorithm for esti-
mating its parameters. The mathematical model for different parameter values must
reproduce different price developments. This can be used, for example, for checking
the operation of a trading computer program in various markets. These mathemat-
ical models can also be used for the construction of the global model of the World
financial crisis development. The wave character of the model prices corresponds to
the real trading observations of various financial instruments in many trading sites
in the World. The stochastic wave price models have a large set of parameters. In
view of this, they are more flexible and adaptable to the reproduction of any hypo-
thetical market developments than the other models. These price models allow one
to attack in a new way the problem of constructing the investment stock portfolio. In
combination with the Monte Carlo method they give the unique possibility to con-
struct a variety of statistical algorithms for the portfolio construction and, what is
most important, to optimally control the portfolio structure depending on the current
stock market situations.

REFERENCES
1. S. S. Artemiev, I. G. Mikhailichenko, and I. N. Sinitsyn, Statistical modelling of fixed-date fi-
nancial operations. Comp. Center, Siberian Branch of RAN, Novosibirsk, 1996 (in Russian).
2. S. S. Artemiev and M. A. Yakunin, Estimation of the parameters in stochastic differential equa-
tions. Russ. J. Numer. Anal. Math. Modelling (1997) 12, No. 1, 1–12.
3. S. S. Artemiev and M. A. Yakunin, Estimation of parameters in the system of stochastic differ-
ential equations by discrete observations. Sib. Mat. Zh. (1999) 40, No. 5, 987–993 (in Russian).
4. S. S. Artemiev and M. A. Yakunin, Estimation of parameters linearly dependent on the system
of stochastic differential equations. Sib. Zh. Vych. Mat. (1999) 2, No. 1, 1–11 (in Russian).
5. S. S. Artemiev and M. A. Yakunin, Multidimensional model of share price development and the
problem of constructing the investment portfolio. Sib. Zh. Vych. Mat. (2001) 4, No. 1, 13–20 (in
Russian).
6. F. Black and M. Scholes, The pricing of options and corporate liabilities. J. Political Economy
(1973), No. 3, 637–659.
7. A. N. Burenin, Futures, Forward, and Option Markets. Trivola, Moscow, 1994 (in Russian).
8. J. Hull and A. White, The pricing of options as assets with stochastic volatilities. J. Finance
(1987), No. 42, 281–300.
9. Kh. D. Ikramov, Numerical Solution of Matrix Equations. Nauka, Moscow, 1984 (in Russian).
10. U. F. Sharp, G. J. Alexander, and D. V. Baily, Investments. INFRA-M, Moscow, 1997 (in Rus-
sian).
11. V. I. Tikhonov and M. A. Mironov, Markov Processes. Sov. Radio, Moscow, 1977 (in Russian).
12. M. A. Yakunin, Estimation of parameters in the multidimensional Ornstein-Uhlenbeck process.
Thery and applications of statistical modelling. Comp. Center, Siberian Branch of USSR Acad.
Sci., Novosibirsk, 1988, pp. 124–134 (in Russian).

Brought to you by | New York University Bobst Library Technical Services

Authenticated
Download Date | 6/5/15 6:45 PM