See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/51957535

Markov Chains application to the financial-economic time series prediction

Article · November 2011

Source: arXiv

CITATIONS READS
13 6,039

3 authors, including:

Vladimir Soloviev Dmitry Chabanenko

Kryvyi Rih State Pedagogical University, Kryvyi Rih, Ukraine Cherkasy State University, Bogdan Khmelnitskiy
100 PUBLICATIONS 859 CITATIONS 3 PUBLICATIONS 18 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Dmitry Chabanenko on 22 May 2015.

The user has requested enhancement of the downloaded file.

 Markov Chains application to the financial-economic time series
prediction

Vladimir Soloviev∗
Cherkasy National University named after B. Khmelnitsky, Cherkassy, Ukraine

Vladimir Saptsin†
arXiv:1111.5254v1 [q-fin.ST] 22 Nov 2011

Kremenchug National University named after M. Ostrogradskii, Kremenchuk, Ukraine

Dmitry Chabanenko‡
Cherkasy National University named after B. Khmelnitsky, Ukraine

In this research the technology of complex Markov chains is applied to predict

financial time series. The main distinction of complex or high-order Markov Chains
and simple first-order ones is the existing of aftereffect or memory. The technology
proposes prediction with the hierarchy of time discretization intervals and splicing
procedure for the prediction results at the different frequency levels to the single
prediction output time series. The hierarchy of time discretizations gives a possibility
to use fractal properties of the given time series to make prediction on the different
frequencies of the series. The prediction results for world’s stock market indices is
presented.

Contents

1. Introduction 2

2. Analysis of prominent publications relevant to the subject 3

3. Aims of the paper, problem statement 3

∗
Electronic address: vnsoloviev@rambler.ru
†
Electronic address: saptsin@sat.poltava.ua
‡
Electronic address: chdn6026@mail.ru
 2

4. Classical modeling problems of ESE systems dynamics 4

5. Modern concepts in ESE systems modeling 7

6. Markov chains prediction technology 8

7. Prediction construction algorithm 10

8. States in complex Markov chains and approaches for defining them 11

9. Step-by-step prediction procedure. Defining the most probable state on

the next step, prediction scenarios 13

10. Time increments hierarchy and splicing procedure 15

11. Results of stock indices prediction 17

12. Conclusions and further work 22

References 22

1. INTRODUCTION

Successful modeling and prediction of processes peculiar to complex systems, such as

ecological, social, and economical (ESE) ones, remain one of the most relevant problems as
applied to the whole complex of natural, human and social sciences ([1–6]).
The diversity of possible approaches to modeling such systems and, usually, more than
modest success in the dynamics prediction, compel us to look for the reasons of failure, find-
ing them not only in details, but also in the axiomatics, which relates to problem statement,
chosen modeling methods, results interpretation, connections with other scientific directions.
With the appearance of quantum mechanics and relativity theory in early twentieth
century new philosophical ideas on physical values, measuring procedures and system state
have been established, the ones that are completely different from Newtonian notions [7, 8].
For more than 70 years basic concepts of classical and neoclassical economic theories have
been discussed by leading scientists, generating new approaches [9]. The general systems
theory has acquired recognition in the middle of the 20th century giving way to development
 3

of the new, systemic, emergent, and quantum in essence approach to investigation of complex
objects, which postulates the limited nature of any kind of modeling and is based upon fixed
and closed system of axioms [10].
However, the development of this new philosophical basis of ESE systems modeling is
still accompanied with numerous difficulties, and new principles are often merely declared.
Current research is devoted to investigation and application of the new modeling and
prediction technology, suggested in [11, 12], based on concepts of determined chaos, complex
Markov chains and hierarchic (in terms of time scale) organization of calculating procedures.

2. ANALYSIS OF PROMINENT PUBLICATIONS RELEVANT TO THE

SUBJECT

Prediction of financial-economic time series is an extremely urgent task. Modern ap-

proaches to the problem can be characterized by the following directions: 1) approximation
of a time series using an analytical function and extrapolation of the derived function to-
wards future – so-called trend models [13]; 2) investigation of the possible influence various
factors might have on the index, which is being predicted, as well as development of econo-
metric or more complicated models using the Group Method of Data Handling (GMDH)
[3, 14]; 3) modeling future prices as the decisions-making results using neuronal networks,
genetic algorithms, fuzzy sets [14–16].
Unfortunately, these techniques don’t produce stable forecasts, what can be explained by
complexity of the investigated systems, constant changes in their structure. Although we
are trying to join these directions in one algorithm, it is the latter option that we prefer,
with it consisting in creating a model adequate to the process generating a price time series
[17]. This very approach gives a chance to approach the complexity of the system, which
generates the observed series, develop the model and use its properties as the prognosis.

3. AIMS OF THE PAPER, PROBLEM STATEMENT

Assume the time series is set by a sequence of discrete levels with constant step of time
sampling ∆t. We need to generate variants of the time series continuation (prognosis sce-
narios) according to the relations between the sequences of absolute and relative changes
 4

discovered with the help of complex Markov chains.

4. CLASSICAL MODELING PROBLEMS OF ESE SYSTEMS DYNAMICS

Another peculiar feature of ESE systems, apart from complexity, is a memory, including
the long-term one, as well as nonlinear and unstable nature of interactions and components,
which makes it harder to predict their future behavior.
Unfortunately, mathematical models based on differential equations have no memory
(there is no aftereffect), while for models with memory, where integral interrelations are
used, it is not always possible to take into account nonlinearity (the integration procedure
is linear by definition).
In reality, in the Cauchy problem future systems behavior is defined by its initial state
and doesn’t depend on the way the system reached its current state. However, it is hardly
true that future behavior of a real socio-economic or socio-ecological system can be predicted
by giving an immediate time “slice” of a variables set that describe its state.
Let us consider possible ways to take into account past events while modeling ESE sys-
tems’ dynamics, which goes beyond the boundaries of classical differential and integral equa-
tions.
Functional differential lagging equation can serve as a simple example of the dynamic
model with memory, where present time is defined by the state variable x(t) and depends
on the past state x(t − τ ) with constant time lag τ = const:

x(t) = f (x(t − τ )) ; t ≥ t0 , (1)

where f (x) is the known function, with initial conditions being set for the half-interval
t0 − τ ≤ t < t0 by the function φ(t):

x(t) = φ(t); t0 − τ ≤ t < t0 . (2)

Given the 2 equation 1 has the only solution, defined by recurrent ratios:

f (φ(t − τ )) ; if t0 ≤ t < t + τ ;







f (f (φ(t − τ ))) ; if t0 + τ ≤ t < t + 2τ ;

x(t) = (3)
f ((f (φ(t − τ )))) ; if t0 + 2τ ≤ t < t + 3τ ;







· · ·

 5

Using Dirac delta function, as defined by ratios:

Z+∞
δ(t) = 0, if x 6= 0; δ(t)dt = 1, (4)
−∞

we can formally rewrite equation 1 in the integral form:

Zt
x(t) = dt1 f (x(t1 )) H(t1 , t); H(t1 , t) ≡ δ (t1 − (t − τ )) ; t ≥ t0 . (5)
−∞

Delta function is not a function in the conventional interpretation and is related to the
class of generalized functions that were mathematically described only in the middle of
the last century [18] (physics started using this function much earlier). Its classical form is
considered to be a limit of the “peak” sequence, with its centre set in the point of origin. The
afore-mentioned “peaks” indefinitely converge widthway, indefinitely increase throughout the
height and have a unit area.
An approximate classic integral analogue of the equation 5 can be derived by substituting
δ(t) with an ordinary function - some specific narrow enough “peak” of a unit area, a certain
finite width v ∆t as well as a finite height . The derivative of the Fermi function is one of
the possible examples:

1 1
Φ(t) = −t

; δ(t) ≈ −t

t


. (6)
1 + exp θ
θ 2 + exp θ
+ exp θ

If the system’s state in the moment t, x(t), is defined not by one, as in 1, but k (k =
2, 3, 4, · · · ) of her past states x(t − τ1 ), x(t − τ2 ), · · · x(t − τk ) in the following moments of
time (t − τ1 ), (t − τ2 ), · · · , (t − τk ) respectively (τ1 = const, τ2 = const, · · · , τk = const, τ1 >
τ2 > · · · > τk > 0), then instead of (1), (2), (5) we get:

x(t) = f (x(t − τ1 ); x(t − τ2 ); · · · ; x(t − τk )) ; t ≥ t0 ; (7)

x(t) = φ(t); t0 − τ1 ≤ t < t0 ; (8)

Rt Rt Rt
x(t) = dt1 dt2 · · · dtk f (x(t1 ), x(t2 ), · · · , x(tk )).
−∞ −∞ −∞ (9)
δ ((t1 − (t − τ1 )) δ ((t2 − (t − τ2 )) · · · δ ((tk − (t − τk )) ; t ≥ t0 .
 6

Therefore if the system’s state in the time moment t depends on the infinite sequence
of its past states, the integral analogue of the functional differential lagging equation will,
generally speaking, contain an integral of the infinite multiplicity. At the same time the
infinite amount of past states can relate to both finite (t − τ1 ; t) (short-term memory) and
infinite (−∞; t) (long-term memory) time span.
Pay attention that the classic integral lagging equation is one of the Volterra type [19]:
Zt


x(t) = F x(t̃); t; t̃ dt̃, (10)
−∞


where F x(t̃); t; t̃ - is an arbitrary (generally nonlinear) function of variables x(t̃); t; t̃,
which allows to take into account system’s memory of its past states only in the additive
approximation, which becomes evident, if the right section 10 is rewritten in the following
way:

Rt
Rt
Rt1

F x(t̃); t; t̃ dt̃ ≡ F x(t̃); t; t̃ dt̃ + F x(t̃); t; t̃ dt̃ + · · · =
−∞ t1 t2
(11)
F x(t˜1 ); t; t˜1 · (t − t1 ) + F x(t˜2 ); t; t˜2 · (t1 − t2 ) + · · · ;

t > t1 > t2 > · · · ; t˜1 ∈ [t1 , t] ; t˜2 ∈ [t2 , t1 ] ; · · ·

In connection with it note that the equation 9 in case of an additive dependency of
contemporaneity on the past, i.e. in case:

f (x(t − τ1 ); x(t − τ2 ); · · · ) ≡ f1 (x(t − τ1 )) ; f2 (x(t − τ2 ))) ; · · · (12)

becomes a particular case of the equation 10 with the following integrand:

F x(t̃); t; t̃ ≡ f1 x(t̃) δ t̃ − (t − τ1 ) + f2 x(t̃) δ t̃ − (t − τ2 ) + · · · (13)

Meaningful analysis of nonlinear models dynamics with memory, in which the future is
defined by the infinite amount of states in the past is generally possible only in case of a
discrete representation. The results of such analysis will be approximated, i.e. will contain
uncertainty, which has to be considered endogenous, i.e. internal, and peculiar to this very
system.
With a certain level of time sampling, models with memory both 7 and 10 becomes:

x(n + 1) = f (x(n); x(n − 1); x(n − 2) · · · ) . (14)

 7

To take into account and quantify the uncertainties, observed in ESE as well as other
complex systems probability models are normally used. However their application is based
on doubtful hypotheses, while the statistical interpretation of the results is not always infor-
mative enough and results might not correspond with the real process occurring within the
system. In particular, the well-known problem of 1/f –noise (look for example [20]), closely
connected to the presence of long-term memory in complex systems, implies the absence of
the mean temporary value (as a limit of a certain time span converging to infinity, which
serves as the basis for averaging) for any process occurring in such kind of system. Therefore
such processes can’t have a rigorous statistical substantiation.

5. MODERN CONCEPTS IN ESE SYSTEMS MODELING

New approaches to modeling and prediction of complex nonlinear systems dynamics with
memory are based on the use of determined chaos and neural networks technologies (cf.
e.g. [15, 21, 22]). Both investigation and realization of such techniques has become possible
only with the appearance of quick-operating computers. Use of the recurrent computational
process has become the general feature for all these technologies:

xn+1 = fn (fn−1 (· · · (f1 (x1 ) · · · ))) , n = 1, 2, · · · , (15)

where fi (xi ) is a certain nonlinear mapping of a multi-dimensional vector xi , i - discrete,

real or fictitious, time. Identification of the model 15 is reduced to the determination of
functions fi (xi ), while the differences between the models of determined chaos and neural
networks are connected with the function type and methods of its definition (neural net-
work models normally use rather narrow class of fi (xi ) mappings [16]). Generally speaking,
stability or convergence of the process 15 is not required, whereas a single-step set of vector
xi components as well as their time dynamics can be of great interest.
For the particular case of the model 15, introducing corresponding lagged variables, a
model 14 can be transformed.
Both determined stable processes, described by integro-differential equations, and random
processes, which also include complex Markov chains (CMC), can be formally considered
as separate extreme cases of determined chaos models realization 15. Given the sampling
scale, which tends to zero, if such a tendency makes sense and corresponding limits exist, we
 8

derive classical differential and integral problem statement. Finite ∆t allows to get models
with discrete time, which in the general case in the corresponding phase space (which also
includes lagged variables) can produce both measurable sets (discrete or continuous) that
allow probabilistic interpretation and those of the special structure – fractals [23], that can’t
be always interpreted in that way.
Various digital generators of so-called random sequences used in imitational modeling
can be an example of determined chaos models that allow probabilistic interpretation.
Let us note that in reality there are no accurate procedures that would give an opportunity
to distinguish a “real” random sequence from the pseudorandom one.

6. MARKOV CHAINS PREDICTION TECHNOLOGY

Suppose there is a sequence of a certain system discrete states. From this sequence we
can determine transitions probabilities between the two states. Simple Markov chain is a
random process, in which the next state probability depends solely on the previous state
and is independent from the rest of them. Complex Markov chain, unlike the simple one,
stands for the random process, in which the next state probability depends not only on the
current, but also on the sequence of several previous states (history). The amount of states
in history is the order of the Markov chain.
Theory of simple Markov chains is widely presented in literature, for example [24]. As
for the high order Markov chains, modern literature [25] can offer us a mere definition.
Developing complex or high order Markov chain’s properties is not widely presented in
modern scientific publications. It’s necessary to mention the papers [26, 27] where properties
of complex Markov Chains are developed, but no prediction algorithm is proposed there.
The development of prediction method, based on complex Markov chains, is proposed in
this paper.
Markov chain of the higher order can be brought to a simple Markov chain by introducing
the notion of a “generalized state” and including a series of consequent system’s states into
it. In this case, tools of simple Markov chains can be applied to the complex ones.
Investigated dynamic series is a result of a certain process. It is assumed that this process
is determined, which implies the existence of a causal dependence of further states on history.
It is impossible to fix and analyze the infinite history, which puts obstacles in the way of an
 9

accurate detection of this influence and making precise predictions.

The problem consists in the maximal use of information, which is contained in the known
segment of the time series, and subsequent modeling of the most probable future dynamics
scenario.
The observed process is described as a time series of prices pt with the given sampling
time span ∆t
pti = p(t0 + i∆t). (16)

Discrete presentation of the time series is in fact a way of existence of this very system.
New prices are formed on the basis of contracts or deals, made on the market in certain
discrete moments of time, while the price time series is a series of the averaged price levels
during the chosen time intervals. While making a decision each trader, who is an active
part of the pricing system, works solely with discrete series of the chosen time interval (e.g.
minute, 5-minute, hourly, daily etc.). For ∆t → 0 the accuracy of data presentations reaches
a certain limit, since for relatively small ∆t the price leaps in the moment of deal, while
staying unchanged and equal to the last deal during the time between the two deals. Hence,
the discreteness of time series has to be understood not only as a limited presentation of
activity of the complex financial system, but also as one of the principles of its operation
[11, 12, 28–30].
The time series of initial conditions has to be turned into a sequence of discrete states.
Let us denote the amount of chosen states as s, each of them being connected to the change
in the quantity of the initial signal (returns). For example, consider the classification with
two states, first of which corresponds to positive returns as the price increases, while the
second one – to negative as it descends. Generally all possible increments of the initial time
series are divided into s groups. Ways of division will be discussed further.
Next we develop predictions for the time series of sampled states. For the given order
of the Markov chain and the last generalized state the most probable state is chosen to
be the next one. In case if ambiguity occurs while the state of maximum probability is
being evaluated, an algorithm is used that allows reducing the amount of possible prediction
scenarios. Therefore we get the series of predicted states that can be turned into a sampled
sequence of prognostic values.
Evaluation of increments, prediction, and subsequent restoration are conducted for the
given hierarchy of time increments t. To use the given information as effectively as possible,
 10

the prediction is conducted for time increments t = 1, 2, 4, 8, ..., or a more complex hierarchy
of increments and subsequent “splicing” of the results derived from different prediction
samplings.
The procedure of prediction and splicing is iterative and conducted starting from smaller
increments, adding a prediction with the bigger time increment on every step.
As the sampling time step t increases, the statistics for the investigation of Markov chains
decreases, whereas the biggest sampling step, which takes part in the prognostication, limits
itself. To supplement the prediction with the low-frequency component the approximation
of zero order is being used in the form of a linear trend or a combination of a linear trend
and harmonic oscillations [31, 32].

7. PREDICTION CONSTRUCTION ALGORITHM

Let us consider the consequence of operations, required for the prognostic time series
construction. To do this we need to set the following parameters:
1) The type of time increments hierarchy (simple – powers of two, complex – product of
powers of the first simple numbers).
2) Values of s – the amount of states and r – the order of the Markov chain. These
parameters can be individual for every sampling level; finding of optimal parameters is done
experimentally.
3) Threshold values δ, and minimal number of transitions Nmin .
Prediction construction algorithm includes the following steps:
1) Generating hierarchy of time increments - t sequence. The maximal of them has to
correspond to the length of a prognostic interval Nmax .
2) For every time increment ∆t, as the increments increase, a prediction of states and
restoration of the time series along the prognostic states is conducted. Current stage includes
following actions:
2.1. Evaluating increments (returns) of the series with ∆t sampling.
2.2. Transforming the time series of increments into the series of state numbers (1..s).
2.3. Calculating transition probabilities for generalized states.
2.4. Constructing the series of prognostic states using the procedure of defining the most
probable next state.
 11

2.5. Restoring the value series from the state series with ∆t sampling.
2.6. Splicing the prediction of ∆t sampling with the time series derived from splicing of
the previous layers (with the lesser step ∆t). In case if the current time series is the first
one, the unchanged time series will come as a result of splicing.
3) To splice the last spliced time series with the continuation of the linear trend, created
along all previously known points.
The time series, spliced with the linear trend, is the result of prediction. Let us consider
the stages of the given algorithm in detail.

8. STATES IN COMPLEX MARKOV CHAINS AND APPROACHES FOR

DEFINING THEM

In everything that concerns current technology, states are connected to the measuring of
a prognostic value. There is a number of ways to classify returns in states, from which the
following are suggested. One of them is the classification based on the homogeneity principle
as concerning the amount of representatives in classes; based on the homogeneity principle
of deviation, as well their combinations for different deviation modules.
Increment or returns of the time series serves as the basis for states classification [32, 33].
Absolute ra and relative rt increments of the time series are considered:

ra = pt − pt−∆t , (17)

pt − pt−∆t
rt = , (18)
pt
where pt – is the input time series of price dynamics, ∆t – sampling interval, which is
chosen for subsequent analysis. It is known that mathematical expectation of the returns
time series equals zero, whereas variation comes as the measure of time series volatility.
Based on returns values rt classification and transformation of values to the time series of
discrete states are conducted. One of the classification principles is homogeneity according to
the amount of class representatives. This classification divides the set of all increments into s
groups equal in number. Calculated with the given sampling, time series increments are then
systematized in growth and divided into equal parts. Thus we define limit values {rlim,i },
which are used afterwards during transformation of the returns into class numbers. Large
 12

number of identical states can cause certain problems, such as identical bounds of several
neighbouring states. It creates a number of states with no representatives, which makes
correction of the division a necessary action. In that way we will reach the largest possible
homogeneity in state division. Classification is conducted along the following algorithm
[34, 35]:

st = (i | rlim,i−1 > rt > rlim,i ) (19)

where st is the number of state, which corresponds to the moment of time t, for which the
returns level was computed rt ; i is the number of state [1 . . . s], which is characterized by
the interval [rlim,i−1 , rlim,i ] corresponding to the calculated returns level rt .
Apart from the returns interval, given by the aforementioned values [rlim,i−1 , rlim,i ], a
mean returns value is chosen for every state ravg,i , which will be used in time series values
transformation according to the prognostic discrete states.
Another way of dividing the time series into states implies dividing the interval of returns
values into equal parts, from minimal to maximal deviation. In this case homogeneity
according to the amount of representatives in states does not occur. In fact this method
differs from the previous one in terms of defining limit values {rlim,i }. Possible combined
ways of division, in case of which the limit value, dependent on standard deviation, is used
instead of maximal and minimal value, and division is conducted homogenously according
to the deviation.
Since the real causal dependence is unknown during the process, to find it adequate
state classification, which would allow to reveal vital dependencies of the time series, is
required. We suggest a couple of ways to divide the time series into states, which in the first
place allow to preserve adequate transition probabilities between states, as well as prevent
averaged deviations inside the states from affecting the accuracy of the derived prediction.
To check the efficiency of division we conduct the sampling procedure and classify the
increments according to each hierarchy. Having completed that, we restore the time series
using known states for each hierarchy and finish the splicing procedure. Since the state series
correspond to the initial time series, we get the curve, with deviation, caused exclusively
by the state averaging mistake (quantum mistake). Thus, having set a certain value of
state numbers s and carried out sampling, restoring, and sampling procedures (excluding
prediction), we get absolute sampling (quantum) mistake.
 13

Increasing the number of states, we improve the accuracy of restoration, however one
should remember, that the choice of the quantum levels is limited by the fact that the
transition probabilities definition with sufficient accuracy is required, which is confirmed by
artificial test time series prediction experiments.

9. STEP-BY-STEP PREDICTION PROCEDURE. DEFINING THE MOST

PROBABLE STATE ON THE NEXT STEP, PREDICTION SCENARIOS

Predicting procedure uses the most probable state as the next one under current circum-
stances. Probability matrix of state transitions is used for the afore-mentioned purpose.
In this case, you have to take into account that probabilities are calculated with a certain
mistake. We cannot precisely compute the probabilities, since it is impossible to derive an
infinite time series, and only a part of the time series is known – the known part serves as
the basis for probabilities. The second important aspect implies the case of several states
with maximal probability.
To prevent the omission of the states, for which the probabilities are computed with a
mistake, one should add a state with maximal probability to the states, which are located in
the distance of δ from the maximal one. The value of parameter δ depends on the probability
evaluation mistake and requires experimental refinement.
If δ > 0, the number of states with maximal probability increases in comparison to the
value δ = 0. Let us call a couple of neighbouring states with maximal probability a cluster.
Cluster states with average deviation values are supposed to have the largest probability.
To predict the dynamics, let us confine ourselves to one or two most probable states. To
define them a following algorithm is suggested:
1) If levels (discretized increments) create several clusters (cluster is a group of several
neighbouring levels – cluster elements, minimal cluster is a single isolated level) with maximal
probability, we choose the largest cluster.
2) If the number of cluster elements is odd, as kmax we choose a central element.
3) If the number of cluster elements is even, we consider two central cluster elements and
choose as kmax the one, which is closer to the centre of distribution.
4) If two central cluster elements are equidistant to the centre of distribution, we consider
both cases as possible variants of kmax values (bifurcation point).
 14

5) If there are several clusters of maximal size, we consider them as new elements, which
can also form clusters that will undergo the same steps 1)-4).
This principle is based on the following ideas:
1) If there are two neighbouring states of maximal probability, it is better to take the
one, which is closer to the centre of distribution, in order to minimize the risk of occurrence
of false linear trends in the prediction.
2) If levels of maximal probability are not the neighbouring ones, at least two variants
have to be considered, as it can be connected to the bifurcations that should not be omitted.
3) If the prediction is carried out according to 1) (on all stages of the hierarchy), we
receive a certain approximation of the lower limit of the prediction, whereas in case of 2) –
we get an approximation of the upper limit.
Hence this algorithm can adequately restore the case of possible bimodal probability
distribution, it is proposed to consider 2 prediction scenarios.
In case of the complex Markov chains, probability of the next state depends not only on
the previous state, but also on the sequence of r states, which have occurred before given.
In this case, it is necessary to calculate transition probabilities from the sequence of r states
into the r + 1 state. Formally, these probabilities can be written into the rectangular table
of (rs , s) size.
Having generalized the notion of “present state” and included a sequence of r preceding
states into it, we can reduce Markov chains of r order to the chain of the first order. Thus
transition probabilities can be written into rectangular matrices of (rs , rs ), that come as
transition probability matrices for generalized states.
The process of prediction implies the following: the last state is chosen (in case of Markov
chains of an order r > 1 a sequence of r latest states is taken). The probability of transition
from current state to all possible states is defined. From all possible states a state with
maximal probability is chosen. It is possible that several states with maximal probability
occur, which can be explained by the bimodal probability distribution. The process of
decision-making in this case is described later.
The chosen most probable state is taken as the next prognostic state and the procedure
is repeated for the next (last added) state. Thus we receive a time series of prognostic states
for the given sampling time ∆t.
Further according to the received state sequence and known initial value the time series is
 15

being restored for the given time sampling ∆t. In this case every state implies ∆t points of
the time series. On the stage of state classification every state was connected to the average
increment ravg,i , which is added to the value of the last point in the time series, and the
next discrete point is computed. Intermediate points are filled as linear interpolation of two
known neighbouring points. Algorithm of yt time series values restoration according to the
initial price pt and a series of average increments ravg,ik , corresponding to the prognostic
states sk , can be given by a sequence of calculations:

yt = p t ,
yt+1 = yt + ravg,i1 /∆t = pt + ravg,i1 /∆t,
yt+2 = yt+1 + ravg,i1 /∆t = pt + 2ravg,i1 /∆t,
...
yt+∆t−1 = yt+∆t−2 + ravg,i1 /∆t = pt + (∆t − 1)ravg,i /t,
(20)
yt+∆t = yt+∆t−1 + ravg,i1 /∆t = pt + ∆travg,i /∆t = pt + ravg,i1 ,
yt+∆t+1 = yt+∆t + ravg,i2 /∆t = pt + ravg,i1 + ravg,i2 /∆t,
...
Pn−1
yt+n∆t−1 = yt+n∆t−2 + ravg,in /∆t = pt + ravg,ik + (∆t−1)
k=1 ∆t
ravg,ik ,
Pn
yt+n∆t = yt+n∆t−1 + ravg,in /∆t = pt + k=1 ravg,ik .

10. TIME INCREMENTS HIERARCHY AND SPLICING PROCEDURE

Time series increments will be computed with different steps. For example, analogous
to the discrete Fourier transform, time increments are equal the powers of 2 are considered.
First, we calculate increments as a remainder of two nearest neighbouring time series values,
then next nearest values are considered with the step of 2, 4, 8, 16 etc. Let us mark this
difference in time as ∆t.
For every ∆t we conduct an increment time series transformation leading to a time series
of states. Further we predict the future sequence of states and restore the time series with
the given sampling rate according to the prognostic series of states.
Time series, received as a result of restoring for different ∆t, undergo the splicing proce-
dure, which gives out an actual prognostic time series.
Thus an increment hierarchy is chosen, where each one is responsible for its own sampling
rate, which serves as a basis for predicting, restoring and splicing.
 16

The splicing process implies the following. The procedure is iterative. With every next
(along with the increasing step) sampling time the series corrects itself, driving the predic-
tion, formed under lower ∆t, to its actual point. Transformations that are conducted during
splicing can be written down in the form of the following calculations.
Suppose the splicing procedure has been finished for all time increments ∆t < Deltati ,
the prediction has been done under the ∆ti sampling according to formulae 20, and as a
result a time series yi has been derived. Let us consider the iterative splicing procedure of
the received series yi with the series, acquired during all preceding splicing procedures gi .
Since the series yi contains system points only in moments aliquot to ∆ti , and other
points of the series are interpolated, the process of splicing implies the substitution of these
interpolated points with the values of system points from previous ∆t < ∆ti , which are
contained in the series of results of previous splicing procedures gi . Splicing algorithm can
be written in the sequence of computations:

zt = gt = pt ,
zt+1 = gt+1 + (yt+∆ti − gt+∆ti ) /∆ti ,
zt+2 = gt+2 + 2 (yt+∆ti − gt+∆ti ) /∆ti ,
...
zt+∆t−1 = gt+∆t−1 + (∆ti − 1) (yt+∆ti − gt+∆ti ) /∆ti ,
zt+∆t = gt+∆t + (∆ti ) (yt+∆ti − gt+∆ti ) /∆ti = yt+∆ti ,
(21)
zt+∆t+1 = gt+∆t+1 + ((yt+2∆ti − gt+2∆ti ) − (yt+∆ti − gt+∆ti )) /∆ti ,
zt+∆t+2 = gt+∆t+2 + 2 ((yt+2∆ti − gt+2∆ti ) − (yt+∆ti − gt+∆ti )) /∆ti ,
...
(∆t−1)


zt+n∆t−1 = gt+n∆t−1 + ∆t
(yt−n∆t − gt−n∆t ) − yt−(n−1)∆t − gt−(n−1)∆t ,



zt+n∆t = gt+n∆t + (yt−n∆t − gt+n∆t ) − yt+(n−1)∆t − gt+(n−1)∆t =
= gt+(n−1)∆t − yt+(n−1)∆t − yt−n∆t .
 17

11. RESULTS OF STOCK INDICES PREDICTION

In this section we offer the results of stock indices prediction. The stock’s indices
databases are available from [36]. Point 2000 indicates the starting moment of the prog-
nosis: March 24, 2011. The green line on the next figures indicates real indice’s or price’s
values. Our software for time series forecasting by the proposed methods is available from
our website: http://kafek.at.ua/MarkovChains1_2_20100505.rar.

Figure 1: Stock indices prediction. a) Dow Jones Industrial Average - DJI (USA). b)
FTSE 100 (Great Britain)

Figure 2: Financial companie’s share prices forecasting. a) Morgan Stanley (USA). b)

BNP Paribas (France)
 18

Prediction time series with different input learning set’s length are shown at the fig.3
and 4. Prediction series for DJI at the figure 3 are more correlated, than FTSE index at
the figure 4. At the subplot b) of the above mentioned plots the mean value and standard
deviations of the prediction’s series are presented. The time of prediction series beginning
on the next figures is the point 1000 and correspond to October 14, 2011.

Figure 3: Dow Jones Industrial Average - DJI (USA). a) Prediction series, calculated with
different learning set’s length. b) Mean value and standard deviation for prediction series.

Figure 4: FTSE 100 index prediction. a) Prediction series, calculated with different
learning set’s length. b) Mean value and standard deviation for prediction series.
 19

The normalization procedure is proposed in order to compare indices and it’s prediction
series with different absolute values. The normalized values calculated with the following
formula:
y(t) − min (y(t))
yn (t) = . (22)
max (y(t)) − min (y(t))
Normalized prediction time series are shown at the fig.5 (America), fig.6 (Europe, devel-
oped countries), fig.7 (Europe, PIIGS), fig.8 (Asian markets). All the figures contain mean
time series, which are weighted average of countrie’s stock indices predictions, weihted with
GDP values [37] for the corresponding countries.

Figure 5: Normalized mean values for the prediction series of America’s stock indices.
Brazil (BVSP), Mexico (MXX), Canada (GSPTSE), Argentina (MERV), USA (S&P 500)
 20

Figure 6: Normalized mean values for the prediction series of European

stock indices. Developed countries: FTSE (Great Britain), DAX
(Germany) FCHI (France), Netherlands (AEX)

Figure 7: Normalized mean values for the prediction series of European

stock indices. Portugal (PSI20), Italy (FTSEMIB), Ireland (ISEQ),
Greece (GD) and Spain (IBEX).
 21

Figure 8: Normalized mean values for the prediction series of Asian

stock indices. China (SSEC, HSI), Korea (KS11), Japan (NIKKEI),
India (BSESN), New Zealand (NZ50).

Figure 9: Mean values of normalized World’s powerful economies

indices prediction series.
 22

12. CONCLUSIONS AND FURTHER WORK

Current paper suggests an algorithm of time series prediction based on complex Markov
chains. Hierarchy of time increments principle allows to use the information, which is con-
tained in the time series during the prognosis construction, to its fullest. Experimental work
on stock market indices time series prediction shows the efficiency of the algorithm and
confirms the relevance of further research of the offered method.

[1] A. A. Samarskii and A. P. Mikhailov. Mathematical Modelling: Ideas. Methods. Examples.

Fizmatlit, Moscow, 2001.
[2] V. V. Bohoboyaschyy, K. R. Kurbanov, P.B. Paly, and V.M. Shmandiy. Principles of Fore-
casting in Ecology: Textbook (in Ukrainian). Center navchalnoji literaturi, Kyiv, 2004.
[3] O. G. Ivakhnenko. Grouping method of data handling - the concurrent of stochastic approxi-
mation methods (in ukrainian) Automatika, 3(3):58–72, 1968.
[4] J. V. Andersen, S. Gluzman, and D. Sornette. Fundamental framework for “technical analysis”
of market prices. European Physical Journal B, 14:579–601, March 2000.
[5] D. Horvath and R. Pincak. From the currency rate quotations onto strings and brane world
scenarios. arXiv:1104.4716 , April 2011.
[6] M. Repasan and R. Pincak. The string prediction models as application to financial forex
market. arXiv:1109.0435 , September 2011.
[7] P. V. Elyutin and V. D. Krivchenkov. Quantum Mechanics with tasks. Nauka, Moscow, 1976.
[8] L.D. Landau and E.M. Lifshitz. Quantum mechanics: non-relativistic theory. Teoretich-
eskafizika (Izd. 3-e) (Landau, L. D, 1908-1968). Butterworth-Heinemann, 1977.
[9] J. Sapir. K Ekonomitcheskoj teorii neodnorodnyh sistem - opyt issledovanija decentralizovannoj
ekonomiki. GU VShE, Moskov, 2001.
[10] L. von Bertalanffy. General system theory—a critical review. “General Systems”, VII:1–20,
1962.
[11] K. R. Kurbanov and V. M. Saptsin. Markov chains as technology for social, economic and eco-
logical processes forecasting. In “Problemy ta perspectivy rozvitku regonalno rinkovo ekonomki”
Conference proceedings (in Russian), pages 10 – 14, Kremenchuk, May, 11-13 2007.
 23

[12] V. M. Saptsin. Experience of using genetically complex Markov chains for the neural network
technology forecasting. Visnyk Krivorizkogo ekonomichnogo institutu KNEU, 2 (18):56 – 66,
2009.
[13] Y. P. Lukashin. Adaptive Methods of Time Series Forecasting: Textbook. Finance and Statis-
tics, Moscow, 2003.
[14] Y. P. Zaichenko. Fuzzy models and techniques in intelligent systems. Monograph(in Russian).
Slovo, Kiev, 2008.
[15] A. A. Ezhov and S. A. Shumsky. Neurocomputing and its application in economics and business
(Series “Textbooks” of Economic-Analytical Institute MEPI) / Ed. prof. V. V. Kharitonov.
MEPI, Moscow, 1998.
[16] I. V. Zayencev. Neural networks: basic models. Textbook for the course “Neural networks”
for 5-th grade students (in Russian) Physical Electronics Department, Faculty of Voronezh
State University Voronezh State University, Voronezh.
[17] D. M. Chabanenko. Detection of short- and long-term memory and time series predic-
tion methods of complex Markov chains. In Visnyk Natsionalnogo tehnichnogo universitetu
“Kharkivsky politehnichny institut”. Zbirnik Naukovyh pratz. Tematichny vypusk: Informatika
i modelyuvannya (in Ukrainian), number 31, pages 184 – 190. NTU KHPI, Kharkov, 2010.
http://www.pim.net.ua/ARCH F/V pim 10.pdf
[18] S. L. Sobolev. Selected topics from the theory of functional spaces and generalized functions.
Nauka, Moscow, 1989.
[19] J. Hale. Theory of Functional Differential Equations. Springer-Verlag, New York, Heidelberg,
Berlin, 1977.
[20] M.J. Buckingham. Noise in electronic devices and systems. Ellis Horwood series in electrical
and electronic engineering. E. Horwood, 1983.
[21] Hanz-Valter Lorenz. Nonlinear Dynamical Economics and Chaotic Motion. Springer-Verlag,
1989.
[22] E.E. Peters. Chaos and order in the capital markets: a new view of cycles, prices, and market
volatility. Wiley finance editions. Wiley, 1996.
[23] J. Feder. Fractals. Physics of solids and liquids. Plenum Press, 1988.
[24] V. I. Tikhonov and V. A. Mironov. Markov Processes. Soviet Radio, Moscow, 1977.
[25] G.A. Korn and T.M. Korn. Mathematical handbook for scientists and engineers: definitions,
 24

theorems, and formulas for reference and review. Dover books on mathematics. Dover Publi-
cations, 2000.
[26] Andrian E. Raftery. A model for high-order markov chains. Journal of the Royal Statistical
Society., 1985.
[27] Adrian Raftery and Simon Tavare. Estimation and modelling repeated patterns in high order
markov chains with the mixture transition distribution model. Appl. Statist., 43(1):179–199,
1994.
[28] V. M. Saptsin and V. N. Soloviev. Relativistic Quantum Econophysics. New paradigms of Com-
plex systems modeling: Monograph. http://kafek.at.ua/sol sap monogr.rar Brama-Ukraine,
Cherkassy, 2009.
[29] V. Saptsin and V. Soloviev. Relativistic quantum econophysics - new paradigms in complex
systems modelling. arXiv:0907.1142v1 [physics.soc-ph].
[30] V. D. Derbentsev, A. A. Serdyuk, V. N. Soloviev, and O. D. Sharapov. Synergeti-
cal and econophysical methods for the modeling of dynamic and structural characteris-
tics of economic systems. Monograph (in Ukrainian). Brama-Ukraine, Cherkassy, 2010.
http://kafek.at.ua/Monogr.pdf
[31] V. M. Saptsin and D. N. Chabanenko. Fourier-based forecasting of low-frequency components
of economical dynamic’s time series. In Problemy ekonomichnoyi kibernetiki: Tezy dopovidey
XIV Vseukrayinskoyi Naukovo-praktichnoyi konferentsiyi.(in Ukrainian), page 132, Kharkiv,
Oct 8-9, 2009 2009. KhNU imeni VN Karazina.
[32] D. M. Chabanenko. Discrete Fourier-based forecasting of time series. Sistemni tehnologii.
Regionalny mizhvuzivsky zbirnik naukovyh pratz (in Ukrainian), 1(66):114 – 121, 2010.
http://www.nbuv.gov.ua/portal/natural/syte/2010 1/15.pdf
[33] V. N. Soloviev. Mathematical economics. A textbook for self-study (in Ukrainian). CHNU,
Cherkassy, 2008. http://kafek.at.ua/Posibnyk Soloviev.rar
[34] Vladimir Soloviev, Vladimir Saptsin, and Dmitry Chabanenko. Prediction of fi-
nancial time series with the technology of high-order markov chains. In Working
Group on Physics of Socio-economic Systems (AGSOE), Drezden, 2009. http://www.dpg-
verhandlungen.de/2009/dresden/agsoe.pdf
[35] V. Soloviev, V. Saptsin, and D. Chabanenko. Financial time series prediction with the tech-
nology of complex markov chains. Computer Modelling and New Technologies, 14(3):63–67,
 25

2010. http://www.tsi.lv/RSR/vol14 3/14 3-7.pdf

[36] Yahoo! finance database. http://finance.yahoo.com
[37] Economic indicators : Gdp per capita (current prices, national cur-
rency) 2010. http://www.economywatch.com/economic-statistics/economic-
indicators/GDP Per Capita Current Prices National Currency/2010/

View publication stats