0% found this document useful (0 votes)
62 views21 pages

Forecasting Exchange Rates A Chaos Based Regression Approach

This document summarizes a research article that proposes a new model for forecasting exchange rates based on chaos theory and support vector regression. The model involves reconstructing the phase space from time series data and using a genetic algorithm to select optimal embedding parameters. The researchers applied this method to foreign exchange rate data and compared its forecasting accuracy to ordinary support vector regression. They found the proposed chaos-based method performs comparably to existing approaches for exchange rate forecasting.

Uploaded by

2k20csbs24
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
62 views21 pages

Forecasting Exchange Rates A Chaos Based Regression Approach

This document summarizes a research article that proposes a new model for forecasting exchange rates based on chaos theory and support vector regression. The model involves reconstructing the phase space from time series data and using a genetic algorithm to select optimal embedding parameters. The researchers applied this method to foreign exchange rate data and compared its forecasting accuracy to ordinary support vector regression. They found the proposed chaos-based method performs comparably to existing approaches for exchange rate forecasting.

Uploaded by

2k20csbs24
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 21

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

net/publication/276300553

Forecasting Exchange Rates:

Article in International Journal of Rough Sets and Data Analysis · January 2015
DOI: 10.4018/ijrsda.2015010103

CITATIONS READS

9 3,021

4 authors:

Ahmed Radhwan Mahmod Kamel


King Abdulaziz University King Abdulaziz University
3 PUBLICATIONS 17 CITATIONS 54 PUBLICATIONS 312 CITATIONS

SEE PROFILE SEE PROFILE

Mohamed Yehia Dahab Aboul Ella Hassanien


King Abdulaziz University Cairo University
80 PUBLICATIONS 525 CITATIONS 1,195 PUBLICATIONS 28,572 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ahmed Radhwan on 17 August 2017.

The user has requested enhancement of the downloaded file.


38 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

Forecasting Exchange Rates:


A Chaos-Based Regression Approach
Ahmed Radhwan, Department of Computing and Information Technology, King Abdulaziz
University, Jeddah, Saudi Arabia
Mahmoud Kamel, Department of Computing and Information Technology, King Abdulaziz
University, Jeddah, Saudi Arabia
Mohammed Y. Dahab, Department of Computing and Information Technology, King Abdulaziz
University, Jeddah, Saudi Arabia
Aboul Ella Hassanien, Cairo University, Cairo, Egypt, & Computers and Information Faculty,
Beni Suef University, Beni Suef, Egypt, & Scientific Research Group in Egypt (SRGE), Giza,
Egypt

ABSTRACT
Accurate forecasting for future events constitutes a fascinating challenge for theoretical and for applied
researches. Foreign Exchange market (FOREX) is selected in this research to represent an example of finan-
cial systems with a complex behavior. Forecasting a financial time series can be a very hard task due to the
inherent uncertainty nature of these systems. It seems very difficult to tell whether a series is stochastic or
deterministic chaotic or some combination of these states. More generally, the extent to which a non-linear
deterministic process retains its properties when corrupted by noise is also unclear. The noise can affect a
system in different ways even though the equations of the system remain deterministic. Since a single reliable
statistical test for chaoticity is not available, combining multiple tests is a crucial aspect, especially when
one is dealing with limited and noisy data sets like in economic and financial time series. In this research,
the authors propose an improved model for forecasting exchange rates based on chaos theory that involves
phase space reconstruction from the observed time series and the use of support vector regression (SVR) for
forecasting.Given the exchange rates of a currency pair as scalar observations, observed time series is first
analyzed to verify the existence of underlying nonlinear dynamics governing its evolution over time. Then, the
time series is embedded into a higher dimensional phase space using embedding parameters.In the selection
process to find the optimal embedding parameters,a novel method based on the Differential Evolution (DE)
geneticalgorithm(as a global optimization technique) was applied.The authors have compared forecasting
accuracy of the proposed model against the ordinary use of support vector regression. The experimental results
demonstrate that the proposed method, which is based on chaos theory and genetic algorithm,is comparable
with the existing approaches.

Keywords: Chaos theory, Differential Evolution, Foreign Exchange Rates, Forecasting, Support Vector
Regression, Forecasting, Time Series

DOI: 10.4018/ijrsda.2015010103

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 39

1. INTRODUCTION corrupted by noise is also unclear. The noise can


affect a system in different ways even though
Accurate forecasting of future events consti- the equations of the system remain determin-
tutes a fascinating challenge for theoretical istic. Since a single reliable statistical test for
researchers and practitioners. One of the largest chaoticity is not available, combining multiple
domains that share this problem is economy or tests is a crucial aspect, especially when one
more specifically financial markets. Predicting is dealing with limited and noisy data sets like
a financial series, as a stock market index or an in economic and financial time series. In this
exchange rate, remains however a very chal- research, a preliminary study was carried out
lenging task. The foreign exchange (FOREX) with the aim of detecting non-stationarity and
market (our case-study in this research) is one non-linearity in exchange rates time series. An
of the largest financial markets in the world. analysis of the underlying dynamical system
For more than four decades, Box and Jen- was also applied to study the existence of a
kins’Auto-Regressive Integrated Moving Aver- chaotic behavior. Finally, different techniques
age (ARIMA) model (Box & Jenkins, 1970) has of forecasting based on both theories of Chaos
been widely used for time series forecasting. and stochastic process were executed then a
However, some problems arise when forecasting comparison of the prediction accuracy was held
financial time series with ARIMA models, as between the two forecasting methods.
follows. First is the characteristic linear limita- The remainder of this paper is organized as
tion of ARIMA models, in contrast to real-word follows: Section 2 presents a review of existing
financial time series, which are often nonlinear [ approaches for research efforts that have been
(Lin et al., 2012), (Huang et al., 2010), (Ding et focused on the hypothesis that financial markets
al., 2009)& (Gradojevic & Yang, 2006) ] and are behavior is non-linear. The problem definition
rarely pure linear combinations. Second is the is described in section 3. The proposed method
robustness limitation of ARIMA models—the for a non-linear system and the applicability of
ARIMA model selection procedure depends chaos theoryare presented in section 4.Section
greatly on the competence and experience of 5 discusses how to apply the proposed method
the researchers to yield desired results. Unfor- on financial markets data especially such as
tunately, choice among competing models can FOREX data. Section6 provides the results and
be arbitrated by similar estimated correlation discussion of the proposed system. Finally, we
patterns and may frequently reach inappropriate summarize our main results and outline future
forecasting results (Lin et al., 2012). directions of further methodological develop-
Chaos theory is dealing with a time series ments based on our proposed technique.
as a realization of a nonlinear dynamical system
with a degree of determinism. Random input is
not the only possible source of irregularity in 2. RELATED WORK
a system’s output: nonlinear, chaotic systems
In the past two decades, considerable research
can produce very irregular data with purely
efforts have been focused on the hypothesis
deterministic equations of motion in an au-
that financial markets behavior is non-linear.
tonomous way. Of course, a system which has
In this section, we will briefly review some of
both, nonlinearity and random input, will most
the previous work in this area. Indeed, there
likely produce irregular data as well (Kantz &
is a continuous stream of publications report-
Schreiber, 2003).
ing more or less successful attempts to apply
. It seems very difficult to tell whether
nonlinear algorithms.
a series is stochastic or deterministic chaotic
Meade (2002) Evaluated evidence for a
or some combination of these states. More
non-linear generating process by an analysis
generally, the extent to which a non-linear de-
of the comparative accuracy of short term
terministic process retains its properties when

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
40 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

forecasts of FX rates. Forecasts were generated based forecasting models using Standard Back
by a linear AR-GARCH model and four non- propagation (SBP), Scaled Conjugate Gradient
linear methods, including three nearest neighbor (SCG) and Back propagation with Bayesian
methods and locally weighted regression. Five Regularization (BPR) for Australian Foreign
data frequencies were used: daily, four-hourly, Exchange to predict six different currencies
two-hourly, hourly and half-hourly. Using root against Australian dollar. Five moving aver-
mean square error as a measure, significantly age technical indicators were used to build the
greater accuracy than a no-change forecast was models. These models were evaluated using
achieved for two-hourly and higher frequency three performance metrics, and a comparison
data sets. These results were supported by was made with the best known conventional
simulated trading based on forecast direction. forecasting model ARIMA. All the ANN based
No evidence was found that the FX rate behavior models outperform ARIMA model (Yu et al.,
is better represented by a non-linear generating 2004). Proposed a novel nonlinear ensemble
process than by a linear model. forecasting model integrating generalized linear
Boero and Marrocu (2002) compared the auto regression (GLAR) with artificial neural
forecasting performance of different models for networks (ANN). They compared the new
the returns of three exchange rates in terms of model’s performance with the two individual
the US dollar, namely the French franc (FF/$), forecasting models —GLAR and ANN— as
the German mark (DM/$) and the Japanese yen well as with the hybrid model and the linear
(Y/$). The relative performance of non-linear combination models. Empirical results obtained
models of the SETAR, STAR and GARCH types reveal that the prediction using the nonlinear
is contrasted with their linear counterparts. The ensemble model is generally better than those
results show that if attention is restricted to mean obtained using the other models presented in
square forecast errors, the performance of the the study.
models, when distinguishable, tends to favor In Strozzi (2002), Strozzi & Zaldivar
the linear models. The forecast performance (2005), and Strozzi & Comenges (2006) they
of the models is evaluated also conditional on applied state space reconstruction techniques
the regime at the forecast origin and on density to estimate state space volume and its varia-
forecasts. This analysis produces more evidence tion. They defined a trading methodology by
of forecasting gains from non-linear models. considering a sort of acceleration in a high-
Qia and Wu (2002) Employed neural net- dimensional state space system as a kind of
work (NN) to study the nonlinear predictability momentum indicator similar to those used in
of exchange rates for four currencies at the 1-, financial technical analysis. This trading meth-
6- and 12-month forecast horizons. They found odology has been applied to high-frequency
that neural network model with market funda- exchange rates between the US Dollar and 18
mentals cannot beat the random walk (RW) in other foreign currencies from the Euro zone.
out-of-sample forecast accuracy. In general, the They concluded that, in terms of prediction
model performed more poorly when it becomes power, high-frequency foreign exchange time
more complex or when the forecast horizon series have a different behavior from a random
lengthens. Their overall results were more on walk, i.e. are more predictable. In this sense a
the negative side and suggested that neither certain amount of determinism is embedded in
nonlinearity nor market fundamentals appear the analyzed financial time series that made their
to be very important in improving exchange prediction more accurate than a random walk.
rate forecast for the chosen horizons. Gradojevic & Yang (2006) Employed
In Kamruzzaman & Sarker (2003a), Kam- a non-parametric method to forecast high-
ruzzaman & Sarker (2003b) and Kamruzzaman frequency Canada/U.S. dollar exchange rate.
& Sarker (2004) they developed and investi- The non-linear models outperformed random
gated three Artificial Neural Network (ANN) walk and linear models based on a number

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 41

of recursive out-of-sample forecasts. The jerk differential equation, which is in principle


artificial neural network (ANN) model was capable of generating chaos. The model was
consistently better in RMSE to random walk econometrically estimated in continuous time
and linear models for the various out-of-sample with Euro/Dollar data and examined for the
set sizes. The empirical results they got showed possible presence of chaotic motion. Their
that optimal ANN architecture is superior to results indicated that the possibility of chaotic
random walk and any linear competing model dynamics has to be rejected.
for high-frequency exchange rate forecasting. Lin et al. (2012) tried to address the nonlin-
In Torkamani et al. (2007) test of strange ear and non-stationary characteristics of finan-
attractor and biggest Lyapunov exponent prob- cial time series such as foreign exchange rates
ability of chaos on several foreign exchange through proposing a hybrid forecasting model
rates vs. IRR (Iranian Rial) has been investi- using empirical mode decomposition (EMD)
gated. Results showed that data in this market and least squares support vector regression
have complex chaotic behavior with big degree (LSSVR) for foreign exchange rate forecasting.
of freedom. EMD was used to decompose the dynamics of
Ding et al. (2009) studied the chaos behav- foreign exchange rate into several intrinsic mode
ior in purchase price of the RMB to the Euro. function (IMF) components and one residual
They figured out that Euro exchange rate refuses component. LSSVR was constructed to forecast
to accept the normal distribution and that Euro these IMFs and residual value individually, and
exchange rate undulation is nonlinear system; then all these forecasted values are aggregated
it does not follow the random walk hypothesis to produce the final forecasted value for foreign
proposed by effective market theory. The results exchange rates. Empirical results showed that
obtained through the R/S analysis showed that the proposed EMD-LSSVR model outperforms
Euro exchange rate undulation is indifferent the EMD-ARIMA (autoregressive integrated
with the effective market theory, it’s undulation moving average) as well as the LSSVR and
is not a simple linear system, it is a complex ARIMA models without time series decom-
nonlinear system which has chaos characteristic. position.
At the same time, the GARCH model examina-
tion implemented accept the non-linear square
difference dynamic distribution. 3. PROBLEM DEFINITION
Huang et al. (2010) implemented a chaos-
It seems very difficult to tell whether a series
based model to predict the foreign exchange
is stochastic or deterministic chaotic or some
rates. In the first stage, the delay coordinate em-
combination of these states. More generally,
bedding was used to reconstruct the unobserved
the extent to which a non-linear deterministic
phase space (or state space) of the exchange
process retains its properties when corrupted
rate dynamics. In the second stage, support
by noise is also unclear. The noise can affect
vector machines (SVMs) were constructed for
asystem in different ways even though the
forecasting. Compared with traditional neural
equations of the system remain deterministic.
networks, pure SVMs or chaos-based neural
Since a single reliable statistical test for
network models, the proposed model performed
chaoticity is not available, combining multiple
best. The root mean-squared forecasting errors
tests is a crucial aspect, especially when one is
were significantly reduced.
dealing with limited and noisy data sets like in
Federici & Gandolfo (2011)developed
economic and financial time series.
a continuous time exchange rate model that
Time series can be analyzed using two
allows for heterogeneity of the agents beliefs,
different model types:
in order to explore non-linearity and possible
chaotic behavior. The theoretical model contains
an intrinsic non-linearity that gives rise to a • Stochastic model (Random Process)

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
42 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

◦◦ Deal with Stationary Time Series mean and the variance. More subtle quantities
(ARMA model) such as spectral components, correlations or
◦◦ Deal with non-Stationary Time Series nonlinear statistics may be needed to detect
(ARIMA model) less obvious non-stationarity.
• Chaotic model (Non-linear Dynamical In this section, we will give some general
System) concepts and discuss particular techniques to
◦◦ Deterministic non-linear system. Time address the question of how non-stationarity
series derived from such systems seem can be detected for a given data set.
stochastic when analyzed with linear
techniques 4.1.1. Running Variance

We perform a very simple test of stationarity


The first step in analyzing the FOREX
by computing the sample mean and variance
dataset is to determine whether it is a random
for consecutive segments of the data set.
or deterministic non-linear. If the FOREX time
Within each segment, the standard error of the
series is deterministic non-linear system N

estimated mean s = ∑sn / N is given by


4. A PROPOSED METHOD n =1

(s − s ) , 
2
BASED ON CHAOS THEORY
N

Standard Error =
∑ n =1
(1)
4.1. Detecting Non-Stationarity N (N − 1)

Quite generally, a scientific measurement of 4.1.2. Space Time Separation Plot


any kind is in principle more useful the more
it is reproducible. We need to know that the The idea behind is that in the presence of tem-
numbers we measure correspond to properties poral correlation the probability that a given
of the studied object, up to some measurement pair of state points in the reconstructed state
error. In the case of time series measurements,
reproducibility is closely connected to two dif- {
space, s (ti ), s (ti − ∆t ), s (ti − 2∆t ), … , }
ferent notions of stationarity. The weakest but has a distance smaller than r , i.e.si − s j < r ,
most evident form of stationarity requires that does not depend only on the position of the
all parameters that are relevant for a system’s state but also on the time that has elapsed be-
dynamics have to be fixed and constant during tween them. This dependence can be detected
the measurement period (and these parameters by plotting the number of neighbor points as a
should be the same when the experiment is function of two variables, the time separation
reproduced). and the spatial distance.
Strong violations of the basic requirement In principle, one can create for each time
that the dynamical properties of the system separationan accumulated histogram of spatial
underlying a signal must not change during distances. In the case of power-law noises the
the observation period can be checked simply only points with smallspatial separation are
by measuring such properties for several seg- dynamically near neighbors, i.e. the series is
ments of the data set. Transition probabilities, non-recurrent in phase space. In this casethe
correlations, etc., computed, for example for the contour curves do not saturate. In the case of
first and second half of the data available, must stationarity, we will find saturation in the plot.
not differ beyond their statistical fluctuations. In experimental chaotic systems, it is not
Characteristics with known or negligible statisti- common for a parameter drift to result in no
cal fluctuations are preferable for this purpose. visible drift in the mean or the distribution of
The statistically most stable quantities are the values. Linear correlations and the spectrum

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 43

may also be unaffected. Only the nonlinear 2. Construct the discrete Fourier transform
dynamical relations and transition probabilities 1
N
−2 ði (m −1)(n −1)/N
Z (m ) = X (m ) + iY (m ) = ∑z e ,
change appreciably. This can be detected by N n =1
n

comparing prediction errors with respect to a (3)


nonlinear model for different sections of the 3. Construct a set of random phases
data set. (Kantz & Schreiber, 2003)
N
4.2. Testing for Nonlinearity ∅m ∈  0,ð , m = 2, 3, … (4)
with Surrogate Data 2

If the dynamics that has generated the time 4. Apply the randomized phases to the Fourier
series is unknown or if the data are noisy, it is transformed data
important to investigate whether the amount
of nonlinear deterministic dependencies is
Z (m ) =
'

worth analyzing further or whether the series



 N
can be considered as stochastic. Hence, one of 
 for m = 1 and m = +1
the first steps before applying any nonlinear  ( )


Z m 2

 N
 Z (m ) e m
i ∅
analysis on a data set is to investigate if the use for m = 2, 3,…,

 2
of such advanced techniques is justified by the  Z (N − m + 2) e N −m +2
−i∅
 N N

 for m = + 2, + 3, …, N
data. The main reason behind this reasoning is 

 2 2
that linear stochastic processes can create very (5)
complicated looking signals and that not all the
structures that we find in a data set are likely to 5. Construct the inverse Fourier transform of
be due to nonlinear dynamics going on within
Z (m )
'

the system. The method of surrogate data has


become a useful tool to address the question if N
the irregularity of the data is most likely due to 1 2 ði (m −1)(n −1)/N
Z (n ) = x (n ) + iy (n ) =
' ' '

N ∑Z ' m
e ,
nonlinear deterministic structure or rather due m =1

to random inputs to the system or fluctuations (6)


in the parameters.
The method of calculating surrogate data 4.3. Phase Space Reconstruction
sets with the same mean, variance, and power (Embedding Theorem)
spectrum but otherwise random is as follows:
First construct the Fourier transform of the ex- The essential problem in nonlinear time series
perimental time series data, then randomize the analysis is to determine whether or not a given
phases, then take the inverse Fourier transform. time series is a deterministic signal from a low-
An explicit algorithm for achieving this is as dimensional dynamical system. If it is, then
follows (Akay, 2000): further questions of interest are: What is the
dimension of the phase space supporting the
1. Input the experimental time series data data set? Is the data set chaotic?
The key to answering these questions is
x (t j ), j = 1, … N into a complex array embodied in the method of phase space re-
construction. Having stressed the importance
of phase space for the study of systems with
z (n ) = x (n ) + iy (n ), n = 1, … N (2) deterministic properties, we have to face the first
problem: what we observe in an experiment is
not a phase space object but a time series, most
where x (n ) = x (tn ) andy (n ) = 0. likely only a sequence of scalar measurements.

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
44 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

We therefore have to convert the observations state space reconstruction. However, Takens’
into state vectors. This is the important problem theorem is true for the unrealistic case of an
of phase space reconstruction which is techni- infinite, noise-free, number of points. Takens
cally solved by the method of delays (or related showed that, in this case, the choice of the time
constructions). delay is not relevant, and gave indications only
Takens’ embedding theorem (Takens, on the choice of the embedding dimension.
1981) provides theoretical foundation for the Nevertheless, in real applications, the
analysis of time series generated by non-linear proper choice of the time delay τ and the calcu-
dynamical systems. It states that a map exists lation of an embedding dimension, m, are both
between the originalstate space and a recon- fundamental for starting to analyse the data. As
structed state space. The theorem shows that a matter of fact, a lot of research on state space
the dynamical properties of the system from reconstruction has centered on the problems
the true state space are preserved through the of choosing the time delay and the embedding
embedding transformation. Therefore, a d-di- dimension which we can call the parameters of
mension chaotic attractor, a phase space d of the reconstruction for delay coordinates.
the attractor can be reconstructed from a time If the time delay chosen is too small, there is
almost no difference between the elements of the
series {x i }
N
, where N is the length of the
i =1 delay vectors, since all points are accumulated
time series, by using delay coordinate defined around the bisectrix of the embedding space:
as: this is called redundancy (Casdagli et al., 1991).
However, when τ is very large, the different
  coordinates may be almost uncorrelated. In this
Z i = x i , x i −ô , …, x i −(m −1)ô  ∈ m , (7)
  case the reconstructed trajectory may become
very complicated, even if the underlying ‘’true’’
trajectory is simple: this is called irrelevance.
whenever m ≥ 2d + 1 . In Eq. (7) ô is called
Determining the time delay and the em-
the delay time and m is the embedding dimen-
bedding dimension is considered as one of the
sion. The values of ô and m are used to trans-
most important steps in nonlinear time series
form the univariate time series into the phase
modeling and prediction. A number of methods
space vectors Z i stacked as have been developed in determining the time
delay and the minimum embedding dimension
z  x x 0−ô … x 0−(m −1)ô  since the early beginning of nonlinear time
 0   0 
z  x x 1−ô … x 1−(m −1)ô  series study. We will discuss some of them in
  1
Zu =  1  =  , the next two subsections.
 2  x 2
z x 2−ô … x 
2−(m −1)ô 
    4.3.1. Finding Time delay ( ô ) and
      Embedding Dimension ( m )
(8)
The dimension, where a time delay reconstruc-
Different choices of τ and m yield different tion of the system phase space provides a
reconstructed trajectories. Takens’ theorem has necessary number of coordinates to unfold the
been generalized by (Sauer et al., 1991) to the dynamics from overlaps on itself caused by
case where the attractor is a strange attractor projection, is called the embedding dimension,
with a fractal dimension D. The embedding of m . This is a global dimension, which can be
a strange attractor using time delay coordinates different from the real dimension.
is one to one if m≥2D+1. The usual method for choosing the mini-
The embedding theorem is important mum embedding dimension is to compute some
because it gives a rigorous justification for the invariants of the attractor. By increasing the em-
bedding dimension used for the computations,

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 45

one notes when the value of the invariant stops this projection the topological structure is no
changing. Then we can say that the geometric longer preserved.
properties of the system are unfolded. Points are projected into neighborhoods of
other points to which they would not belong in
False Nearest Neighbours: The method of false higher dimensions. These points are calledfalse
nearest neighborsis designed to determine neighbors. If now the dynamics is applied, these
how many dimensions are sufficient to false neighbors are not typically mapped into
embed a particular time series (Kennel et the image of the neighborhood, but somewhere
al., 1992). else, so that the average ‘‘diameter” becomes
quite large.
Their idea is quite intuitive. Suppose the The basic idea behind false nearest neigh-
minimal embedding dimension for a given time bors is that points in a state space should be close
to each other because their dynamical state is
series is m 0 . This means that in a m 0 -dimen-
similar, not because they have been projected
sional delay space the reconstructed attractor close to each other as an artifact of constructing
is a one-to-one image of the attractor in the the embedding with a dimension which is too
original phase space. Especially, the topologi- low. Figure 1 presents a geometric explanation
cal properties are preserved. Thus the neighbors of the concept that is at the core of the false
of a given point are mapped onto neighbors in nearest neighbors’ technique (Boker, 1996).
the delay space. Due to the assumed smoothness The time series is first embedded into a
of the dynamics, neighborhoods of the points one dimensional state space. For each point in
are mapped onto neighborhoods again. Of the state space, its nearest Euclidean neighbor
course the shape and the diameter of the neigh- is found. In the example, the line at the bottom
borhoods is changed according to the Lyapunov of Figure 1 represents the one dimensional state
exponents. But, suppose now you embed in an space and the nearest neighbor of the yellow
m -dimensional space with m < m 0 . Due to point is the green point. Next, the time series is
embedded into a two dimensional state space as

Figure 1. A geometric explanation of the false nearest neighbors’ algorithm

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
46 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

represented by the oval in the middle of Figure (D ) (D )


1. The yellow point and the green point are no xi − xj
longer near to each other, so the green point is < A , (10)
σX
labeled a “false nearest neighbor” because it
was only near to the yellow point due to the
projection of the time series onto the line. where ÃX is the standard deviation of x over
Next, the nearest neighbor for each point in the whole time series, then x i is declared to
the two dimensional state space is found. Now
have a false nearest neighbor.
the nearest neighbor to the yellow point is the
The resulting percentage of false nearest
red point. The time series is now embedded
neighbors for each embedding dimension isthen
into a three dimensional state space as repre-
plotted against the corresponding embedding di-
sented in the hemisphere at the top of Figure
mension to create a false nearest neighbors plot.
1. The yellow point and the red point are still
A one-to-one embedding can be obtained
near to each other and so the red point is not a
for any value of the delay time τ>0. However,
false nearest neighbor. This process continues
very small delay times will result in near-linear
until either there are no further false nearest
reconstructions with high correlations between
neighbors or the data set becomes so sparse in
consecutive phase space points and very large
a high dimensional space that no points can be
delays might obscure the deterministic structure
considered to be near neighbors to begin with.
linking points along a single degree of freedom.
The false nearest neighbors’ algorithm can
If the delay time is commensurate with a char-
be summarized as follows:
acteristic time in the underlying dynamics, then
this too may result in a distorted reconstruction.
1. Find the nearest neighbor for each point First local minimum of the Average mu-
in an embedding of dimension D . tual information: (Fraser & Swinney, 1986)
2. Find the percentage of those nearest neigh- suggested to use the average mutual information
bors which do not remain within a ratio of
(AMI) function, I (τ ) , as a kind of nonlinear
their original distance in an embedding of
dimension D + 1 correlation function to determine when the
values of s (n ) and s (n + τ ) are independent
enough of each other to be useful as coordinates
(D +1) (D +1) in a time delay vector but not so independent
xi − xj as to have no connection which each other at
> T , (9)
(D )
xi − xj
(D ) all. For a discrete time series, I (τ ) can be
calculated as,

where x i
(D )
− xj
(D )
is the distance between the I (τ ) =

 (
 P s n ,s n + τ
() ( )) 

two points in dimension D and x i
(D +1) (D +1)
− xj ∑ P (s (n ), s (n + τ )) log   ,
n ,n +T
2
( ) (
 P s (n ) P s (n + τ ) ) 

is the distance between the two points in dimen- (11)
sion D + 1 .

3. If there is no nearest neighbor x j for a


( )
where P s (n ) refers to individual probability

point x i such that ( )


and P s (n ), s (n + ô ) is the joint probability
density. To determine P s (n ) ( ) we simply
project the values taken from s (n ) versus n

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 47

foreign exchange rates. Table 1 summarizes the


back onto the s (n ) axis and form a histogram
foreign exchange rates considered.
of the values. Once normalised, this gives us The data span over a period of more than
( )
P s (n ) . For the join distribution of s (n ) and 17 months starting from December 1, 2011 with
s (n + ô ) we form the two-dimensional histo- a total number of 17568 readings (instances)
per each currency pair. Each instance includes
gram in the same way (Strozzi, 2002).
the date/time data (Year, Month, Day, Hour and
4.4. Testing for Nonlinearity Minute). Figure 2 represent the close price of
half hourly exchange rates for one pair of the
If the dynamics that have generated the time six currency pairs considered in analysis and
series are not known or if the data are noisy, it forecasting.
is important to investigate whether the amount
of nonlinear deterministic dependencies is worth 5.2. Detecting Non-Stationarity
analyzing further or whether the series can be
We have performed a very simple test of sta-
considered as stochastic. Hence, one of the first
tionarity on the selected exchange rates. We
steps before applying any nonlinear analysis
have computed the sample mean and variance
on a data set is to investigate if the use of such
for 73 consecutive segments of the dataset per
advanced techniques is justified by the data.
each currency pair, each segment containing
240 data points. Within each segment, the
5. CASE STUDY AND DATASET standard error of the estimated mean
N

5.1. FOREX Dataset s = ∑sn / N is given by


n =1

The datasets used in this research have been


(s − s ) ,
N 2

exported from FOREX market with a script


Standard Error =
∑ n =1
code especially coded for this purpose using
N (N − 1)
the MQL5 programming language. MQL5 is the
built-in programming language introduced by (12)
MetaQuotes for the development of automated
trading strategies. It allows creating Expert According to the way explained in (Kantz
Advisors (trading robots), Custom Indicators, & Schreiber, 2003), the running standard de-
Scripts and Libraries. The dataset exported viation showed fluctuations from on segment
consists of open, low, high and close bid quotes to another across all currency pairs, as can be
for half hourly periodicity of 6 currency pairs’ seen in AUD/USD (Figure 3).

Table 1. Foreign exchange rates considered

No. Currency Pair Symbol Currency Pair Name


1 AUD/USD Australian Dollar Against US Dollar
2 EUR/AUD Euro Against Australian Dollar
3 EUR/CAD Euro Against Canadian Dollar
4 EUR/CHF Euro Against Swiss Franc
5 EUR/USD Euro Against US Dollar
6 GBP/USD British Pound Against US Dollar

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
48 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

Figure 2. AUD/USD foreign exchange time series

Figure 3. Running standard deviation for AUD/USD

5.3. Phase Space Reconstruction in FOREX. We therefore have to convert the


(Embedding Theorem) observations into state vectors. This is the im-
portant problem of phase space reconstruction.
What we observe in an experiment is not a phase Takens’ embedding theorem (Takens,
space object but a time series, most likely only a 1981) provides theoretical foundation for the
sequence of scalar measurements as is the case analysis of time series generated by non-linear

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 49

dynamical systems. It states that a map exists embedding parameters. So, the optimization
between the original state space and a recon- function of DE algorithm aimed to get the least
structed state space. The theorem shows that prediction error, Mean Squared Error (MSE),
the dynamical properties of the system from given different combinations of embedding pa-
the true state space are preserved through the rameters. We’ve used a number of 11,900 states
embedding transformation. Therefore, a d-di- of phase space reconstructed using embedding
mension chaotic attractor, a phase space d of parameters then, partitioned this number to
the attractor can be reconstructed from a time 80% training set and 20% testing set per each
time series in our dataset. In each iteration of
series {x i }
N
, where N is the length of the
i =1 the differential evolution algorithm, the MSE
time series, by using delay coordinate defined was calculated and finally the best embedding
as: parameters that give the minimum prediction
error were selected. Table 2 summarizes the
  values of time delay and embedding dimension
Z i = x i , x i −τ , …, x i −(m −1)τ  ∈ m , (13)
  obtained by analyzing the foreign exchange
rate time series.
We have to state here that the intervals of
whenever m ≥ 2d + 1 . In Eq. (2.7) τ is called
possible values used in the DE technique for time
the time delay and m is the embedding dimen-
delay and embedding dimension were chosen to
sion. The values of τ and m are used to
be (2 - 400) and (1 - 15) respectively as proposed
transform the univariate time series into the
in the work of (Strozzi & Comenges, 2006).
phase space vectors Z i stacked as Figure 3 shows for the (AUD/USD) the
Distribution of Forecasting Mean Squared
z  x x 0−τ … x 0−(m −1)τ  Error that was calculated per each time delay
 0   0 
z  x x 1−τ … x 1−(m −1)τ  and embedding dimension tested. The highest
  1
Zi =  1  =  , point in the histogram refers to the optimum
 2  x 2
z x 2− τ … x   (14)
2−(m −1)τ  combination of time delay and embedding
   
      dimension that give the least forecasting mean
squared error.
i = 1, …, N
5.4. Testing for Nonlinearity
5.3.1. Finding Time Delay ( ô ) and
Embedding Dimension ( m ) If the dynamics that have generated the time
series are not known or if the data are noisy, it
Through our investigation in the literature for
is important to investigate whether the amount
the methods used to get the optimal time delay
of nonlinear deterministic dependencies is worth
and embedding dimension, it has been found
analyzing further or whether the series can be
that there are two very popular methods to do
considered as stochastic. Hence, one of the first
that and most of researchers have used them.
steps before applying any nonlinear analysis
First one is the Average Mutual information
on a data set is to investigate if the use of such
(AMI) technique which is used to get the time
advanced techniques is justified by the data.
delay. The second one is the False Nearest
The principle of a statistical test is the
Neighbor (FNN) technique which is used to
comparison of a parameter measured on a given
get the embedding dimension.
realization to the distribution of the same param-
Instead of using the two common methods
eter under presumption of the null hypothesis.
of finding the embedding parameters, we’ve
In case the measured parameter does not match
contributed by using the differential evolution
this distribution at a certain level of significance,
algorithm (as a global optimization technique)
we reject the null hypothesis. (Akay, 2000).
in the selection process to find the optimal

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
50 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

Table 2. Time Delay and Embedding Dimension found for the Foreign exchange rate time series
data sets

Dataset Time Delay ( τ ) Embedding Dimension ( m )

AUD/USD 325 8
EUR/AUD 176 7
EUR/CAD 179 7
EUR/CHF 244 3
EUR/USD 151 11
GBP/USD 320 5

One way to establish a suitable distribution program (surrogates) within TISEAN (Hegger
is given by the concept of surrogate data. We et al., 1999), (Schreiber & Schmitz, 2000) to
generate artificial data sets that are consistent generate the 19 surrogate datasets.
with the null hypothesis with respect to the For the reconstruction of phase space from
following two properties: (they are random & original time series and its corresponding sur-
they pass the same hypothetical linear system rogate datasets we’ve used the pre-calculated
as our measured time series). See Figure 4 for embedding parameters. After the reconstruction
an example. We then estimate the parameter on of the phase space we’ve considered a number
this data ensemble. of 11900 states for training and testing. We then
In order to test the null hypothesis, we split the phase space states into two half’s one
have created 19 surrogate datasets for each half for training (5950 state) and the other for
foreign exchange time series. We have used the testing (5950 state). SVR was used to forecast

Figure 4. Distribution of forecasting mean squared error for AUD/USD using DE algorithm

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 51

test values. The statistic we have used to reject Regression (SVR) over the original time series
the null hypothesis is Mean Absolute Error of exchange rates (pure-SVR), and the phase
(MAE) in the nonlinear prediction of the original space that is reconstructed using the same time
time series and the median of it across the (19) series (Chaos-SVR), to see the effect of using
corresponding surrogate datasets. To visualize the chaos theory on foreign exchange rates
the whole process of nonlinearity detection, after the analysis hold confirms the existence
please refer to Figure 5 and Figure 6. of nonlinearities in the datasets. We have used
Table 3 summarizes the results. As it can the LIBSVM toolbox (Chang & Lin, 2011) to
be seen, 5 datasets out of 6 reject the null hy- implement our forecasting technique over the
pothesis, which means that in all these cases considered foreign exchange rates datasets.
the nonlinear prediction error was lower in
the original time series than in the surrogate 6.1. Data Partitioning
datasets. This means that these time series do
have some dynamical nonlinearity. For the purpose of comparing forecasting ac-
curacy across multiple currency pairs, we have
started with the reconstruction of phase space
6. RESULTS AND DISCUSSION from currency datasets using the embedding
parameters we have calculated by the differ-
Our main goal of this research was to get the ential evolution algorithm. The second step
most accurate predictions of financial time was considering a number of (15000) phase
series; in our case we have considered the space states per each currency pair. Then we
foreign exchange rates, through investigating have partitioned this number of states to 10
the existence of nonlinearities and investi- partitions each has a number of (1500) states.
gating the nonlinear forecasting techniques. Each one of those 10 partitions was divided into
We’ve decided to hold a comparison between two sets, training and testing data set by 80/20
forecasting performance of Support Vector principle. 80% training set consists of (1200)

Figure 5. Original time series (AUD/USD) in blue and three surrogate time series

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
52 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

Figure 6. Flow chart of the detection process of nonlinearities, where (N) refers to the number
of surrogate datasets considered (19 surrogate datasets were considered in this work)

Table 3. Results from the surrogate data test. Rejection of the null hypothesis means that the
non-linear prediction error found in the original time series is lower than in the surrogate time
series generated by a stationary Gaussian linear process

Dataset Original Time Series (19) Surrogate Time Series H0


AUD/USD(MAE) 5.1063e-04 6.4099e-04 Rejected
EUR/AUD(MAE) 8.7503e-04 6.7445e-04 Don’t Reject
EUR/CAD(MAE) 5.8105e-04 7.4722e-04 Rejected
EUR/CHF(MAE) 2.1573e-04 6.5370e-04 Rejected
EUR/USD(MAE) 6.8341e-04 7.8386e-04 Rejected
GBP/USD(MAE) 5.7166e-04 7.2303e-04 Rejected

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 53

states and 20% testing set consists of (300) time series. This method was based on the use
states. The states were fed to SVR as feature of differential evolution technique together
vectors. Then, the median error was calculated with the prediction error as a tuner for the
across the 10 partitions. We have used the same optimization function. After getting the embed-
way to calculate the error over the original time ding parameters we were able to compare the
series taking into consideration the use of the prediction error of the two models used and we
same intervals of the time series that have been could conclude that: due to the inherent chaotic
used to reconstruct the phase space. behavior in exchange rate dynamics, the phase
space reconstruction could effectively extracts
6.2. Comparison of the time series dynamical features and therefore
Forecasting Accuracy enhances the prediction power of traditional
models.
The performance comparisons under median Future improvements to the work presented
of MSE are listed in Table 4, where bold-faced here might consist of applying multivariate
numbers indicate the minimum errors in each phase space reconstruction by using the time
exchange rate. series of open price, high price, and low price
As seen from the above table we could say available in the dataset in addition to the close
that the results we’ve got using the datasets price that was used in this research. Another
considered in this work refer to: forecasting good work that could be done here is investigat-
in phase space outperformed forecasting of ing hybrid techniques that uses both hypothesis
original time series in almost all the exchange of chaos and the hypothesis of stochastic process
rates covered in this research. and compare it to our work. Also it might be
Figure 7 presents actual rates (green line), useful to apply this scientific work in the real
predicted values (red line) and prediction er- market by building a system for traders that
rors (blue line) of (SVR) model over EUR/ could be an expert advisor or some indicator that
AUD where Figure 8 presents the same lines uses the predicted values to do some real trad-
for (Chaos-SVR) model over the same period ing automatically or giving advices to traders.
of time for EUR/AUD. These are the results
of one partition (out of ten) considered for the
same currency pair. ACKNOWLEDGMENT
This paper was funded by the Deanship of
7. CONCLUSION Scientific Research (DSR), King Abdulaziz
University, Jeddah, KSA. The authors, there-
In this work, we’ve proposed a new method of fore, acknowledge with thanks DSR technical
estimating the embedding parameters needed and financial support.
for the reconstruction of phase space from a

Table 4. Model performance under median (MSE)

Dataset SVR Chaos-SVR


AUD/USD (MSE) 7.6433e-07 6.4411e-07
EUR/AUD (MSE) 2.0491e-06 9.3641e-07
EUR/CAD (MSE) 1.0375e-06 8.5606e-07
EUR/CHF (MSE) 3.2388e-07 1.9654e-07
EUR/USD (MSE) 1.2385e-06 9.5249e-07
GBP/USD (MSE) 9.7126e-07 9.7100e-07

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
54 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

Figure 7. EUR/AUD forecasting performance of SVR

Figure 8. EUR/AUD forecasting performance of Chaos-SVR

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 55

REFERENCES Huang, S. C., Chuang, P. J., Wu, C. F., & Lai, H.


J. (2010). Chaos-based support vector regressions
Akay, M. (2000). Nonlinear Biomedical Signal for exchange rate forecasting. Expert Systems with
Processing: Dynamic Analysis and Model- Applications, 37(12), 8590–8598. doi:10.1016/j.
ing (Vol. II). New York: Wiley-IEEE Press; eswa.2010.06.001
doi:10.1109/9780470545379 Kamruzzaman, J., & Sarker, R. A. (2003a). Compar-
Boero, G., & Marrocu, E. (2002). The Performance ing ANN Based Models with ARIMA for Prediction
of Non-linear Exchange Rate Models a Forecasting of Forex Rates. ASOR BULLETIN, 22(2), 2–11.
Comparison. Journal of Forecasting, 21(7), 513–542. Kamruzzaman, J., & Sarker, R. A. (2003b). Forecast-
doi:10.1002/for.837 ing of Currency Exchange Rates using ANN: A Case
Boker, S. M. (1996). Linear and Nonlinear Dynami- Study. IEEE Int. Conf. Neural Networks & Signal
cal Systems Data Analytic Techniques and an Ap- Processing, Nanjing, China(Vol. 1, pp. 793-797).
plication to Developmental Data. PhD Dissertation, doi:10.1109/ICNNSP.2003.1279395
Charlottesville, Virginia. Kamruzzaman, J., & Sarker, R. A. (2004). ANN-
Box, G. E., & Jenkins, G. M. (1970). Time Series Based Forecasting of Foreign Currency Exchange
Analysis: Forecasting and Control. John Wiley & Rates. Neural Information Processing-Letters and
Sons. Reviews, 22(2), 49–58.

Casdagli, M., Eubank, S., Farmer, J. D., & Gibson, Kantz, H., & Schreiber, T. (2003). Nonlinear Time
J. (1991). State space reconstruction in the presence Series Analysis (2nd ed.). New York: Cambridge
of noise. Physica D. Nonlinear Phenomena, 51(1), University Press; doi:10.1017/CBO9780511755798
52–98. doi:10.1016/0167-2789(91)90222-U Kennel, M. B., Brown, R., & Abarbanel, H. D. (1992).
Chang, C. C., & Lin, C. J. (2011). LIBSVM: A library Determining embedding dimension for phase-space
for support vector machines. ACM Transactions reconstruction using a geometrical construction.
on Intelligent Systems and Technology, 2(3), 1–27. Physical Review A., 45(6), 3403–3411. doi:10.1103/
doi:10.1145/1961189.1961199 PhysRevA.45.3403 PMID:9907388

Ding, Y., Li, S., & Li, L. (2009). An Analysis on the Lin, C. S., Chiu, S. H., & Lin, T. Y. (2012). Em-
Chaos Behavior of currency Exchange Rate Undu- pirical mode decomposition–based least squares
lation. First International Workshop on Education support vector regression for foreign exchange rate
Technology and Computer Science, Wuhan, Hubei forecasting. Economic Modelling, 29(6), 2583–2590.
(Vol. 2, pp. 599-602). doi:10.1109/ETCS.2009.394 doi:10.1016/j.econmod.2012.07.018

Federici, D., & Gandolfo, G. (2011). The Euro/Dol- Meade, N. (2002). A comparison of the accuracy of
lar Exchange Rate: Chaotic or Non-Chaotic? CESifo short term foreign exchange. International Journal
Area Conference on Macro, Money and International of Forecasting, 18(1), 67–83. doi:10.1016/S0169-
Finance, Munich. 2070(01)00111-X

Fraser, A., & Swinney, H. (1986). Independent coordi- Qia, M., & Wu, Y. (2002). Nonlinear prediction of
nates for strange attractors from mutual information. exchange rates with monetary fundamentals. Journal
Physical Review A., 33(2), 1134–1140. doi:10.1103/ of Empirical Finance, 10(5), 623–640. doi:10.1016/
PhysRevA.33.1134 PMID:9896728 S0927-5398(03)00008-2

Gradojevic, N., & Yang, J. (2006). Non-linear, Sauer, T., Yorke, J. A., & Casdagli, M. (1991). Em-
Non-parametric, Non-fundamental Exchange Rate bedology. Journal of Statistical Physics, 65(3-4),
Forecasting. Journal of Forecasting, 25(4), 227–245. 579–616. doi:10.1007/BF01053745
doi:10.1002/for.986 Schreiber, T., & Schmitz, A. (2000). Surrogate time
Hegger, R., Kantz, H., & Schreiber, T. (1999). Practi- series. Physica D. Nonlinear Phenomena, 142(3),
cal implementation of nonlinear time series methods: 346–382. doi:10.1016/S0167-2789(00)00043-9
The TISEAN package. Chaos: An Interdisciplin- Strozzi, F. (2002). Application of Non-linear Time
ary Journal of Nonlinear Science, 9(2), 413–435. Series Analysis Techniques to High Frequency Cur-
doi:10.1063/1.166424 PMID:12779839 rency Exchange Data. Serie Metodi quantitativi 14 .

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
56 International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015

Strozzi, F., & Comenges, J. M. Z. (2006). Towards a Torkamani, M. A., Mahmoodzadeh, S., Pourroostaei,
non-linear trading strategy for financial time series. S., & Lucas, C. (2007). Chaos Theory and Application
Chaos, Solitons, and Fractals, 28(3), 601–615. in Foreign Exchange Rates vs. IRR (Iranian Rial).
doi:10.1016/j.chaos.2005.08.006 International Journal of Humanities and Social
Science, 1(3).
Strozzi, F., & Zaldivar, J. M. (2005). Non-linear
forecasting in high-frequency financial time series. Yu, L., Wang, S., & Lai, K. K. (2004). A novel
Physica A: Statistical Mechanics and its Applications, nonlinear ensemble forecasting model incorporating
353, 463–479. doi: 10.1016/j.physa.2005.01.047 GLAR and ANN for foreign exchange rates. Com-
puters & Operations Research, 30(10), 2523–2541.
Takens, F. (1981). Detecting strange attractors in doi::10.1016/j.cor.2004.06.024
turbulence. Lecture Notes in Mathematics, 898,
366–381. doi:10.1007/BFb0091924

Ahmed Hamdy Radhwan was born in 1983, Jeddah, Saudi Arabia. He received the B.S. degree
in 2005 from the Department of Computer Science in the Faculty of Computing and Information
Technology, King Abdul Aziz University (KAU), Jeddah, Saudi Arabia. He received M.S. degree
in 2014 from the Department of Computer Science in the Faculty of Computing and Information
Technology, King Abdul Aziz University (KAU), Jeddah, Saudi Arabia. Currently, he is preparing
to get enrolled in the Ph. D. program at King Abdul Aziz University (KAU). His research interest
includes artificial intelligence, pattern recognition, time series forecasting, chaos theory and
brain computer interface.

Mahmoud I. Kamel was born in 1955, Cairo, Egypt, Bsc. from Electronic and communication
department (1978) Cairo university. PhD. Systems and Computer Engineering, 1991, Al-Azhar
University. Visiting Professor, University of Al Ain - United Arab Emirates, Al Ain, United Arab
Emirates(1993). 1993 - 2002: Consultant, Research Center, Cairo University, Cairo, Egypt. 2002-
2012 King Abdulaziz University (Computer Science Department). Research Interests:- Industrial
Automatic Control, Modeling and Simulation, Artificial Intelligence, Pattern Recognition, Brain
Computer Interface, text mining, information retrieval.

Mohammed Yehia Dahab is an assistant professor at the Department of Computer Science in


the Faculty of Computing and Information Technology, King Abdul Aziz University (KAU), Jed-
dah, Saudi Arabia. He served as the Chairman of the agricultural expert systems development
department for 2 years (2006-2009) at The Central Laboratory for Agricultural Expert Systems
(CLAES), Ministry of Agriculture Egypt. In addition, he served as a knowledge engineer in The
International Center for Agricultural Research in Dry Areas (ICARDA) for a year (2005-2006).
His main research interests include pattern recognition, prediction, natural language processing,
expert systems, knowledge bases, text mining and information retrieval.

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
International Journal of Rough Sets and Data Analysis, 2(1), 38-57, January-June 2015 57

Aboul Ella Hassanein, Dean of the Faculty of Computing and Information - University of Beni
Suef, the founder and head of the Egyptian Scientific Research Group (SRGE) and a professor of
information technology at the Faculty of Computer and Information, Cairo University. Professor
Hassanien has more than 500 scientific research papers published in prestigious international
journals and he has more than 30 books in the topics of data mining and medical images and
intelligent systems address social networks and smart environment. Hassanien is a collaborative
researcher member of the Computational Intelligence Laboratory in the Department of Electri-
cal and Computer Engineering at the University of Manitoba. Professor Abo is the chair of the
Computer Science and Information Technology Division at the Egyptian Syndicate of Scientific
Professions (ESSP). His research interests include Computational intelligence, medical image
analysis, security, animal identification and multimedia data mining. Prof. Hassanien won sev-
eral awards including the best researcher of the youth award of astronomy and Geophysics of
the National Research Institute - Academy of Scientific Research, Egypt in 1990 and received
a scientific excellence in the humanities from the University of Kuwait for the 2004 award and
received the superiority of scientific - University Award Cairo university in 2013 and was hon-
ored in Poland as the best scholar in the Arab smart technology . Hassanien is the founder and
head of Africa Scholars Association in information and communication technology and he has
many contributions to the fields of society and the environment service, where more than 40
organized workshops in most universities in the governorates of Egypt has been honored by a
number of university presidents and governors on activity in the service of society. He has also
many activities that support the disability and children of the Internet, where he disseminate the
importance of the role of information and communication technology in the service of people
with disabilities and the protection of children online in addition to organizing celebrations for
persons with disabilities that the most important celebration of the Egyptian disabled – “under
the partners in one world without barriers” and smartphones techniques which is the most
important scientific workshops and their role in the development of visually impaired skills in
various different areas of life, and a conference entitled Effect of communications and social
networking technology to integrate disabled people logo the art of etiquette in dealing with people
with special needs - and the issue of the future of disability in the light of changes.

Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.

View publication stats

You might also like