Financial time series forecasting using neuro fuzzy approach for the Bucharest Stock Exchange.
Trifan, Alina Lucia
Abstract: The motivation of this study is the research of an
originally engineering field, but with important implications and
applications for economics, in general and finance, in particular.
Forecasting financial time series is the goal of many studies that
combine concepts from various disciplines, in terms of classical
theories or of the latest approaches, forming a point of great interest.
Financial time series forecasting using neural networks, fuzzy systems,
neuro fuzzy, genetic algorithms leads the research into the area of
intelligent technology in an attempt to characterize the dynamic,
hyperactive, catalyzing system of the capital markets, the behavior of
the investors acting in this environment, the specific relationships.
Key words: fuzzy logic, neuro fuzzy systems, financial time series,
forecasting, artificial neural networks
1. INTRODUCTION
The current business environment has become dependent on
intelligent problem-solving techniques, continually being researched and
developed methods and models in which the neural network rules are
combined with genetic algorithms, fuzzy logic, neuro fuzzy systems and
fuzzy expert systems.
Neural networks cannot be assinged an initial set of knowledge and
have to follow a learning algorithm, a process depending on a time
variable component that can not always ensure good results. Thus, the
disadvantage of neural networks in terms of lack of transparency in the
process of collecting, handling and processing of input data in output
data led to the development of fuzzy expert systems, a special case of
expert systems.
While neural networks have the information represented in the form
of specific links, called weights, fuzzy systems are based on fuzzy
logic, representing the information into fuzzy sets.
Fuzzy logic systems have the ability to interpret imprecise data
and this is the reason why they can be useful in the decision making
process. Handling and explaining inaccurate data and arguing decisions
in linguistic form in the context of available factors is an advantage
of fuzzy expert systems, but they can not automatically acquire the
linguistic rules underlying the formulation of such decisions.
Neuro fuzzy systems are structures that combine in a hybrid
intelligent technology the advantages of neural networks--their ability
to learn and to adapt, to those of fuzzy logic--the ability of
management of the human specific reasoning at linguistic level,
transparency, interpretation of the generated models and handling
uncertain data.
2. LITERATURE REVIEW
Fuzzy expert systems have been applied in numerous studies to solve
various prediction problems (Bezdek, 1993; Bolloju, 1996; Kaneko, 1996;
Al-Shammari & Shaout, 1998; Kee, 2002; Yang et al., 2004; Huarng and
Yu, 2005; Hassan et al., 2007; Tai-Liang Chen et al., 2008; Hakan,
2010).
Combining the learning abilities of the artificial neural networks
and fuzzy logic, neuro fuzzy approaches have emerged. Many studies
exploit hybrid neuro fuzzy systems, obtaining encouraging results and
proposing different architectures (Gupta, 1994; Brown & Harris,
1994; Pedrycz, 1995; Buckley & Hayashi, 1995; Dash et al., 1995; Lie
& Sharaf, 1995; Studer & Masulli, 1997; Padmakumari et al.,
1999; Mitra & Hayashi, 2000; Kulkami, 2001; Lee et al., 2002;
Craiger et al., 2003; Kim et al., 2004; Dusan, 2004; Radeerom et al.,
2008; Atsalakis & Valavanis, 2009; Venugopal et al., 2009; Chaudhuri
et al., 2009; Jagric et al., 2010; Ebrahimpour et al., 2011).
3. CASE STUDY
3.1 The structure of fuzzy logic systems
A fuzzy logic system contains blocks, fuzzy sets used to sort
incoming data by category, a process called fuzzification, decision
rules that are applied to each set and a mechanism, called
defuzzification to generate an output from the results of the rules.
In the fuzzification stage to each point (data) of each set is
assigned a degree of membership (DOM) determined by the membership
function, that may have different geometric shapes (bell-shaped,
triangular, trapezoidal, singleton) and may be heterogeneous (e.g. fuzzy
Markov chain).
Each rule inherits a degree of membership that is the result of the
composition of the inputs degrees of membership. These rules may be
introduced by a human factor before running the system or in the case of
FAM model the rules are learned online.
Defuzzification is a procedure that creates a real non-fuzzy result
by combining the results of the fuzzy rules.
[FIGURE 1 OMITTED]
The two main steps of the inference process are aggregation and
composition. Aggregation is a process of assessment of the rules from
the previous part IF of the conditional instruction and composition is
similar to the instruction conducted in a final decision instruction
THEN.
During aggregation to each condition in the IF rule is assigned a
degree of truth based on the degree of affiliation of the corresponding
linguistic term. By applying the minimum function (min) or arithmetic
operator product (prod) to the truth degrees of the conditions can be
determined the degree of truth of IF part, that is assigned as degree of
truth to the THEN part.
Although systems based on fuzzy logic are getting good results on
the forecast, the construction process of a fuzzy system is subjective
and choosing the membership functions and rules for each scenario
follows a heuristic pattern. Fixing the rules and selecting the
membership functions in this manner are subject to intuition, trial and
error.
3.2 The hybrid neuro fuzzy model
Having considered the advantages and disadvantages of artificial
neural networks and fuzzy expert systems, the next step has resulted in
the emergence of a hybrid model, combining the ability of learning,
adaptation and generalization of neural networks with transparency and
functionality of fuzzy systems, creating the hybrid neuro fuzzy systems.
Hybrid neuro fuzzy systems are endowed with specific modeling
techniques of neural networks (the unique learning ability of a
black-box type pattern) and of fuzzy logic structures (transparent
interpretation and analysis of fuzzy IF-THEN rules).
With the ability to acquire and interpret linguistic IF-THEN rules
by human experts, fuzzy neural networks with basis functions provide the
architecture that combines in a uniform manner both numerical
information and language skills.
Kosko (1992) proposed a remarkable pattern of neuro fuzzy system
called FAM (Fuzzy Associative Memory), which involves the use of
triangular membership functions and the scheme was implemented in the
Trajectory Error Learning model as a feedforward component.
A FAM component contains two subcomponents: the rules of fuzzy
logic and pattern relationships, the template entry/exit, called
associative memory.
Three types of FAM have been used in various applications: Mamdani
fuzzy model, Sugeno fuzzy model and fuzzy model of Tsukamoto.
3.3 Building fuzzy logic and neuro fuzzy system
Building a fuzzy logic system
The fundamental idea of building a fuzzy system is substituting
numerical values specific to neural networks with linguistic terms
features of the fuzzy logic that will be converted to quantitative
signals belonging to the interval [0,1].
The frame for building the indicators K, D, KD, KDM1, KDP1 follows
the methodology implemented in Trifan (2010).
Neuro fuzzy approach
The basic idea of a neuro fuzzy system is to determine the
parameters of a fuzzy system using methods of learning specific to
artificial neural networks.
In order to achieve this approach this study uses the FAM method,
characterized by fuzzy rules with associated weights, allowing the error
backpropagation (EBP) algorithm specific to backpropagation neural
networks to combine with fuzzy logic. This technique is useful for
generating and optimizing the membership functions and the weights
associated with each rule in the data set.
4. RESULTS
Even if models based on AI have achieved better results, consisting
in superior values of performance compared to conventional models, they
remain difficult to understand, apply and interpret, leaving open other
possible options for better choices and combinations of parameters
(learning rates, momentum, the choice of certain algorithms, linguistic
variables, membership functions).
5. CONCLUSION
In the global crisis context the current financial climate hardly
finds predictability in classical models, even in the latest approaches.
What stands out in favor of artificial neural networks, fuzzy
logic, hybrid neuro fuzzy systems is their power to picture and
replicate human features (learning, modeling, generalization, patterns,
sorting).
The additional features of the individual categories that use
artificial intelligence technology provide them the capacity to solve
specific problems. Such capabilities and restrictions have caused
advantages and disadvantages of using one or another existing technology
and have created smart hybrid systems able to solve many complex
problems.
6. REFERENCES
Atsalakis, G.S.; Valavanis, P. (2009). Forecasting stock market
short-term trends using a neuro-fuzzy based methodology, Expert Systems
with Applications, 36, pp. 10696-10707
Ebrahimpour, R.; Nikooc, H.; Masoudniad, S.; Yousefie, M.; Ghaemif,
M.S. (2011). Mixture of MLP-experts for trend forecasting of time
series: A case study of the Tehran stock exchange, International Journal
of Forecasting, 27 (3), pp. 804-816
Hassan, R.; Nath, B.; Kirley, M.A. (2007). Fusion model of HMM, ANN
and GA for stock market forecasting, Expert Systems with Applications,
33, pp. 171-180
Huarng, K.; Yu, H.K. (2005). A type-2 fuzzy time-series model for
stock index forecasting, Physica A, 353, pp. 445-462
Kosko, B. (1992). Neural Networks and Fuzzy Systems: A Dynamical
Systems Approach to Machine Intelligence, Prentice Hall, pp. 323-333
Radeerom, M.; Srisa-an, C. and Kulthon Kasemsan, M. L. (2008).
Prediction Method for Real Thai Stock Index Based on Neurofuzzy
Approach, Trends in Intelligent Systems and Computer
Engineering--Lecture Notes in Electrical Engineering, 6, pp. 327-347
Trifan, A.L. (2010). Financial time series forecasting using neural
networks: A case study of the Bucharest Stock Exchange, Annals of DAAAM
for 2010 & Proceedings of the 21st International DAAAM Symposium,
pp. 1381-1383, ISBN 978-3-901509-73-5, ISSN 1726-9679
Yang, M.-S.; Hwang, P.Y.; Chen, D.H. (2004). Fuzzy clustering
algorithms for mixed feature variables, Fuzzy Sets and Systems, 141, pp.
301-317
Tab. 1. Properties of linguistic values and variables
Linguistic variable Linguistic value
Name Value Type
[0,100] input very_small, small,
medium_small,
medium,
medium_high,
high, very_high
KDM1 [-1,1] input negative, positive
KDP1 [-1,1] output strong_decrease,
small_decrease,
stable,
small_growth,
strong_growth
Tab. 2. Comparative average results for statistics and
performance indicators
Model [R.sup.2] [[bar.R].sup.2] h (N = 15144) HR RMSE
Linear 0.7444 0.7432 12,160 0.8030 1.1010
Regression
ANN -- -- 13,262 0.8757 0.8598
Neuro -- -- 13,781 0.9100 0.7807
fuzzy