Versión de HTML Básico
Elsy Gómez-Ramos y Francisco Venegas-Martínez
Analíti a
k
6
Revista de Análisis Estadístico
Journal of Statistical Analysis
1 Introduction
The efficient market hypothesis states that stock prices
come from a random walk, which implies that the stock
returns are not predictable for the public. However, there
exists significant empirical evidence that rejects such a hy-
pothesis. For example, there are studies that focus on the
persistence and long memory in the volatility of stock mar-
kets (Sharth and Medeiros, 2009; Venegas-Martínez and
Islas-Camargo, 2005), as well as others that sustain calen-
dar effects (McNelis, 2005). These studies leave the possi-
bility open to predict the behavior of those markets and,
surprisingly, the number of research papers supporting the
possibility to forecast the prices of this kind of markets is
vast and growing.
Traditionally, econometrics has provided a widely ran-
ge of tools like the GARCH model for forecasting stock
prices and exchange rates. However, the rigidity (linear
in mean) and the violation of assumptions (non-negativity
of the coefficients) of such symmetric models have been
discussed in many studies; these models cannot account
for leverage effects, although they can account for volati-
lity clustering (volatility appears in groups), leptokurtosis
(kurtosis excess), and fat tails (extreme values have a big-
ger probability than that obtained from the Normal distri-
bution).
The above facts have motivated the use of more flexi-
ble models in order to capture in a better way the finan-
cial markets behavior (Brooks, 2006; McNelis, 2005). Some
of these models come from Artificial Intelligence (AI) that
is characterized by its flexibility and capability to integra-
te different methodologies that somehow try to emulate
the biological systems behavior. Within this field, we can
find the Artificial Neural Networks (ANNs) that attempt to
emulate the human brain functions; see, for instance, An-
derson (2007).
There are many potential advantages offered by the
ANNs, for instance: i) non-linearity, that is, the neural pro-
cessor is basically non-linear, ii) input-output mapping,
in other words, through supervised learning the network
learns according to the examples, iii) adaptability, that is
to say, the network has the ability to adapt their synap-
tic weights even in real time, iv) response capacity, in ot-
her words, in the context of pattern classification the net-
work not only provides a pattern selection but also the re-
liability of decision making, v) fault tolerance due to the
massive interconnection, vi) integrated large scale, that is,
its parallelism makes it potentially faster for certain tasks
and thus capturing complex behaviors, vii) uniformity in
the analysis and design, that is to say, the same notation is
used in all fields engaged with networks, and viii) neuro-
biology analogy (Haykin, 1994). In general, the ANNs are
data-drive, self-adaptive and non-linear methods that do
not require specific assumptions about the underlying mo-
del.
Yet, there have been severe criticisms in applications
of networks in finance, we may mention, for instant that:
a) the estimated coefficients obtained by the network do
not have a real interpretation, b) there are no specific tests
available in order to consider that a model is adequa-
te, and c) the results are satisfactory inside the sample,
but outside the sample are poor (Brooks, 2006). Despite
of these critiques, the ANNs have been successfully ap-
plied in some specific finance areas. For example, the clas-
sification of areas proposed by Mender
et al
. (1996) is ba-
sed on the decision-making (credit analysis, mortgage risk,
project management, investment portfolios, price analysis,
and corporate bankruptcy). While in Burrell and Folarin
(1997) are mentioned applications in other specific areas
(financial analysis, corporate bankruptcy, risk assessment,
stock markets forecasting). There is also another available
simplified classification in three groups: credit assessment
(credit rating, credit risk, and bond pricing), portfolio ma-
nagement (optimal portfolio selection, and portfolio selec-
tion) and forecast and planning (predicting corporate ban-
kruptcies; see Bahrammirzae, 2010). Yet, one of the most
attractive applications in finance is forecasting financial ti-
me series, especially stock indices and exchange rates; in
this regard, several investigations consider stock indices
and exchange rates as indicator for the future conditions
of the economic and financial system.
Since their application in finance in the early 90’s, the
ANNs have become popular, partly because they are con-
sidered as non-parametric models from a statistical point
of view. This feature makes them quite flexible in mode-
ling real-world phenomena where observations are gene-
rally available, but there is not a theoretical relationship or
specification, especially for non-linear functions (Haykin,
1994; Mehrotra
et al
. 2000).
One of the most known networks is the MLP, which
is characterized for being a universal approximator and
classifier. The construction of the MLP for financial and
economic series forecasting is described in Kaastra and
Boyd (1996) and Mehrotra
et al
. (2000). Also, the ANNs
performance has been compared with traditional models
in finance in Burrell and Folarin (1997), Hamid and Iqbal
(2004), Khashei and Bijari (2011), McNelis (2005), and Pali-
wal and Kumar (2009). When performance focuses on the
various fields of AI in finance applications, see, for instan-
ce, Bahrammirzaee (2010) and Rada (2008).
Unlike previous studies, which analyze the performan-
ce of various networks with traditional models in many
areas, this research makes a comparison among different
types of networks to forecast particularly stock indices (or
stocks) and/or the exchange rates. In general, we can iden-
tify about five groups of networks used as approximators
and/or classifiers: (1) Feedforward Networks, like MLP,
(2) Recurrent Networks, (3) Polynomial Networks, (4) Mo-
dular Networks, and (5) Support Vector Machine. In this
paper, we shall analyze several research works that apply
the MLP and other type of network for the stock indices
and/or the exchange rate. The objective is to assess the per-
formance when applying different types of networks in re-
lation to MLP.
The research is organized as follows. In section 2, we
introduce the MLP. In section 3, we present an overview of
each networks group. In section 4, we carry out a compara-
tive analysis of the ANNs. In section 5, we assess the per-
8
Analítika,
Revista de análisis estadístico
, 3 (2013), Vol. 6(2): 7-15