CIRJE-F-165

Can the Neuro Fuzzy Model Predict

Stock Indexes Better than its Rivals?
Chin-Shien Lin
Providence University
Haider Ali Khan
University of Denver / CIRJE, University of Tokyo
Chi-Chung Huang
Providence University
August 2002

Discussion Papers are a series of manuscripts in their draft form.

They are not intended for

circulation or distribution except as indicated by the author. For that reason Discussion Papers may
not be reproduced or distributed without the written consent of the author.

Can the neuro fuzzy model predict stock indexes

better than its rivals?

Chin-Shien Lin
Associate Professor
Department of Finance
Providence University
200 Chungchi Rd., Shalu, Taichung Hsien, 433 Taiwan, R.O.C., e-mail:
cslin@pu.edu.tw

Haider A. Khan
University of Denver
Denver
Co. 80208 USA
Tel. 303-871-4461/2324
Fax 303-871-2456
hkhan@du.edu

Chi-Chung Huang
Graduate School of Business Administration
Providence University
Revised, March 2002.

Can the neuro fuzzy model predict stock indexes

better than its rivals?

Abstract
This paper

develops a model of a trading system by using neuro fuzzy

framework in order to better predict the stock index.

Thirty well-known stock

indexes are analyzed with the help of the model developed here. The empirical
results show strong evidence of nonlinearity in the stock index by using KD
technical indexes.

The trading point analysis and the sensitivity analysis of

trading costs show the robustness and opportunity for making further profits
through using the proposed nonlinear neuro fuzzy system. The scenario analysis
also shows that the proposed neuro fuzzy system performs consistently over time.

Key words: linear, nonlinear, KD indexes, buy and hold, neuro fuzzy

1. Introduction
Accurate predictions of stock market indexes are important for many reasons.Chief
among these are the need for the investors to hedge against potential market risks, and
the opportunities for market speculators and arbitrageurs to make profit by trading
indexes. Clearly, being able to accurately forecast stock market index has profound
implications and significance to both researchers and practitioners.
The most commonly used techniques for stock price forecasting are regression
methods and ARIMA models (Box and Jenkins, 1970). These models and methods
have been used extensively in the past. However, they fail to give an accurate
forecast for some series because of their linear structures and some other inherent
limitations. Although there are ARCH/GARCH models (Engle, 1982; Bollerslev,
1986) to deal with the non-constant variance, still some series cannot be explained or
predicted satisfactorily.

Recent research in the area of neural network technique has

shown that neural networks possess the properties required for relevant applications,
such as nonlinear and smooth interpolation, ability to learn complex non-linear
mappings, and self-adaptation for different statistical distributions.
However, neural network cannot be used to explain the causal relationship
between the input and output variables. This is because of the essentially black box
like nature of the many existing neural network algorithms. A neural network

cannot be initialized with prior knowledge. The network usually must learn from
scratch. The learning process itself can take very long with no guarantee of success.

On the other hand, the fuzzy expert system approach has been applied to
different forecasting problems (Bolloju, 1996; Kaneko, 1996; Al-Shammari and
Shaout, 1998), whereby the operator's expert knowledge is used for prediction.
Although the fuzzy-logic-based forecasting shows promising results, the process to
construct a fuzzy logic system is subjective and depends on somewhat heuristic
processes. The choices of membership functions and rule base have to be developed
heuristically for each scenario.

The rules fixed in this way may not always yield the

best forecast, and the choice of membership functions still depends on trial and error.

With these advantages and disadvantages of neural network and fuzzy logic, a
neuro-fuzzy framework has emerged by combining the learning ability of the neural
network and the functionality of the fuzzy expert system.

Its application can be

found in the work of Dash et al. (1995), Lie and Sharaf (1995), Studer and Masulli
(1997), and Padmakumari et al. (1999). Such a hybrid model is expected to provide
humanly understandable fuzzy meanings through the creation of more reliable
knowledge base through the learning ability of neural network.

Now, some researchers such as Jacobs and Levy (1989) have made the
interesting claim

that the the stock market is not an ordered system

4

that can be

explained by simple rules, nor is it

a totally random system for which no

predictions are possible. In fact, they claim that the market is a complex system, in
which only portions of the system's behavior could be explained and predicted by a
set of complex relationships among the variables.

Recognizing the complex characteristics of the stock market leads us to ask

if the predictability of the various indexes could be improved by using nonlinear
models with endogenous learning capabilities in a fuzzy real world environment. The
answer turns out to be yes, and

the specific modeling approach we use

demonstrates the advantages of the neuro-fuzzy technique. In particular, our modeling

of learning via the neuro fuzzy approach leads to better predictions in the case of
utilizing the KD technical indexes to describe the stock index movement.

The

purpose of this paper is to show this concretely through an investigation of the relative
profitability of this proposed KD based neuro-fuzzy trading system.

Specifically, then, the major contributions of this study are (1) to demonstrate
and verify the predictability of stock index return by applying neuro fuzzy technique
to KD index estimates; (2) to compare the performance of linear and nonlinear models
based on KD indexes; (3) to show the robustness of this proposed KD based neurofuzzy model; (4) and to show the existence of market opportunities for further
profitability via results from this proposed model and its profitability consistency over
5

time.

The remainder of this paper is organized as follows.

work about the technical analysis is reviewed.

In section 2, the past

Section 3 describes how the KD

based neuro fuzzy trading system is constructed and how the alternative benchmark
models are formulated.

The empirical results are shown in section 4.

Finally

section 5 provides some concluding remarks.

2. Literature Review

In general the approaches to predict stock price could be roughly categorized into two
kinds, fundamental analysis and technical analysis. Fundamental analysis is based
on macroeconomic data, such as exports and imports, money supply, interest rates,
inflationary rates (Fama and Schwert 1977, Campbell 1987, and Fama and French
1988a, 1988b), foreign exchange rates, unemployment figures, and the basic financial
status of companies such as dividend yields, earnings yield, cash flow yield, book to
market ratio, price-earings ratio, lagged returns, and size. (Basu, 1977; Fama and
French, 1992; Lakonishok, Shleifer and Vishny, 1994).
Technical analysis is based on the rationale that history will repeat itself and
that the correlation between price and volume reveals market behavior. Prediction is
6

made by exploiting implications hidden in past trading activities, and by analyzing

patterns and trends shown in price and volume charts (Epps, 1975; Smirlock and
Starks, 1985; Rogalski, 1978; Bohan, 1981; and Brush, 1986).
Basically the test of weak form efficient market is to test whether there exists
excess return by using technical analysis.

There have been some researches

claiming the existence of the weak form of efficient market (Fama 1965; Fama and
Blume 1966; Jensen 1967). Also there exist some researches claiming that the weak
form efficient market does not exist (Sweeney 1986; Brock, Lakonishok and LeBaron
1992; Bessembinder and Chan 1995). So far the research remains inconclusive.
One of the most commonly used methods in technical analysis is the moving
average filter rule.

The criterion is that the buying signal happens when the short

term moving average line breaks through the long term moving average line from
down, and the selling signal happens when the short term moving average line breaks
through the long term moving average line from up. The logic behind this rule is to
identify periods when expected returns deviate from unconditional means
(Bessembinder and Chan 1995).

Although Fama and Blume (1966) and Jensen and

Benington (1970) concluded that the filter rules are not useful, Brock et al. (1992) and
Sweeney (1986) showed the non-trivial ability to predict the price changes by using
the filter rules. On the other hand, Bailey et al (1990) and Pan et al. (1991) present

evidence that prices in some stock markets exhibit substantial deviations from random
walk behavior.
The reasons for the different empirical results can come from the different
samples, different technical indexes, or different rules. However, in this paper we
consider that there exists some relationship between the technical indexes and stock
price, the question is that how to use the information to explore this relationship.

In

other words, the specification of the function form is a difficult problem. A data
driven method to construct a model can be an effective way.
Similar to the filter rules, KD technical rules proposed by Lane (1957), is
trying to capture the period when expected returns deviate from unconditional means
by using K and D indexes instead of the moving averages. Essentially K and D
indexes with the advantages of momentum, relative strength, and moving average,
and with the consideration of the highest and the lowest prices, are expected to be
capable of capturing the short-term variance.
too simple to be effective.

However, the KD filter rules could be

Besides, the parameters for the rules are arbitrary.

Therefore, this paper is trying to develop a model based on the knowledge contained
in KD technical rules by using neuro-fuzzy.

Investment performance is simulated

based on the signals produced by the system.

Thirty world wide known stock

indexes are used as the testing sample. The standard regression model, GARCH-M,

and neural network are used to derive comparative results on relative prediction
performances.

3. Methodology
3.1 KD Trading System1
The commonly used K D indicators are calculated as follows.
RSVt = (Ct-L9) * 100/(H9-L9)

(1)

Kt

(2)

Dt

= 1/3 * Kt

1/3 * RSVt + 2/3 * Kt-1

+ 2/3 * Dt-1

(3)

where RSVt is the raw stochastic value for period t, Ct is the closing price for
day t, H9 and L9 are the highest price and the lowest prices for the latest nine days
respectively, K t and Dt are the values for K and D on day t.
available, 50 are used as the initial values for both in general.

If K and D are not

The trading rules for

the KD trading system are as follows.

Rule 1.

If D is greater than 80 and K breaks through D from up then sell out.

KD indicators were originally developed by Lange(

); but they have become popular only in

recent years. Elder(1987) points out:

The logic of this index is based on the observation that, as prices rise, daily closes tend to occur nearer
the high end of their recent range. When prices trend higher or are flat but the daily closes begin to sag
lower within that range, they signal internal market weakness and its readiness for a trend reversal to
the downside. The opposite occurs in down trends; They are confirmed when the closing prices are
near the bottom of the recent range. When closing prices move higher within a range, they show
internal strength.

Rule 2.

If D is less than 20 and K breaks through D from down then buy in.

Rule 3. When the slope of K is flat, the market trend is likely to change.
Rule 4. When the stock price reaches the new highest (lowest), K and D is not
reaching the new highest (lowest), the market trend is likely to change.
This is a so- called an expert system. The parameters, 80 and 20, are just the rule of
thumb values. The obvious question that one can ask is: can this expert system
beat the market?

In addition to implementing the original KD expert system, we

also try to fit different models that are commonly used in the literature, namely,
regression, ARCH_M, neural network, and neuro fuzzy, to describe the stock index
movement by considering it as a pattern recognition problem.
To capture the spirit of the KD trading system, we need to choose the
appropriate variables to describe the above KD rules.

The cross- over phenomenon

when K breaks through D from up is depicted in Figure 2.

Let K and D represent

the level of K and D, K_D the difference between K and D, and K_D_1 the K_D of
the previous day.

If the cross over phenomenon happened, as depicted in Figure 2,

then K_D_1 would be greater than 0 and K_D would be less than 0. Similarly, when
K breaks through D from down, then K_D_1 would be less than 0 and K_D would be
greater than 0. Therefore, we use K_D_1 and K_D to describe the cross over

See also Murphy(1986).

10

phenomenon.

In other words, K, D, K_D_1 and K_D are used to capture the

relationship described in rules one and two.

Let KS represent the slope of K, and KT, DT, and PT indicate the trend of K,
D, and P respectively.

We use KS to capture the relationship describe in rule 3, and

KT, DT, and PT to capture the relationship described in rule 4.

Let Pt be the stock

price for day t, and trendt denote the rate of return of day t. Then trendt is calculated
as ( Pt Pt 1 ) / Pt 1 . Totally we have 7 input variables, K, D, K_D, K_D_1, KS, KT,
and DT, and one output variable, TREND, to describe the KD system.

The

independent variables for predicting the index returns are all observable on or before
the last day of the day preceding the day to be forecasted.

In other words, only

observable, but not future, data are used as inputs to the forecasting models.

3.2 The construction of a Fuzzy Logic system

To facilitate the exposition, we only explain the model that describes the first two
rules, the crossover phenomenon. The complete system is constructed according to
exacly the same logic.

A fundamental idea of the fuzzy system is that we no

longer say, for instance that IF K is greater than 80.

Instead, we will describe the

value of K, for instance, to be very_low, low, medium_low, medium, medium_high,

high, very_high.

In other words, all the input and output variables will be translated
11

into fuzzy ordinary linguistic terms.

Table 1 summarizes the variables and their

linguistic terms. K, D, and Trend are described by 7, 7, and 5 terms respectively.

K_D, and K_D_1 are both described by 2 terms. Each term is described by a
membership function.

A membership function, expressed as u A (x) , describes the

extent to which an object x belongs to a fuzzy set (term) A.

There are many different kinds of membership functions. Popular ones are Z, S,

and (Von Altrock 1997: p. nos?). In our case, we used Z, S, and for our
experiments.2

Fig. 1(a), 1(b), 1(c), 1(d), 1(e) shows the membership functions for K,

D, K_D, K_D_1, and Trend.

Consider the case where the value of K is 80, D is 60, K_D is 0.6 and K_D_1
is 0.4.

It can be found in Figure 1(a) that the degree of K being high is 0.6 and the

degree of K being very_high is 0.4. Besides, the degrees of K for other linguistics
terms are all 0. The membership function for K equal to 80 can be expressed as
follows.
uvery _ low (80) = 0.0 ,
umedium _ high (80) = 0.0 ,

ulow (80) = 0.0 , umedium _ low (80) = 0.0 , umedium (80) = 0.0 ,
uhigh (80) = 0.6 ,

uvery _ high (80) = 0.4

Similar to K, the values of the linguistic terms for the other variables can be
found from Fig. 1(b), 1(c), and 1(d) as follows.

The Z, S,lambda and pi functional forms were all tried in order to choose the most appropriate one.It

12

For variable D:
uvery _ low (80) = 0.0 , ulow (60) = 0.0 ,

umedium _ low (60) = 0.0 , umedium (60) = 0.2 ,

umedium _ high (60) = 0.8 , uhigh (60) = 0.0 , uvery _ high (60) = 0.0
For variable K_D:
unegative (1) = 0.0 , u positive (1) = 1.0
For variable K_D_1:
unegative (1) = 1.0 , u positive (1) = 0.0
After the numeric values have been translated into linguistic values, a much
more sophisticated rule, for example, can be obtained as follows:
IF K is high, D is medium, K_D is positive and K_D_1 is negative, then
Trend is high_inc.

(1)

This is so called an inference rule. Each rule consists of two parts, IF part
and THEN part, describing the extent to which

the real object satisfies the

condition and the response of this system respectively.

The operator proposed by

Zimmermann and Thole (1978) to represent logical connectives and is the

minimum value among all the validity values.
is summarized in table 2.

The validity of each term for rule (1)

Therefore, the validity value of the IF part is equal to

min{0.6, 0.2, 1.0, 1.0} = 0.2, which also indicates the degree of validity for the

turns out that the last three are the most appropriate. (Why?)
13

THEN part.

In other words, the validity extent of the systems response TREND

is high_inc is 0.2.
Let us assume that we have, say five, inference rules. Using these five
inference rules, we obtain the following inferences:

(1) The validity extent of TREND is high_dec is 0.0.

(2) The validity extent of TREND is small_dec is 0.3
(3) The validity extent of TREND is steady is 0.0
(4) The validity extent of TREND is small_inc is 0.2
(5) The validity extent of TREND is high_inc is 0.3

Note that there are the following five fuzzy set membership functions:
high_dec, small_dec, steady, small_inc, and high_inc. To facilitate discussion, let us
denote high-dec, small-dec, steady, small_inc, and high_inc by f 1 , f 2 , f 3 , f 4 , and f 5
respectively.

For each membership function f i , let M i denote the value of TREND

which achieves the maximum value of f i .

choose the medium of the interval.

If the values are within an interval, we

For instance, by consulting Fig. 1(e), we have

the following mapping.

M1 = -0.83

14

M2 = -0.33
M3 = 0.0
M4 = 0.33
M5 = 0.83

Let the validity extent of TREND belongs to f i be denoted as U i , then the

value of TREND will be determined by the following formula:
5

TREND= U i M i
i =1

Let us assume that U i ' s be 0.0, 0.3, 0.0, 0.2, and 0.3. We will have
TREND = 0.0 (0.83) + 0.3 (0.33) + 0.0 (0.0) + 0.2 0.33 + 0.3 (0.83) = 0.20 .
This means that predicted trend is 0.20 for the next day.
Buying signals are recognized when the predicted trend is greater than a
predetermined threshold value, and selling signals are recognized when it is less than
another predetermined threshold value. Usually both threshold values are set equal
to 0.

Stocks are bought in when signal is greater than 0, and the stocks are held until

the trend is less than 0.

Buying signal is ignored when there are stocks on hold, and

selling signal is ignored when there are no stocks on hold.

In this paper, short sell strategy is not considered. Obviously, the buy and

sell decisions become more complicated whenshorting is allowed.

15

Note, for example, that using the rule IF K is high, D is medium, K_D is
positive and K_D_1 is negative, then Trend is high_inc., we will obtain the validity
extent of TREND is high-inc as 0.2. However, there are 7 terms for K and D, 2
terms for K-D and K-D-1 and 5 terms for Trend.
start with.

Therefore, we have two problems:

the inference rules which are not practical?

We have 7*7*2*2*5=980 rules to

(1)How can we eliminate some of

(2)How can we use the remaining

inference rules to obtain a precise value of TREND?

Among all of these 980 rules, some of them are not valid in practical sense.
For instance, the following rule obviously makes no sense:
IF D is very_high, K is very_high, K_D is positive and K_D_1 is positive,
then Trend is high_inc.

(2)

Such a rule must be eliminated. Besides, how can we determine the

membership functionthat is appropriate?

The training method of neural network can

be used to solve both these problems, i.e., to refine the membership functions and to
eliminate the irrelevant inference rules.

3.3 Going to a Neuro Fuzzy Formulation

Basically the idea of a neuro-fuzzy system is to find the parameters of a fuzzy
system by means of learning methods obtained from neural networks. Many

16

alternative ways of integrating neural nets and fuzzy logic have been proposed
(Buckley and Hayashi 1994, Nauck and Kruse 1996, Lin and Lee 1996) which have
much in common, but different in implementation aspects.

The most common

approach is to use so-called Fuzzy Associative Memories (FAMs). A FAM is a

fuzzy logic rule with an associated weight. A mathematical framework exists that
maps FAMs to neurons in a neural net.4

This enables the use of a modified error

back propagation algorithm with fuzzy logic.5

This approach can help to generate

and optimize membership functions and the associated weight of each rule from
sample data.

In our experiments, we implemented the FAM approach to construct

the model.6

3.4 Benchmark Models

A linear regression model is based on constructing a linear relationship between
dependent and independent variables. GARCH-M is also a linear model, but with
the additional nonlinear consideration about the residual variance. Neural network is

See for example, Eric B. Baum (1988), Neural Nets for Economists and the references therein as

well as the references in von Altrock(1996).

5

For more details on the mathematical foundation and relevant derivations, refer to (Kosko 1992)
6

Please refer to Tong and Bonissone (1984), Zimmermann (1987), and Klir

and Yuan (1995) for the details for the implementation of fuzzy logic, and refer to
(Von Altrock 1997) for the details for the implementation of neuro fuzzy.

17

a model mainly to map the nonlinear relationship among the variables that allows for
endogenous learning. Using the terminology introduced before, we can characterize a
neuro fuzzy system as an expert system with different weights associated with each
rule where the fuzziness of ordinary language is also modelled explicitly.
With the same input and output variables, these models are specified as
follows.

The linear regression:

Trend t +1 = 0 + 1 K _ D + 2 K _ D _ 1 + 3 K + 4 D + 5 KS + 6 KT
+ 7 PT + 8 DT + t
where t ~ ID(0, t2 ) , this is the benchmark model for this research.

(1)
The buying

signal is recognized when Trend > 0.

The GARCH-M (1,1):

Trend t +1 = 0 + 1 K _ D + 2 K _ D _ 1 + 3 K + 4 D + 5 KS + 6 KT
1

(2)

+ 7 PT + 8 DT + ht 2 + t
where

t ~ N0, htand h t = 0 + 1ht 1 + 2 t21 . .

The neural network model is:

Trend t +1 = F2 ( w2 F1 ( w1 x))

(3)

where F1 and F2 are the transfer functions for hidden node and output node,
respectively.

The most popular choice for F1 and F2 are the sigmoid function,

18

F ( x) =

1
1 + e x , representing the activation function adopted in the calculation

process,

w1 and w2 are the matrices of linking weights from input to hidden layer

and from hidden to output layer, respectively, x is the vector of input variables, K, D,
K_D, K_D_1, KS, KT, and DT. Basically a three-layer MLP is implemented in this
paper with learning rate equal to 0.1 and momentum equal to 0.7. For details of this
procedure and its mathematical background please refer to (Azoff 1994, Beltratti et al.
1996).
The purpose of neural network training is to estimate the weight matrices, w1
and w2 , in equation (3) such that an overall error measure such as the mean squared
errors (MSE) or sum of squared errors (SSE) is minimized.
MSE =

1
(a j TREND j ) 2
N

MSE can be defined as

(4)

where aj and TREND j represent the target value and network output for the j th day
respectively, and N is the number of days in training data set.

4. Empirical Results
4.1 Sample
The data set was obtained from Taiwan Economic Journal Data Bank (TEJ). Thirty
world wide known stock indexes, as listed in table 1, are used for

19

testing the

predictive powers

of alternative models.

1992/1/1 to 2000/9/30.

The data series include the period from

It is divided into two periods.

The first period (in-sample

data set) runs from 1992/1/1 to 1999/9/30 (1600 observations), while the second
period (out-sample data set) runs from 1999/10/1 to 2000/9/30 (350 observations).
The in-sample data set is used to determine the parameters of the models and the
outsample data set is used for model validation.

4.2 Transaction Simulation

Essentially, the investment simulation

implemented here is

based on the signal

produced by these systems.It is asummed that if the predicted value, TREND, turned
out to be higher than the threshold value, a portfolio of stocks interlocked with the
stock index was purchased; if the predicted value was lower than the threshold value,
the portfolio was sold. This would seem to reflect the rationality embodied in the
profit making activities in the stock market adequately.

4.3 Performance Evaluation

In this paper we use RMSE, direction prediction, and rate of return to compare the
performance of these different models. RMSE is calculated as follows.
n

(a TREND )
i =1

(4)

20

where n is the number of days in testing data set.

If the predicted value is greater than the threshold value and the actual stock
price movement for the next day is up too, then it is counted as one time
prediction.

correct

Direction prediction percentage (hit rate) is calculated as follows:

Hit_rate =

h
,
n

(5)

where h is the number of correct prediction.

The rate of return for each strategy is calculated as follows:
1

R = ((1 + r1 ) (1 + r2 ) L (1 + rn )) n n 1

(6)

where ri is the daily return for day i . The daily rate of return for cash on deposit
is 0.05/250=0.0002 (i.e. 5% for yearly rate of return).

We do the unit root test before constructing the GARCH-M models. The unit root
hypothesis

is rejected at significance level 0.05 for all thirty series by the

Augmented Dickey Fuller (ADF) test.

stationary.
model.

In other words, these thirty series are all

No differencing of any series is therefore necessary when fitting the

The residuals of the GARH-M model are all white noise, as is readily found

by checking the ACF and PACF.7

On the other hand, many of the assumptions for the linear regression model( for
example, the normality assumption, the constant variance assumption, and the
assumption of non existence of autocorrelaton )

21

4.4 Empirical Results

In order to fix notation for facilitating the presentation, we use BH, KD, REG, GM,
NN, and NF to denote buy and hold strategy, traditional KD trading system,
regression model, Garch_M model, neural network model, and neuro fuzzy model
respectively.

Based on the predictions of each model on these thirty stock indexes,

the paired test of the RMSE is listed in table 3.

The number in the cell shows the

difference of RMSE of the model at the row and the RMSE of the model at the
column. For example, 0.0004 is the difference between GARCH-M and regression.
The number in parenthesis is the p-value of the paired test.

It can be seen that neuro

fuzzy has the biggest RMSE among the methods.

In addition to the RMSE, we also show the direction prediction.

Table 4

lists the basic statistics for the correct prediction percentage of all four models for
thirty stock indexes.

It can be seen that neuro fuzzy has 58.03% correct prediction

for the next days stock movement direction during the test period, neural network
55.77%, GARCH-M 52.83%, and regression 52.47%.

Table 5 lists the paired test

among these models. The number in the table represents the difference between the
correct prediction percentage of model at the row position and the model at the

are not satisfied for most of the series.

However, we still fit the regression models

for each series as our benchmark.

22

column position. It can be seen that neuro fuzzy has the highest hit rate among all
these models.
Based on the signals produced by each model, transactions are implemented
for each model and the corresponding rate of return are calculated.

Table 6 lists the

basic statistics of the yearly rate of return of each model. Neuro fuzzy can achieve
yearly rate of return as 27.17%, neural network 19.47%, GARCH-M 12.2%,
regression 9.84%, traditional KD 9.56%, and buy and hold 9.35% respectively.
Table 7 shows the paired test among these methods.

It can be seen that the rate of

return of neuro fuzzy is significantly greater than those of the other methods.
In addition to the statistical test of the yearly rate of return, Figure 3 also
shows the cumulative wealth for each model for Landon FT 100 Index, which is
typical for the other indexes. Neuro fuzzy is significantly the best one among these
models.
It is interesting to note that neuro fuzzy is the most profitable model though its
RMSE is not the least. This result is similar to Leung, Daouk, and Chens work
(2000).

It implies that the financial forecasters and traders could focus on accurately

predicting the direction instead of minimizing the MSE or RMSE.

Since the threshold values of buying and selling signals cane influence the rate

23

of return, we do the sensitivity analysis on the trading points as follows.

We divide

the signal range from 0 to 1 into 20 points with the interval equal to 0.05 as the
alternative buying threshold values.

Therefore we have 21 alternative threshold

values, 0, 0.05, 0.10, 0.15, K , 1.00. And we do the same processing for the selling
signal range from 1 to 0 (-1.0, -0.95, -0.90, K ,0). Totally we have 441 (21*21=441)
combinations.

With the produced signal from the proposed model, we use each

combination as the threshold values. For example, one alternative is (0.05,-0.15),

which means that if the signal is greater than 0.05, then buy in the stocks.
signal is less than 0.15, then sell out the stocks.
transactions.

If the

Otherwise, do not do any

Therefore, we can have a rate of return associated with each

combination.
Since the profitability of a threshold value combination in the training data set
does not promise the profitability in the testing data set, we need to show the
robustness of the trading points. We calculate the rate of return for each threshold
value combination on training data set and testing data set.

The empirical results

show that the average is close to 75 percent that if the threshold value combination is
profitable in training data set, it will also be profitable in the testing data set.

The

detailed simulation result is shown in table 8. There are 8 out of 30 series with
profitable percentage within 80% to 100%, 18 within 60% to 80%, 3 within 40% to

24

60%, and only 1 within 20% to 40%, showing the robustness of the proposed model
for the parameters.
In addition to the robustness testing, a sensitivity analysis of the influence of
transaction cost on the profitability of each model is also conducted.

Transaction

costs consist of commission fee 0.13% and trading tax 0.3% in Taiwan.

The

sensitivity analysis is run by setting trading tax equal to 0.0%, 0.1%, 0.3%, 0.5%, and
1.0% respectively.

The simulation results are shown in Table 9.

It can be seen that

neuro fuzzy can consistently beat the traditional KD strategy and buy and hold
strategy and the rate of return is decreasing only a little bit as the transaction costs
increases.
Besides, the testing data set is arbitrarily divided into two different scenarios,
bull market and bear market (or sluggish market) to see the influence of the scenario
on the profitability performance.

A bull market is defined as the period before the

highest point of the testing period.

A bear market or a sluggish market is defined as

the period after the highest point during the testing period. Since some series have
no turning points during the testing period, therefore the sample size for the different
scenarios testing will be different.

The basic statistics of the simulation results are

shown in table 10. Table 11 and 12 show the paired test among the methods when
the market is a bear market and a bull market respectively.

25

It can be seen from table

11 that neuro fuzzy could significantly beat buy and hold strategy and traditional KD
when the market is a bear market.

Table 12 shows that neuro fuzzy is significantly

better than the traditional KD but not significantly better than buy and hold when the
market is a bull market.8

5. Summary and Conclusions

This paper uses 30 well known stock indexes to examine the linear and nonlinear
predictability of stock market returns with KD technical trading rules.

The

empirical results show strong evidence of nonlinear relationships among the key
variables

in the stock market. Empirically, this is demonstrated most clearly by the

nonlinear neuro fuzzy model that was used along with several others to predict returns
by using KD technical indexes. The rate of return of the proposed neuro fuzzy
system is significantly greater than that of the other methods.

In addition to the

robustness shown by the trading point analysis, the sensitivity analysis of transaction

A possible explanation for this is that for the extreme case, that is the case with no
turning points at all for the testing period, buy and hold will be the worst for the bear
market and be the best for the bull market. However, when there are many turning
points during the testing period, it is a different matter. Generally, the more
turning points there are, the better the Neuro Fuzzy mode will be in prediction
performance. For this paper, NF is better than buy and hold but not significantly so
during the testing period.

26

costs also shows the profitability of the proposed system.

Finally, the scenario

analysis shows that though the rate of return of neuro fuzzy system is not significantly
greater than that of buy and hold strategy when the market is a bull market, it is

the

best in a statistical sense when the market is a bear market or a sluggish market. This
conclusion is important for both theory and strategy. Theoretically it shows that the
efficient market hypothesis need not hold in the short run, but with learning the
possibility of a convergence to the long run efficient market equilibrium can not be
ruled out. Strategically, our approach shows that the neuro fuzzy model may allow
investors to earn higher returns when there is a bear market which is far from the
effient market equilibrium.

27

References
1. Al-Shammari, M. and Shaout, A., Fuzzy logic modeling for performance
appraisal systems, A framework for empirical evaluation, Expert Systems with
Applications, 14, (1998), 323-238.
2. Azoff, E. M., Neural network time series forecasting of financial markets, 1994,
John Wiley & Sons Ltd.
3. Bailey, W., Stulz, R., and Yen, S., 1990, Properties of daily stock returns from the
Pacific Basin stock markets: Evidence and implications, in: Rhee, S. and Chang,
R., eds., Pacific-Basin capital markets research (North-Holland, Amsterdam).
4. Basu, S. , The investment performance of common stocks in relation to their
price -earnings ratios: A test of the efficient market hypothesis, Journal of
Finance 32,( 1977) , 663-682.
5. Baum,Eric B.

(1988), Neural Nets for Economists, in Anderson,P., Arrow, K.

and Pines, D., eds. The Economy as an Evolving Complex System, New york:
Addison-Wesley.
6. Beltratti, A., Margarita, S., and Terna, P., Neural networks for Economic and
Financial Modelling, 1996, Internal Thomson Publishing Inc.
7. Bessembinder, H., and Chan K., The profitability of technical trading rules in the
Asian stock markets, Pacific-Basin Finance Journal, 3, 1995, 257-284.
8. Bollerslev, T., Generalized Autoregressive Conditional Heteroskedasticity ,
Journal of Econometrics, 31(1986) , 307-327.
9. Bolloju, N. , Formulation of qualitative models using fuzzy logic, Decision
Support Systems, 17(1996) , 275-298.
10. Box, G. and Jenkins, G., Time Series Analysis: Forecasting and Control,
Holden-Day, San Francisco, CA, 1970.

28

11. Brock, W., Lakonishok, J., and LeBaron B .,Simple Technical Trading Rules
and the Stochastic Properties of Stock Returns, Journal of Finance , Vol.47(1992),
1731-1764.
12. Buckley, J. J. and Hayashi, Y. (1994) Fuzzy Neural Networks: A Survey,
Fuzzy Sets and Systems, 66, 1-13.
13. Campbell, J. (1987). Stock returns and the term structure, Journal of Financial
Economics 18, 373-399.
14. Dash, P.K., Liew, A. C., Rahman, S., and Dash, S.,Fuzzy and Neuro-fuzzy
Computing Models for Electric Load Forecasting, Engng Applic. Artif. Intell.
Vol. 8, No. 4, 1995, 423-433.
15. Engle, R. F., Autoregressive Conditional Heteroskedasticity with Estimates of
the Variance of U.K. Inflation, Econometrica, 50(1982), 987-1008.
16. Fama, E. F., The Behavior of Stock Market Prices, Journal of Business , Vol.33,
January(1965),34-105.
17. Fama, E. F. and Blume,M.E., Filter Rules and Stock Market Trading, Journal
of Business, Vol.39, January(1966), 226-241.
18. Fama, E., & French, K. Permanent and temporary components of stock prices,
Journal of Political Economy 96, 1988a, 246-273.
19. Fama, E., & French, K. Dividend yields and expected stock returns, Journal of
Financial Economics, 1988b, 22, 3-25.
20. Fama, E., and French,K., The cross-section of expected stock returns, Journal
of Finance ,vol.47(1992), 427-465.
21. Fama, E., & Schwert, W., Asset returns and inflation, Journal of Financial
Economics, 1977, 5, 115-146.
22. Jensen, M.C., Random Walks : Reality or MythComment, Financial Analysts
Journal,vol.23(1967),77-85.
29

23. Jensen, M.C. and Benington, G.A. , Random Walks and Technical Theories :
Some Additional Evidence , Journal of Finance , vol.25(1970),49-54.
24. Jocobs, B., and K. Levy. ,The complexity of the Stock Market, The Journal of
portfolio Management ,16 (1989),69-76.
25. Kaneko, Takaomi Building a financial diagnosis system based on fuzzy logic
production system, Computers ind.Engng, Vol.31,No.3/4(1996), 743-746.
26. Klir,G.J., and Yuan , B. , Fuzzy sets and fuzzy logic:Theory and Applications,
Upper Saddle River ,NJ:Prentice Hall , 1995.
27. Kosko, B., Neural Networks and Fuzzy Systems. A Dynamical Systems
Approach to Machine Intelligence. NJ :Prentice Hall, Englewood Cliffs, ,1992.
28. Lakonishok, J.,Shleifer, A.,and Viahmy , R. W. Contrarian investment ,
extrapolation, and risk, Journal of finance, 49(1994),1541-1578.
29. Lane, G.,

30. Leung, M. T., Daouk, H., and Chen, A-S., Forecasting stock indices: a
comparison of classification and level estimation models, International Journal of
Forecasting, 16, 2000, 173-190.
31. Lie, T. T. and Sharaf, A. M., A novel neuro-fuzzy based self-correcting online
electric load forecasting model, Electric Power Systems Research, 34, (1995),
121-125.
32. Lin, C.-T and Lee, C. G. (1996) Neural Fuzzy Systems. A Neuro-Fuzzy
Synergism to Intelligent Systems.

Prentice Hall, New York.

33. Nauck, D. and Kruse, R. (1996) Neuro-Fuzzy Systems Research and

Applications outside of Japan, Fuzzy-Neural Networks, Soft Computing series,
pp. 108-134. Asakura Publ., Tokyo.
34. Padmakumari, K., Mohandas, K.P., and Thiruvengadam, S., Long term
distribution demand forecasting using neuro fuzzy computations, Electrical
30

Power and Energy Systems, 21, (1999), 315-322.

35. Pan, M., Chiou, J., Hocking, R. and Rim, H., 1991, An examination of
mean-reverting behavior of stock prices in Pacific-Basin stock markets, in: Rhee,
S. and Chang, R., eds., Pacific-Basin capital markets research, Vol. II
(North-Holland, Amsterdam).
36. Studer, L. and Masulli, F., Building a neuro-fuzzy system to efficiently forecast
chaotic time series, Nuclear Instruments and Methods in Physics Research, 389,
(1997), 264-267.
37. Sweeny, R.J., Beating The Foreign Exchange Market, Journal of Finance ,Vol.
41(1986), 163-182.
38. Tong, R. M. and Bonissone, P.P.Linguistic Solutions to Fuzzy Decision
problems, in Zimmermann, H.-J., Zadeh, L. A., and Gaines, B. R. (Eds.), Fuzzy
Sets and Decision Analysis, Amsterdam (1984), p. 323-334.
39. von Altrock , C., Fuzzy Logic & NeuroFuzzy Applications In Business & Finance,
Upper Saddle River, N.J: Prentice Hall PTR. 1997.
40. Zimmermann, H.-J., Fuzzy Sets, Decision Making, and Expert Systems, Boston:
Kluver Academic Publisher (1987). ISBN 089838-149-5.
41. Zimmerman, H.-J. and Thole, U. On the Suitability of Minimum and Product
Operators for the Intersection of Fuzzy Sets, Fuzzy Sets and Systems, Vol. 2
(1978) , 173-186.

31

Figure 1(a). Membership function of K

Figure 1(b). Membership function of D

32

Figure 1(c). Membership function of K_D

Figure 1(d). Membership function of K_D_1

33

Figure 1(e). Membership function of Trend

34

K
K_D_1 > 0
D

K_D < 0

Figure 2. The cross over phenomenon

35

1400000

NF
NN
garch
reg
KD
BH

1200000

1000000

800000
1999/10/1

2000/1/1

2000/4/1

2000/7/1

Figure 3. The equity curve for each model for Landon FT100 Index

36

Table 1. Properties of linguistic variables and their terms

Variable Name

Type

Minimum

Maximum

Input

100

very_low, low, medium_low,

medium_high, high, very_high

medium,

Input

100

very_low, low, medium_low,

medium_high, high, very_high

medium,

K_D

Input

-1

negative, positive

K_D_1

Input

-1

negative, positive

Trend

Output

-1

high_dec,
high_inc

37

Term Names

small_dec,

steady,

small_inc,

Table 2. The corresponding validity extent of each term for rule (1).
Variable

Values

Membership

Validity

function
K

80

K is high

0.6

60

D is medium

0.2

K_D

0.6

K_D is positive

1.0

K_D_1

-0.4

K_D_1 is negative

1.0

38

Table 3. Paired test of RMSE

REG

GM

NN

-0.021
(0.006) *
0.032
(0.001) *

0.0526
(0.000) *

REG
GM
NN
NF

0.0004
(0.201)
-0.017
(0.049) *
0.036
(0.002) *

*:0.05

39

NF

Table 4. Basic statistics for the correct prediction percentage.

REG
GM
NN
NF

Minimum

Maximum

Average

30
30
30
30

.45
.48
.49
.52

.58
.58
.64
.66

.5247
.5283
.5577
.5803

40

Standard
Deviation
0.024
0.029
0.0035
0.0038

Table 5. Paired test for the correct prediction percentage.

REG
GM
NN
REG
GM
NN
NF

0.004
(0.423)
0.037
(0.000) *
0.056
(0.000) *

0.029
(0.001) *
0.052
(0.000) *

*:0.05

41

0.023
(0.000) *

NF

Table 6. Basic statistics of the yearly rate of return

BH
KD
REG
GM
NN
NF

Minimum

Maximum

Average

30
30
30
30
30
30

-0.32
-0.27
-0.31
-0.22
-0.05
-0.05

0.50
0.38
0.49
0.50
0.53
0.73

0.093
0.096
0.098
0.122
0.194
0.271

42

Standard
Deviation
0.239
0.152
0.174
0.165
0.159
0.200

Table 7. Paired tests of the yearly rate of return

BH

KD

REG

GM

NN

BH
KD

0.002
0.935

REG

0.004
0.002
0.868 0.905

GM

0.028
0.026
0.023
0.409 0.365 0.096

NN

0.101
(0.001) *

NF

0.1782
0.1761
0.1733
0.1498
0.00* 0.00* 0.00* 0.00*

0.099
0.096
0.072
0.00* 0.00* 0.00*

*:0.05

43

0.077
(0.001) *

NF

Table 8 The percentage of the trading point combinations that is also profitable in
the testing data
Percentage

No. of series

80%~100%

60%~80%

18

40%~60%

20%~40%

44

Table 9. Rate of return under different transaction costs

Investment strategy
Transaction cost
NF
BH
0.00
27.17
9.56
0.10
24.87
8.33
0.30
22.65
8.05
0.50
20.53
7.22
1.00
18.28
5.98

45

KD
9.35
9.35
9.32
9.32
9.31

Table 10. Basic statistics of the rate of return under different market scenario
Bear market

BH
KD
NF

Minimum

Maximum

Average

30
30
30

-0.40
-0.30
-0.18

0.27
0.12
0.21

-0.170
-0.073
-0.002

Minimum

Maximum

Average

30
30
30

0.15
-0.01
0.09

0.73
0.38
0.86

0.329
0.156
0.340

Standard
Deviation
0.174
0.107
0.088

Bull market

BH
KD
NF

46

Standard
Deviation
0.149
0.087
0.189

Table 11. Paired test of the yearly rate of return when the market is a bear market
Paired difference

NF - KD
NF - BH
KD - BH

mean

Standard
deviation

Degree of
freedom

0.0710
0.1673
0.0962

0.1008
0.1240
0.1300

3.307
6.329
3.474

21
21
21

*:0.05

47

significance
(two-tailed)
0.003*
0.000*
0.002*

Table 12. Paired test of the yearly rate of return when the market is a bull market
Paired difference

NF - KD
NF - BH
KD - BH

mean

Standard
deviaiton

Degree of
freedom

0.1848
0.0111
-0.1736

0.1551
0.0959
0.1205

5.955
0.582
-7.202

24
24
24

*:0.05

48

significance
(two-tailed)
0.000*
0.566
0.000*