Utilization of Soft Computing Techniques for Constructing Reliable Decision Support

Systems for Dealing Stocks

Norio Baba a), Naoyuki Inoue a), Yan Yanjunb)

a) Information Science, Osaka Kyoiku University, Asahiga-Oka,4-698-1,
Kashiwara City, Osaka Prefecture, 582-8582, Japan.
b) Information Science, Osaka Kyoiku University, Asahiga-Oka,4-698- 1,
Kashiwara City, Osaka Prefecture, 582-8582, Japan. (One year and six months
leave from Harbin Institute of Technology)
e-mail: [email protected]

Abstract: The use of soft computing techniques such system for dealing in the TOPIX (Tokyo Stock Exchange
as NNs, GAS, etc. in the financial market has recently Prices Indexes), which utilizes neural networks and
become one of the most exciting and promising genetic algorithms. In the proposed system, neural
application areas. In this paper, we propose a decision networks are utilized in order to make forecasts of the
support system (DSS) for dealing in the TOPIX highest and lowest prices of TOPIX four weeks in the
(Tokyo Stock Exchange Prices Indexes), which utilizes future. We shall propose a (prototype) decision support
neural networks and genetic algorithms. In the system (DSS) for dealing in the TOPIX which utilizes
proposed system, the neural network is utilized in these two forecasts. Next, genetic algorithms are utilized
order to make a forecast of the TOPIX four weeks in in order to find appropriate parameter values (included in
the future. Genetic algorithm is utilized in order to the (prototype) DSS) which would have a significant
find an effective way of dealing. Several computer effect upon the performance of the total decision support
simulations have been carried out in order to compare system. GAS are also utilized in order to find an
the proposed DSS with the other approaches such as intelligent way of dealing in the TOPIX.
the DSS using traditional technical analysis and The outline of this paper is as follows. In the second
buy-and-hold method. These simulations confirm the section, following a brief explanation concerning the
effectiveness of the proposed DSS. neural network model, input variables, etc., we will
propose the decision support system which utilizes ANNs
I. INTRODUCTION for dealing in the TOPIX. In the third section, we will
One fundamental aspect of artificial neural networks further try to utilize GAS in order to make the proposed
(ANNs) is their capability to build complex nonlinear system more effective. In the fourth section, we will
relationships between input variables and output variables compare the proposed DSS with other approaches such as
directly from the training data. This characteristic feature the DSS using traditional technical analysis and
of ANNs is especially suited to various applications. buy-and-hold method. Finally, the paper concludes with a
Regarding the field of financial engineering, a large summary.
number of researchers have had growing concerns about
applying ANNs 111-[12]. However, despite a large 11. APPLICATION OF ANNs FOR
number of trials for the utilization of ANNs to the field of CONSTRUCTING A DECISION SUPPORT
financial engineering having been done so far, only SYSTEM FOR DEALING IN THE TOPIX
limited success has been reported. In this section, we shall propose a decision support
Recently, other soft computing techniques such as system for dealing in the TOPIX which utilizes artificial
fuzzy engineering, genetic algorithms (GAS), etc. have neural networks. First of all, let us briefly touch upon the
been used in the field of financial engineering [7]- [8]. neural network model and input variables.
Unfortunately, like the trials with ANNs, they have The recent spread of the technologies utilizing the
produced only limited success. Intemet has revolutionized our daily lives. In this decade,
In order to produce a more reliable (intelligent) system, the basic structure of the industry has also been greatly
a number of researchers are trying to construct hybrid changed by the penetration of Intemet Technology.
systems which incorporate the powers of ANNs with
those of the other soft computing techniques mentioned A. Neural Network Model and Input Variables
[91- [ill. Fig.1 shows the neural network model which consists
In this paper, we shall propose a decision support of the eight input units, one hidden layer having fifteen

0-7803-7278-6/02/$10.00 02002 IEEE 2150

units, and two output units. As inputs into the neural
network model, we have chosen the following eight
variables which may significantly effect the changes of
the TOPIX:
1) Currency Rate (Yen - Dollar)
2) TOPIX
3) Changes of the Amount of the Buying Volume by
Foreign Traders
4) Changes of the Amount of the Selling Volume by
_i' I
Foreign Traders Fig.2a. Teacher signal concerning outputs from the upper
5 ) Changes of the TOPIX output unit
6) Turnover in the Tokyo Stock Market
7) Total Amount of Money Used for Dealing in the
Tokyo Stock Market
8) PBR in the Tokyo Stock Market
The reason we have considered the neural network
model with two output units is to predict the increase
rate of the highest price of the T O P E and the decrease
rate of the lowest price of TOPIX up to four weeks in
the future by using the output sequences from the upper
output unit and the lower output unit, respectively. Fig.2b. Teacher signal concerning outputs from the lower
h
output unit

Remark 2B1: For the output function of each neuron in

the neural network model, we have used the bipolar
sigmoid function.
Remark 2B2: In this simulation, we have used the hybrid
algorithm for training the parameters involved in the
neural network model. We could also use MATLAB for
training NNs. We are now using MATLAB in order to
Fig. 1. Neural network model with two output units construct a DSS to deal with NIKKEI-225.

B. Teacher Signals and the 'Ikaining Algorithm C. Prediction of the Highest (and Lowest) Price of TOPIX up to
Four Weeks in the Future
The neural network model (Fig,1) is utilized in order to
By utilizing the output from the trained neural network,
make predictions concerning the increase rate of the
one can make a prediction of the highest (lowest)
- price
- of
highest price of TOPIX and the decrease rate of the
lowest price of TOPIX up to four weeks in the future. Fig. T O P E up to four weeks in the future. Let 01( 0 2 ) be
2a and Fig. 2b explain how two teacher signals are an output from the upper (lower) output
-unit -of the trained
generated. neural network. Then, a prediction XI ( X2 ) concerning
Although the BP method and its modified versions the increase (decrease) rate of the highest (lowest) price
have fiequently been used for various pattern of TOPIX up to four weeks in the future
classification problems, they have several limitations. - - can -be - made by

One of the most important is that it often fails to find the the inverse transformation DI + X I ( 0 2 +X2), as
global minimum of the total error function of the neural
shown in Fig.3a (Fig.3b).
network and falls into a local minimum. The hybrid One can make a prediction of the highest (lowest) price
of TOPIX up to four weeks in the future (by using
algorithm [6] which combines the modified BP method - -
with the random optimization method ensures XI(Xz)) as follows: Let B be the current price of
convergence with probability 1 to a global minimum in a TOPIX. Then, one can obtain a prediction conceming the
compact region of weight vector space. In the proposed highest (lowest) price of TOPIX up to four weeks in the
system, this hybrid algorithm is utilized for training
weights in the neural network model.
future by calculating B( I+ x) (B( I+ )).

0-7803-72784/02/$10.00 02002 IEEE 2151

 -'I
Fig.3a. Prediction concerning the increase rate of the
highest price of TOPIX up to four weeks in the future

Fig.4. An illustration which explains dealings in the

DSS (a=P=O)
We have carried out computer simulations using DSS.
We have so far obtained numbers of successful computer
simulations, which confirm the effectiveness of the
proposed DSS. Due to limitation of space, we are not able
to show you details. Interested readers are kindly to read
the papers [ 113.
Fig.3b. Prediction concerning the decrease rate of the
lowest price of TOPIX up to four weeks in the future 111. UTILIZATION OF GAS FOR MAKING THE
DECISION SUPPORT SYSTEMS MORE
D. A Decision Support System for Dealing TOPIX EFFECTIVE
By utilizing the output from the upper output unit of By the large numbers of computer simulations, we have
the trained neural network, one can make the forecast of confirmed that the proposed decision support system DSS
the highest price of TOPIX up to four weeks in the future. is quite effective in yielding considerable return under the
On the other hand, one can make the forecast of the specific combination of parameter values a and 0.
lowest price of TOPIX up to four weeks in the future by However, in order to check the effectiveness of the DSS,
utilizing the output from the lower output unit. Below, let one has to investigate it regarding various numbers of
us propose a decision support system (DSS) which combinations. Further, the DSS does not specify how to
utilizes this information effectively for dealing in the deal in the TOPIX under the general condition that one
TOPIX: may use only limited funds. In the decision support
DSS system DSS proposed in the previous section;one has to
1. Let a and p be the positive parameters. sell all of the stocks which he had bought at a time when
2. Calculate the average of the prediction of the highest the TOPIX surpassed the specified level'. On the other
price of the TOPIX up to four weeks in the future and hand, one has to buy all of the stocks which he sold at a
the prediction of the lowest price of the TOPIX up to time when the TOPIX went below the other specified
four weeks in the future. Let AVbe the average of the level. However, one should consider the possibility that
above. he could have received greater return by dividing
3. If (Current TOPIX)x(l+a) is below AV, then buy. dealings*. Moreover, the DSS assumes that AV should be
when the TOPIX goes above AV x(l+p), then sell.
4. If Current TOPIX is above Avx(l+p), then sell.
When (TOPIX)x(l+a) drops below AV,then buy. As shown in the Step 2 of DSS 2, one has to sell when the TOPIX has
5 . Otherwise, wait. gone above AV x ( I + p). (One has to buy when TOPIX x (l+a) has gone
In the above decision support system, small positive below AV.)
parameters a and p play important roles for preventing Suppose that you have 10 units of the stock. What kind of dealing
methods do you choose when the price of the stock is going up? Probably,
frequent dealings and ensuring appropriate amount of more than half of the persons who have successfully encountered such a
returns. (Fig.4 shows an illustration which explains situation don't sell all at the same time. Some of them may sell 30 % or
dealings in the proposed DSS when a=P=O.) 50 % of the stocks and wait patiently in order to receive greater return in the
near future.

0-7803-7278-6/02/$10.00 02002 IEEE 2152

kept constant during one cycle of dealing in the TOPIX. strategy. (Due to limitation of the space, we don’t go into
However, this assumption sometimes may become details concerning the chromosomes related to other parts.
harmful. Interested readers are kindly asked to attend our
In this section, we will try to utilize GAS for the presentation.)
following three objectives: When the TOPIX goes up beyond AV and stands in
To utilize GAS in order to find appropriate region A, the two bits in the chromosome indicate the
parameter values a and p. following dealings:
To utilize GAS in order to find successful way of 00: no dealing
dealing in the TOPIX. 01: sell 113 of the stocks bought3 (sell TOPIX by using
To utilize GAS in order to check whether the AV 113 of the amount of money that one wants to
should be kept constant or the current average in~est.~)
estmimate should be used as the value of AV. 10: sell 2/3 of the stocks bought (sell TOPIX by using
2/3 of the amount of money that one wants to
Utilization of GAS invest.)
In the following, we shall show how GAS are utilized 11: sell all of the stocks bought (sell TOPIX by using
in order to find appropriate parameter values a and p and all of the money that one wants to invest.)
appropriate dealing rules for TOPIX. Fig. 5 shows the
chromosome which we have used for the above objective. Computer Simulation Results:
Evaluation of each chromosome was done using the
data from November 1995 to October 1996. As the best
a , 6 Avk B c D E F chromosome, we have found:
0011101110011111100
1 /
/
Prediction ofthe highest price o f TOPIX
up to four weeks in the future We have tried to utilize the above chromosome for
dealings in the TOPIX from November 1996 to October
tA /-. TOPIX 1997. During this period, the TOPIX changed from 1556
,,,,/f- Average AV
yen to 1360 yen. This means that we would have lost
about 13 % of our investments if we had utilized

- Prediction of the lowest price of TOPIX

up to four weeks in the future
buy-and-hold strategy. (That is to say, if we bought
TOPIX using 10 billion yen in the initial week and did not
sell at all until the end of October 1997, then the total
would be only 8.73 billion yen after one year.) However,
Fig.5. Structure of the chromosome we succeeded in making a considerably large return by
following the dealing rule obtained by GAS. (The initial
The first six bits of chromosome have been prepared in amount of money [lo billion yen] increased to 10.98
order to find appropriate parameter values a and p of billion yen.).
three bits each. Each set of three corresponds to the Remark 3: The chromosome that we have found indicates
following parameter values a and 0: the following:
000: 0.000 001: 0.003
1) The parameter value of a should be 0.003.
010: 0.006 011: 0.009
2) The parameter value of p should be 0.018.
100: 0.012 101: 0.015 3) AV should be the current estimate concerning the
110: 0.018 111: 0.021 average of the forecasts four weeks in the future.
The one bit following the above six bits has been
4) When the TOPIX goes beyond AV, stands in the
prepared in order to investigate whether the AV should be region A, and satisfies the fourth condition of DSS,
kept constant during one cycle of dealing in the TOPIX. all of the stocks one has should be sold.
1 : AV should not be kept constant. Instead, it should be
the current estimate of the average.
0: AVshould be kept constant.
The latter 12 bits of chromosome have been prepared in ~ ~ ~ ~~~~~

order to find successful method of dealing in the TOPIX. If one has some TOPIX stocks, then he should sell 1/3 of the
In the following, we briefly touch upon how the stocks that he has.
chromosomes in the part A correspond to the dealing If one has no stocks, then he should sell TOPIX by using the 1/3
of the amount of money that he wants to invest.

0-7803-7278-6/02/$10.00 (82002 IEFX 2153

1V. Computer Simulation Results of Dealing TOPIX (decreasing) trend of turnover is one of the most reliable
by Utilizing Traditional Technical Analysis signs that the stock price will go up (down). They
In the last section, we showed that the proposed DSS frequently utilize the measure of SMVT versus 1 MVT in
utilizing NNs & Gas had behaved intelligently under the order to find the current tendency of the stock market.
bear market (1996.11 - 1997.10). As in the previous subsection, we shall embody the above
In this section, we shall investigate by computer idea by constructing the following DSS utilizing SMVT
simulations whether the proposed DSS approach & 1MVT:
outperforms the approach utilizing traditional technical 1) If the sign of SMVT - lMVT changes from positive
analysis. to negative, then buy TOPIX.
2) If the sign of 5MVT - lMVT changes from negative
A. 13 Weeks Moving Average (13MA) of TOPIX Versus to positive, then sell all of the stocks having been bought.
S(4) Weeks Moving Average (SMA(4MA)) of TOPIX
Traditionally, many stock traders utilized the measure C. Computer Simulation Results
13MA - 5(4)MA in order to detect up and down tendency
of the stock market. Figure 6 and Figure 7 illustrate We have camed out computer simulations for dealings
golden cross and dead cross of two moving averages, (utilizing the DSS proposed in the previous subsections)
respectively. “Golden Cross” is the sign that the price of in the TOPIX from November 1996 to October 1997.
TOPIX is becoming higher and “Dead Cross” is the sign Table 1 and Table 2 show the changes of the initial
that the price of TOPIX is moving downwards. Here, money
we shall embody the above idea by constructing the (10 billion yen) by dealings utilizing the DSS proposed in
following decision support system for dealing TOPIX: the previous subsections. Although all of the proposed
1) If the sign of 13MA - 5(4)MA changes from DSS utilizing traditional technical analysis outperform the
positive to negative, then buy TOPIX. buy-and-hold strategy, they could not surpass the
2) If the sign of 13MA - 5(4)MA changes from performance of the DSS utilizing NNs & Gas
negative to positive, then sell all of the stocks having been
bought Table. 1: Dealings by Using the Changes of the Sign of
TQSX
(13MA-5MA)
t

mi3
Fig.6. An illustrative examde of the golden cross
1a m
b

I c
dam
Fig.7. An illustrative example of the dead cross

B. 5 Weeks Moving Average of Turnover (SMVT) Versus 1

Week Moving Average of Turnover (1MVT)
Many technical analysts believe that increasing

0-7803-7278-6/02/$10.00 02002 IEEE 2154

 Table.2: Dealings by Using the Changes of the Sign of full confirmation concerning the developed DSS, we need
to carry out huge numbers of simulations through a rather
long range of years.
The idea proposed in this paper could be easily applied
for dealing in each listed stock in the Tokyo Stock Market.
(Of course, with several minor modifications, it could
also be applied for dealing in stocks in the foreign
markets.)
In this paper, we have only tried to utilize the two soft
computing techniques ANNs and GAS. If one could
design a decision support system which also utilizes fuzzy
engineering and a reinforcement learning method such as
TD-Learning [12] in addition to NNs and GAS and
succeed in eliciting the compound merits from these soft
computing techniques, it would be helpful for making the
DSS more reliable. These are also left for our future
study.

REFERENCES
[I] A.P. Refenes, Editor, Neural Networks in the Capital Markets,
Wiley, 1995.
[2] A.S. Weigend et al, Editors, Decision Technologies for
Financial Engineering, World Scientific, 1997.
[3] H. White, "Economic prediction using neural networks: The
case of IBM daily stock returns", Proceedings of the IEEE lnt.
Conf. on Neural Networks, Vol. 11, pp. 45 1-458, 1988.
[4] A.P. Refenes et al, "Neural networks in financial engineering:
A study in methodology", IEEE Trans. Neural Networks,
V01.8, pp.1222-1267, 1997.
[ 5 ] N. Baba and M. Kozaki, "An intelligent forecasting system of
Remark 4.C: The computer simulations have been done stock price using neural network", Proceedings of IJCNN, Vol.
using the data only from November 1996 to October 1997. 1, pp. 371-377, 1992.
In order to derive the general conclusion, we should carry [6] N. Baba et al, "A hybrid algorithm for finding the global
out computer simulations for long periods. In the DSS minimum of error function of neural networks and its
applications", Neural Networks, Vol. 7, pp. 1253-1265, 1994.
proposed in this section, we deal stocks after the sign
[7] S . Mahfoud and G. Mani, "Financial forecasting using genetic
(13MV-5MV, and etc.) changes. However if we could algorithms", Applied Artificial Intelligence, Vol. I O ,
construct a system which is able to detect the change of pp.543-565, 1996.
the sign earlier, we could receive higher retum. We also [8] S.H. Chen and W.C. Lee, "Option pricing with genetic
have to compare the performance of the proposed DSS algorithms: The case of European Style of Options",
utilizing NNs & GAS with that of the smarter DSS. These Proceedings of ICGA, pp. 704-71 1, 1997.
are left for our future study. [9] K.N. Pantazopoulos et al, "Financial prediction and trading
strategies using neuro-fuzzy approaches", IEEE Trans. SMC,
Vol. 28, NO. 4, pp. 520-53 I , 1998.
V. CONCLUDING REMARKS [IO] N. Baba et al, "Utilization of neural networks and GAS for
A decision support system (DSS) for dealing in the constructing an intelligent decision support system to deal
TOPIX which utilizes ANNs and GAS is developed in this stocks", Proceedings of the SPIE conference, U.S.A.,
paper. ANN is utilized in order to make forecasts V01.2760, pp. 164- 174, 1996.
conceming the highest price and the lowest price of [ 111 N. Baba et al, "Utilization of neural networks & GAS for
constructing reliable decision support systems to deal stocks",
TOPIX up to four weeks in the future. GA is utilized in Proceedings of IJCNN'2000, Italy, Vol. V, pp. 111- 1 16,2000.
order to find an intelligent way of dealing in the TOPIX. [ 121 N.Baba and H.Suto, "Utilization of artificial neural networks
Computer simulations which have been carried out and TD-learning method for constructing intelligent decision
confirm the effectiveness of the developed DSS using support systems", European Joumal of Operation Research.
NNs & GAS. However, these simulations have been done Vol. 122, pp. 501-508,2000,
during a rather short range of years. In order to execute

0-7803-7278402/$10.00 02002 IEEE 2155