Vol.2, No.

1, 2006 83-99

The Application of Regression Model and Artficial Neural Network for Studying the Taiwan
Stock Index Future
1
2
3

( Received: Nov. 7, 2005First Revision: Apr. 10, 2006Accepted: May. 12, 2006 )

87721

1055
5

Abstract
Taiwan stock market formally established Taiwan Stock Index Future Contract on July
21, 1998, it could be said a milestone for Taiwan finance career's liberalization and
internationalization. It offers investors the new investment product and hedging risk tools, and
also offers speculators and arbitragers the opportunity of making more profit at less fund.
Especially, the higher futures transactions, the more potential of futures market in these few
years. Artificial Neural Network is a tool of information technique that rapidly rises in these
few years. Especially using in finance field, the performance is very outstanding. So this
study tries to use Artificial Neural Network and Regression Analysis of statistical methods in
order to predict the next day closing index of FITX, and then find the suitable prediction
1

83

 Vol.2, No.1, 2006

model to create the better rate of gaining profit. The results of the study are: At the aspect of
the Back-Propagation network, it could be found that the model with hidden layer is not big
difference than the model without hidden layer for the prediction performance. At the aspect
of Regression Analysis, By ruling out three outliers and by stepwise regression analysis to
select the five variables of the closing index, basis difference, up and down, 10 days W%R
and 5 days BIAS into the model. At the aspect of improving Artificial Neural Network, the
five variables selected by using stepwise regression analysis method are regard as input
variables. It could be found that the model without hidden layer is better than the model with
hidden layer for the prediction performance, but the difference is small. At the prediction
performance of three models comparison aspect, It could be found the performance of
improving model of Artificial Neural Network is the better choice, Regression Analysis is the
best one, but Artificial Neural Network is the worst one. So it is not certainly better when
variables are more but create too much complex, on the contrary, decrease some effects
between some variables. Improving model of Artificial Neural Network is to simplify
variables in order to decrease effects and get better prediction performance.
Keywords : Taiwan stock index future, Artificial neural network, Regression analysis

1.
1.1
86
87721(TAIFEX, Taiwan Futures Exchange)
FITX

84

1.2

(FITX)8912
21911218

1.
2.

3.
4.

2.

(2003)

2.1
(processing element, PE)
(layer)(network)(artificial neuron)
()

Yj = f (Wij X i j ), i = 1, ..., p; j = 1, ..., n

i

Xi;
Yj;
f(Transfer Function);
Wij;
j(Threshold)
1
2.2 (Back-Propagation Network, BPN)
Rumelhart McClelland PDP 1986
(layered feed-forward network)
Sigmoid functiontanH function

(the gradient steepest descent method)

85

 Vol.2, No.1, 2006

W1j
X1

Wij

Xi

= f (W ij X i j )

Yj

Xp

Wpj

1
(Back-Propagation Learning Algorithm)

2001

NeuralWorks Professional /PLUS: NW2

Delta-Rule2003a&b

3.
Kimoto Asakawa(1990)

Chiang, Urban,
Baldridge(1995) 1981
1986 101 15

Warner Misra(1996)

(1997)

(2001)
(TAIMEX)
87 9 2 88 12 28 14

86

(2002)

(2002) 680

(2005)

1
1

Cheng Titterington(1994)
1 Cheng
Titterington(1994)

4.

4.1
:

4.2
2
5 20
(2001)
(2004)

87

 Vol.2, No.1, 2006

1990 Kimoto
Asakawa

1995 Chiang,Urban,
Baldridge

1996 Warner
Misra

1997

2001

2002

2002

2005

88

2(2001)9KDMA
RSI
ARBR

5.

(MAE)
(MSE):
(Mean Absolute ErrorMAE)

MAE
MAE = Oi - Ei n
Oi:Ei:n:
(Mean Square ErrorMSE)

MAEMSE

MSE = (Oi - Ei)2 n

5.1
NW2

(MA)
10

(RSI)
5

(KD)
9

(DMI)
10

(BIAS)
5

(W%R)

(VR)
10

10

AR

BR

20

20

89

(MACD)
9

(PSY)
10

 Vol.2, No.1, 2006

5.1.1
(FITX)
891221911218490
89122191
61936291620911218128

5.1.2
4.2
3

(Sigmoid function)
(0,1)(0.2,0.8)

5.1.3
Delta-Rule

NW21
(Lcoef Ratio)0.5(momentum)0.9, 0.4 0.1

X1:
X2:
X3:
X4:
X5:
X6:
X7:10
X8:5RSI
X9:9K
X10:9D

X11:9MACD
X12:10DI
X13:10DI
X14:5BIAS
X15:10W%R
X16:10PSY
X17:10VR
X18:20AR
X19:20BR

Y:

90

5.1.4
(2003b)12
1(2003b)
0

(2003b)4-164-17
01

+/2
++
2(TanH)

5.1.5
(Correlation Coefficient) (Root of Mean Square Error,
RMSE) [-1,1]RMSE
RMSE =(Oi - Ei) 2n1/2
Oi: Ei: n:
RMSE
19 0
1
RMSE 5.1.1 5.1.5
NW2 18 (2003) 35
36 4-2 4-6(2003)

19 20 1

1 0.9
(Sigmoid function)
MAE 72.866 MSE 10231.49
2

5.2
5.2.1
5.13

Yi = 0 + 1X1i + 2X2i + +19X19i +i

iiid N( 0, 2), i=1, 2, , 362.

91

 Vol.2, No.1, 2006

6000

5500
5000
4500
4000

20
0

2/
20 6/20
02
/
20 6/30
02
/
20 7/10
02
/
20 7/20
02
/7
20 /30
02
20 /8/9
02
/
20 8/19
02
/8
20 /29
02
20 /9/
02 8
/
20 9/18
02
/
20 9/28
02
20 /10
0 2 /8
/
20 10/1
02 8
/1
20 0/28
02
20 /11
0 2 /7
/
20 11/1
02 7
/1
20 1/27
02
20 /12
0 2 /7
/1
2/
17

3500

89122191619362
91620911218128

(stepwise regression)
(outlier)359

-25.961

48.858

-0.566

0.572

(X1)

0.997

0.008

128.064

0.000

(X5)

0.655

0.160

4.104

0.000

BIAS (X14)

16.683

4.589

3.635

0.000

0.713

0.282

2.526

0.012

-0.167

0.073

-2.282

0.023

(C)

W%R (X15)
(X6)

92

5
Y= 25.961+0.997X1 +0.655X50.167X6+16.683 X14 +0.713 X15
(Adjusted R2)98.1% (ANOVA)
6
5.2.2
3
90 1 16 90 12 12 91 1 21 339351
361 3

5.1.2 3

(0,1)(0.2,0.8)

3
3
5.2.3

1. Kolmogorov-Smirnov p =0.2 0.05

2.

3.

MAE72.878MSE9709.518
3

195343523.890

39068704.778

3762311.771

353

10658.107

199105835.661

358

93

3665.633

0.000

 Vol.2, No.1, 2006

6000
5500

5000

4500

4000

20
02
/6
20 /20
02
/6
20 /30
02
/7
20 /10
02
/7
20 /20
02
/7
/
20 30
02
20 /8/9
02
/8
20 /19
02
/8
/
20 29
02
20 /9/8
02
/9
20 /18
02
/9
20 /28
02
/
20 10/8
02
/1
20 0/1
02 8
/1
0
20 /28
02
/
20 11/7
02
/1
20 1/1
02 7
/1
1
20 /27
02
20 /12/
02 7
/1
2/
17

3500

3
5.3
5.1 19
5.2

5.1

TanH 21 (2003) 45
46 4-11 4-16(2003)
5 6 1 0.1
(Sigmoid Function)
MAE 73.441 MSE
10075.31 4

0 0

94

6000
5500

5000

4500

4000

20
02
/6
20 /20
02
/
20 6/30
02
/7
20 /10
02
/7
20 /20
02
/7
20 /30
02
20 /8/9
02
/8
20 /19
02
/8
20 /29
02
20 /9/8
02
/9
20 /18
02
/
20 9/28
02
20 /10/
02 8
/
20 10/1
02 8
/1
20 0/28
02
20 /11/
02 7
/
20 11/1
02 7
/1
20 1/27
02
20 /12/
02 7
/1
2/
17

3500

4
5.4
7 MAE MSE

7 MAE MSE

MAE

MSE

72.866

10231.49

72.878

73.441

95

9709.518
10075.31

 Vol.2, No.1, 2006

ANN

ANN

6000
5500

5000

4500

4000

20
02
/6
20 /20
02
/6
20 /30
02
/7
20 /10
02
/7
20 /20
02
/7
/3
20 0
02
/
20 8/9
02
/8
20 /19
02
/8
/2
20 9
02
/
20 9/8
02
/9
20 /18
02
/9
20 /28
02
/
20 10/8
02
/1
20 0/18
02
/1
0
20 /28
02
/
20 11/7
02
/1
20 1/17
02
/1
1
20 /27
02
/
20 12/7
02
/1
2/
17

3500

6.
6.1

1.
MAE 72.866 MSE 10231.49

2.
10 (W%R)5 (BIAS) 5
MAE 72.878 MSE 9709.518

3. 5

MAE 73.441 MSE 10075.31

5.1.4 (2003b)

96

4.

5.

6.

(2002)

2002

6.2

(Warner
and Misra, 1996)

6.3
(2001)8
8
89.12.2191.12.18
(2001) 87.9.288.12.28

19 (2001) 13
9 KD
97

 Vol.2, No.1, 2006

MA RSI
ARBR

TanH 0.9 & 0.1

(2001)

89.12.2191.12.18

87.9.288.12.28

10 5 RSI9
K 9 D 9 MACD
10 DI10 DI5 BIAS
10 W%R10 PSY10 VR
20 AR20 BR (19 )

3MA-6MA6RSI9KDDIF
MACD (13 )

()

Sigmoid Function &

TanH 0 1
Delta Rule 1
0.9,0.4 0.1

Sigmoid Function
0 1
Delta Rule 1
0.4

98

1.

(2003)
EMBA
2. (2002)-

3. (1997)-

4. (2003)

5. (2004)
5
6. (2001)

36 91-109
7. (2005)

117-139
8. (2003a)
9. (2003b) 8
10. (2002)

11. http://www.taifex.com.tw/
12. Cheng, B. , D. M. Titterington, (1994), Neural Networks: A Review from a Statistical
Perspective (including discussion), Statistical Science, l.(1), pp.2-54.
13. Chiang, W. C., T. L. Urban, and G. W. Baldridge (1995), A Neural Network
Approach to Mutual Fund Net Asset Value Forecasting, Omega, Int. J. Mgmt. Sci.
24(2), pp.205-210.
14. Kimoto, T., K. Asakawa (1990), Stock Market Prediction System with Modular
Network, in Proc. IEEE Int. Conf. Neural Network, 1, pp.1-6.
15. Rumelhart, D. E., J. L McClelland and the PDP Research Group (1986), Parallel
Distributed Processing: Explorations in the Microstructure of Cognition, MIT press,
Cambridge, MA.
16. Warner, B., M. Misra (1996), Understanding Neural Networks as Statistical Tools,
The American Statistician, 50(4), pp.284-293.

99