Expert Systems with Applications 38 (2011) 14346–14355

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

Short communication

Forecasting stock indices with back propagation neural network

Jian-Zhou Wang a, Ju-Jie Wang a, Zhe-George Zhang b,c,⇑, Shu-Po Guo a
a
School of Mathematics and Statistics, Lanzhou University, Lanzhou 730000, China
b
Department of Decision Sciences, Western Washington University, Bellingham, WA 98225, USA
c
Faculty of Business Administration, Simon Fraser University, Burnaby, British Columbia, Canada V5A 1S6

a r t i c l e i n f o a b s t r a c t

Keywords: Stock prices as time series are non-stationary and highly-noisy due to the fact that stock markets are
Wavelet de-noising affected by a variety of factors. Predicting stock price or index with the noisy data directly is usually sub-
BP neural network ject to large errors. In this paper, we propose a new approach to forecasting the stock prices via the Wave-
WDBP neural network let De-noising-based Back Propagation (WDBP) neural network. An effective algorithm for predicting the
Stock prices
stock prices is developed. The monthly closing price data with the Shanghai Composite Index from Jan-
uary 1993 to December 2009 are used to illustrate the application of the WDBP neural network based
algorithm in predicting the stock index. To show the advantage of this new approach for stock index fore-
cast, the WDBP neural network is compared with the single Back Propagation (BP) neural network using
the real data set.
Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction without this restrictive assumption have been shown to outper-

form the statistical models empirically by many recent studies
Stock price data are always one of the most important infor- such as Enke and Thawornwong (2005), Hansen and Nelson
mation to investors. Unfortunately, stock prices are essentially (2002), Ture and Kurt (2006) and Zhang (2003). Therefore, AI ap-
dynamic, non-linear, nonparametric, and chaotic in nature. This proaches, such as neural network, fuzzy system, and genetic algo-
implies that the investors must handle the time series which rithm, have been utilized in predicting stock prices in Armano,
are non-stationary, noisy, and have frequent structural breaks Marchesi, and Murru (2005), Chen, Leung, and Daouk (2003), Chun
(Oh & Kim, 2002; Wang, 2003). In fact, stock prices’ movements and Kim (2004), Kim and Han (2000), Shen and Loh (2004), Tha-
are affected by many macro-economical factors such as political wornwong and Enke (2004), Vellido, Lisboa, and Meehan (1999)
events, company’s policies, general economic conditions, com- and Wang (2002a,b).
modity price index, bank rates, investors’ expectations, institu- In this paper, we propose a hybrid forecasting model called
tional investors’ choices, and psychological factors of investors. Wavelet De-noising-based Back Propagation (WDBP) neural net-
Thus forecasting stock price movement accurately is not only ex- work. In such a model, the original data are first decomposed
tremely challenging but also of great interest to investors. into multiple layers by the wavelet transform. Each layer has a
Past works on forecasting stock prices can be classified into two low-frequency and a high-frequency signal component. Then a
categories: statistical and artificial intelligence (AI) models. The Back Propagation (BP) neural network model is established by
statistical approach includes autoregressive integrated moving the low-frequency signal of each layer for predicting the future
average (ARIMA), generalized autoregressive conditional heter- value. To the best of our knowledge, this paper is the first at-
oskedasticity (GARCH) volatility (see Franses & Ghijsels, 1999), tempt to utilizing the WDBP neural network based algorithm
and the smooth transition autoregressive model (STAR) (see Saran- for forecasting the stock prices. The empirical data set of Shang-
tis, 2001). These models are mainly based on the assumption of lin- hai Composite Index (SCI) closings prices from January 1993 to
earity among normally distributed variables. However, the December 2009 are used to illustrate the application of the
linearity and normality assumption may not be satisfied in model- WDBP neural network. Furthermore, the superiority of our
ing the stock price movements. On the other hand, the AI models model is shown by comparing the WDBP neural network with
a single BP neural network.
The rest of this paper is organized as follows. Section 2 focuses
⇑ Corresponding author at: Department of Decision Sciences, Western Washing- on the WDBP neural network algorithm for forecasting stock price.
ton University, Bellingham, WA 98225, USA. Tel.: +1 360 650 2867; fax: +1 360 650
Section 3 presents the experimentation design and numerical
4844.
E-mail addresses: gzhang@sfu.ca, George.zhang@wwu.edu (Z.-G. Zhang).
results. Section 4 concludes.

0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2011.04.222
 J.-Z. Wang et al. / Expert Systems with Applications 38 (2011) 14346–14355 14347

2. The proposed WDBP neural network C

2.1. Artificial neural networks

In general, artificial neural networks (ANNs) possess attributes Output Layer

of learning, generalizing, parallel processing and error endurance.
These attributes make the ANNs powerful in solving complex prob-
lems (see Azadeh, Ghaderi, & Sohrabkhani, 2008). Back-Propaga-
tion (BP) neural networks, a type of ANNs, take inputs only from Hidden Layer
the previous layer and send outputs only to the next layer (see
Fig. 1).
Our study employs a BP neural network which is widely used in
Input Layer
business situations. A three-layer BP neural network is shown in
Fig. 2. The BP process determines the weights for the connections
among the nodes based on data training, yielding a minimized
least-mean-square error between the actual and the estimated val-
ues from the output of the neural network. The connection weights
are assigned initial values first. The error between the predicted
S X T U V
and actual output values is back-propagated via the network for
updating the weights (see Wang, 2009). Theoretically, neural net- Fig. 2. The architecture one hidden layer back propagation neural network.
works can simulate any kind of data pattern given sufficient train-
ing. The neural network must be trained before being applied for
forecasting. Assume that there are n input neurons, m hidden neu- where fo is the activation function, usually a linear function. All
rons, and one output neuron, a training process can be described by weights are assigned with random values initially, and then modi-
the two stages (see Zhang & Wu, 2009): fied by the delta rule according to the learning samples.

(I) Hidden layer stage: The outputs of all neurons in the hidden
layer are calculated by the following steps: 2.2. Wavelet transform
X
n
netj ¼ v ij xi ; j ¼ 1; 2; . . . ; m; ð1Þ Wavelet transform is used in analyzing non-stationary time ser-
i¼0 ies for generating information in both the time and frequency do-
mains. It may be regarded as a special type of Fourier transform at
yj ¼ fH ðnet j Þ; j ¼ 1; 2; . . . ; m: ð2Þ multiple scales and decomposes a signal into shifted and scaled
versions of a ‘‘mother’’ wavelet. The continuous wavelet transform,
Here netj is the activation value of the jth node, yj is the output of
denoted by CWT, is defined as the convolution of a time series x(t)
the hidden layer, and fH is called the activation function of a node,
with a wavelet function w(t) (Goswami & Chan, 1999):
usually a sigmoid function as follows:
Z  
1 1 tb
fH ðxÞ ¼ : ð3Þ CWT wx ðb; aÞ ¼ /wx ðb; aÞ ¼ pffiffiffiffiffiffi xðtÞ  w dt; ð5Þ
1 þ expðxÞ jaj a
(II) Output stage: The outputs of all neurons in the output layer
are given as follows: where a is a scale parameter, b is the translational parameter and 
! is the complex conjugate of w(t). Let a = 1/2s and b = k/2s, where s
X
m
O ¼ fo xjk yj ; ð4Þ and k belong to the integer set Z. The CWT of x(t) is a number at
j¼0 (k/2s, 1/2s) on the time-scale plane. It represents the correlation be-
tween x(t) and w(t) at that time-scale point. A discrete version of
Eq. (5) is thus obtained as

  Z  
k 1 1
t  k=2s
DWT wx ðk; sÞ ¼ /wx ; ¼ xðtÞ  w dt; ð6Þ
2s 2s 1 1=2s

which separates the signal into components at various scales corre-

sponding to successive frequencies.Note that DWT corresponds to
the multi-resolution approximation expressions for analysis of a
signal in many frequency bands (or at many scales). In practice,
multi-resolution analysis is carried out by starting with two chan-
nel filter banks composed of a low-pass and a high-pass filter, and
then each filter bank is sampled at a half rate of the previous fre-
quency. The number of steps of this de-composition procedure will
depend on the length of data. The down sampling procedure keeps
the scaling parameter constant (1/2) throughout successive wavelet
transforms (Li & Kuo, 2008).
Over the past decade, DWT has been well developed and ap-
plied in signal analysis of various fields (Mörchen, 2003). In this
study, DWT is utilized to remove noise from the stock prices data
Fig. 1. Back-propagation neural network. for prediction purpose.
14348 J.-Z. Wang et al. / Expert Systems with Applications 38 (2011) 14346–14355

2.3. The WDBP neural network model algorithm 1. Wavelet decomposition: This step decomposes the signal into
low-pass filter B and high-pass filter D by the wavelet trans-
The WDBP neural network algorithm first decomposes the ori- form. The low-pass filter B reflects the main features of the sig-
ginal data into several layers via the wavelet transform, then estab- nal, and the high-pass filter D represents random factors often
lishes a BP neural network model using the low-frequency signal of called the noise. The main purpose of the wavelet de-composi-
every layer for prediction. The algorithm is described as follows tion is to separate the basic characteristics from the noise of the
and the flowchart is shown in Fig. 3. signal.

No

No

Fig. 3. The algorithm of the proposed WDBP neural network model.

 J.-Z. Wang et al. / Expert Systems with Applications 38 (2011) 14346–14355 14349

6000

4000
Stock closings price

2000

0
0 34 68 102 136 170 204
Time

Fig. 4. The original data.

2.4. Evaluation criteria

Furthermore, three accuracy measures are utilized to evaluate

the forecasting performance relative to the actual price C MP
n . These
are mean absolute error (MAE), root mean-square error (RMSE)
and mean absolute percentage error (MAPE) which are given as
follows:
T 
X 
 MP 
MAE ¼ T 1 C n  C n ; ð7Þ
n¼1

!1=2
X
T
2
RMSE ¼ T 1 ðC MP
n  CnÞ ; ð8Þ
n¼1

T 
X 
 MP MP 
MAPE ¼ T 1 ðC n  C n Þ=C n ; ð9Þ
n¼1
Fig. 5. The six levels of db3 wavelet de-composition.
where C MP
n and Cn are the actual value and predicted value, respec-
2. Main wave extraction: This step extracts the low-pass filter Bi tively, and T is the sample size.
(i = 1, 2, . . . , n) and remove the high-pass filter Di Smaller values of these measures indicate more accurate fore-
(i = 1, 2, . . . , n) in the data. Here n is the number of layers of casted results and if the results are not consistent among three cri-
the wavelet de-composition. teria, we choose the relatively more stable MAPE, as suggested by
3. Normalization: Low frequency data Bi is normalized into Ai. Makridakis (1993), to be the benchmark. In this paper, we use all
4. Training and testing data sets determination: Divide the low-pass three measures to evaluate the forecasting performance.
filter data Ai into two subsets: training Ai(1) and testing Ai(2)
data. A validation must be performed by using Ai(2) to test
3. Experimentation design
how well the network is able to generalize for unknown data.
To cover wide ranges of outcomes, it is necessary to achieve a
3.1. Data preparation
balance between the training and validation data set sizes.
5. Relation estimation: Estimate the relation between input(s) and
The data for our experiments are SCI closing prices, collected on
output(s) through training the BP network using the training set
the Shanghai Stock Exchange (SSE). The total number of values for
Ai(1). To find the appropriate number of hidden nodes, repeat
the SCI closing prices is 204 trading months, from January 1993 to
these steps using different training parameters for networks
December 2009. Fig. 4 shows the original data series. The data set
with 1 to q nodes in their hidden layer. Training continues until
is partitioned into a training set (80%) and a testing set (20%) for
the estimation error is below a threshold.
validation.
6. Validation: Validate the network using Ai(2) and make forecasts.
As the network’s output value is between 0 and 1 (a character-
7. De-normalization: De-normalize the predicted values.
ðiÞ istics of the transfer function), we need to normalize the original
8. Evaluation: Check if PMAPE , the mean absolute percentage error
ðiÞ data series. A stock closings price p is normalized to p0 by
(MAPE) is no more than a threshold. If PMAPE 6 c, the training
is stopped and the algorithm ends; otherwise, set i = i + 1 and pi  pmin
p0i ¼ ; i ¼ 1; 2; 3; . . . ; 204; ð10Þ
go to the Step 4. pmax  pmin
14350 J.-Z. Wang et al. / Expert Systems with Applications 38 (2011) 14346–14355

Low-frequency signals A1 High-frequency signals D1

10000 1000
A1

D1
5000 0

0 -1000
0 34 68 102 136 170 204 0 34 68 102 136 170 204

Low-frequency signals A2 High-frequency signals D2

10000 1000
A2

D2
5000 0

0 -1000
0 34 68 102 136 170 204 0 34 68 102 136 170 204

Low-frequency signals A3 High-frequency signals D3

10000 1000
A3

D3
5000 0

0 -1000
0 34 68 102 136 170 204 0 34 68 102 136 170 204

Low-frequency signals A4 High-frequency signals D4

10000 1000
A4

D4
5000 0

0 -1000
0 34 68 102 136 170 204 0 34 68 102 136 170 204

Low-frequency signals A5 High-frequency signals D5

10000 1000
A5

D5

5000 0

0 -1000
0 34 68 102 136 170 204 0 34 68 102 136 170 204

Low-frequency signals A6 High-frequency signals D6

10000 1000
A6

D6

5000 0

0 -1000
0 34 68 102 136 170 204 0 34 68 102 136 170 204

Time Time

Fig. 6. The de-composition process of the experimental data.

Original signals
6000

3000

0
0 34 68 102 136 170 204
Signals Reconstruction
6000

3000

0
0 34 68 102 136 170 204
-8
x 10 Error signals
5

-5

-10
0 34 68 102 136 170 204

Fig. 7. The re-construction process of the experimental data.

 J.-Z. Wang et al. / Expert Systems with Applications 38 (2011) 14346–14355 14351

3000
Actual Value
BP1 Fitting Value
Error

2000

1000

-1000
0 33 66 99 132 165

Fig. 8. The comparison chart of the actual value and BP1 fitting value.

3000
Actual Value
BP2 Fitting Value
Error

2000

1000

-1000
0 33 66 99 132 165

Fig. 9. The comparison chart of the actual value and BP2 fitting value.

where pmax and pmin are the maximum and minimum value of the obtained by superimposing detail D2 on approximation A2. Further-
original series, respectively. more, A2, A3, A4 and A5 are obtained by the similar iterations.
The appropriate number of levels of wavelet decomposition can
3.2. Wavelet de-noising be determined by the nature of a time series, according to domi-
nant frequency components (Mörchen, 2003), entropy criterion
From Fig. 4, we can see that the observed data are contaminated (Coifman & Wickerhauser, 1992), or application’s characteristics
by a lot of noise. The noise can be removed from the observed data (Li & Shue, 2004). Generally speaking, the number of levels of
by DWT. decomposition depends on the length of the time series. For exam-
There are a variety of wavelets proposed in the literature for ple, a six level decomposition is used in this study.
performing DWT. Each has its own application domain with unique The noise of SCI closings prices sample data in our study has
resolution capability, efficiency, and computational cost etc. In this been reduced by the db3 wavelet’s decomposition. Decomposition
study, the computationally efficient Daubechies (db3) wavelet is and reconstruction processes are shown in Figs. 6 and 7,
used. The Daubechies (db3) wavelet is the common wavelet and respectively.
its decomposition process is shown in Fig. 5.In Fig. 5, Ai and Di
(i = 1, 2, 3, 4, 5, 6) are the approximation and detail components, 3.3. WDBP neural network
respectively. Ai’s (i = 1, 2, 3, 4, 5, 6) represent the high-scale and
low-frequency components of the time series and Di’s Although many different neural network models have been ap-
(i = 1, 2, 3, 4, 5, 6) the low-scale and high-frequency components. plied in finance, we have chosen to use the BP neural network in this
These approximation and detail records are reconstructed from paper due to its popular use in the short-term forecasting situations.
the wavelet coefficients. The first high-pass filter provides the Unlike other BP models directly constructed by the original data, we
detail D1. The first low-pass filter is approximation A1, which is first decompose the original data into many layers by the wavelet
14352 J.-Z. Wang et al. / Expert Systems with Applications 38 (2011) 14346–14355

3000
Actual Value
BP3 Fitting Value
Error

2000

1000

-1000
0 33 66 99 132 165

Fig. 10. The comparison chart of the actual value and BP3 fitting value.

3000
Actual Value
BP4 Fitting Value
Error

2000

1000

-1000
0 33 66 99 132 165

Fig. 11. The comparison chart of the actual value and BP4 fitting value.

transform. Every layer has a low-frequency signal and high-fre- this experiment, we compare our models with the BP network
quency signal. Then we establish a BP neural network model by established by the original data without de-noising processing.
the low-frequency signal of each layer and predict its future. In all This BP without de-noising values versus actual values are shown
BP models, we choose three values of every quarter as an input sam- graphically in Fig. 14. In addition, the comparisons between the ba-
ple, and the total observations (the stock closings prices) are divided sic BP and our six models are presented in Table 1 and are graph-
into 68 groups. The former 55 groups of every layer’s low-frequency ically shown in Fig. 15. Table 2 shows accuracies of these models
data is used for training and the remaining groups for testing. based on MAE, RME and MAPE and indicate that the BP4 model
In this paper, we chose a three-layer BP neural network which has the smallest errors among these models. Hence, BP4 should
has a 3-neuron input layer, a 10 neuron hidden layer, and a 3- be chosen as our WDBP model. The WDBP model selected in this
neuron output layer. Decomposing the original data into six layers, process outperforms the conventional BP model significantly.
we establish the six models shown in Figs. 8–13.
In these Figs. 8–13, we show how BPi values fit the actual values,
where BPi (i = 1, 2, 3, 4, 5, 6) denotes the model based on the ith 4. Concluding remarks
layer of data. The errors are also shown in these figures.
This paper proposes an improved way of forecasting the stock
closings price based on the neural network. In previous studies,
3.4. Results and discussion neural networks are established with the original data for predic-
tion. However, the stock market data are highly random and
In order to verify the proposed WDBP method, we have con- non-stationary, thus contain much noise. The prediction accuracy
ducted a forecasting experiment with the SCI closings prices. In of traditional neural networks without the de-noising process is
 J.-Z. Wang et al. / Expert Systems with Applications 38 (2011) 14346–14355 14353

3000
Actual Value
BP5 Fitting Value
Error

2000

1000

-1000
0 33 66 99 132 165

Fig. 12. The comparison chart of the actual value and BP5 fitting value.

3000
Actual Value
BP6 Fitting Value
Error

2000

1000

-1000
0 33 66 99 132 165

Fig. 13. The comparison chart of the actual value and BP6 fitting value.

3000
Actual Value
BP Fitting Value
Error

2000

1000

-1000
0 33 66 99 132 165

Fig. 14. The comparison chart of the actual value and BP fitting value.
14354 J.-Z. Wang et al. / Expert Systems with Applications 38 (2011) 14346–14355

Table 1
Comparing the predicted values of seven models (the BPi (i = 1, 2, 3, 4, 5, 6) network and the BP network).

Number Time Actual BP BP1 BP2 BP3 BP4 BP5 BP6

1 Jan-07 2786.34 2372.008 2190.09 2426.758 3129.364 2669.715 2592.685 2220.476
2 Feb-07 2881.07 2396.536 2426.313 2497.576 3505.848 2843.667 2580.487 2611.874
3 Mar-07 3183.98 2539.259 2689.835 2578.715 4082.62 3109.071 2613.389 2598.046
4 Apr-07 3841.27 2863.294 2085.804 2462.571 3546.475 3121.006 2619.506 2233.963
5 May-07 4109.65 2970.428 1739.596 2514.216 3985.493 3261.657 2624.032 2706.219
6 Jun-07 3820.7 3287.309 1531.385 2576.392 4743.108 3636.844 2649.681 2783.059
7 Jul-07 4471.03 2978.664 2058.575 2466.067 3262.927 3416.227 2660.398 2246.888
8 Aug-07 5218.82 2974.182 1600.961 2516.305 3541.268 3547.855 2707.301 2808.414
9 Sep-07 5552.3 3077.43 1302.705 2575.313 4092.433 4006.215 2715.889 2993.747
10 Oct-07 5954.77 3160.7 2059.712 2465.43 3261.144 3658.971 2674.954 2253.656
11 Nov-07 4871.78 3191.587 1601.633 2515.893 3532.715 3818.983 2717.764 2913.31
12 Dec-07 5261.56 3345.321 1302.91 2576.907 4078.095 4370.94 2726.515 3283.371
13 Jan-08 4383.39 2671.548 2047.312 2445.434 3246.611 3813.24 2681.93 2257.932
14 Feb-08 4348.54 2324.962 1588.765 2503.397 3509.434 3997.667 2725.288 3012.803
15 Mar-08 3472.71 1995.324 1290.585 2622.052 4043.666 4613.772 2732.402 3578.82
16 Apr-08 3693.11 2700.148 2019.745 2177.668 3697.985 2766.339 2762.913 2269.352
17 May-08 3433.35 2554.911 1560.347 2332.732 5348.569 2851.501 2892.879 3132.847
18 Jun-08 2736.1 2536.928 1263.514 3206.021 8782.262 3328.44 2866.599 3855.249
19 Jul-08 2775.72 2262.053 1545.676 2328.798 3375.344 2562.423 2807.252 2272.906
20 Aug-08 2397.37 2031.621 1063.971 2399.028 4942.449 2589.192 2969.296 3239.813
21 Sep-08 2293.78 1900.214 779.6376 2689.265 8234.721 2918.935 2929.419 4179.23
22 Oct-08 1728.79 2098.349 1479.176 1919.298 2835.693 2267.147 2848.619 2277.322
23 Nov-08 1871.16 2159.389 963.0936 1900.923 3887.044 2256.494 3039.129 3334.65
24 Dec-08 1820.81 2139.749 605.1982 2091.826 6005.922 2514.231 2991.506 4449.536
25 Jan-09 1990.66 1651.648 1989.116 1742.71 2760.343 2211.379 2872.987 2281.169
26 Feb-09 2082.85 1509.184 2158.615 1679.256 3039.229 2149.05 3077.423 3408.811
27 Mar-09 2373.21 1419.823 2351.878 1653.744 3425.232 2244.137 3030.119 4654.961
28 Apr-09 2477.57 2186.201 2117.907 2410.279 2676.114 1761.886 2915.987 2287.577
29 May-09 2632.93 2213.058 2060.366 2490.442 2819.038 1709.009 3159.782 3473.601
30 Jun-09 2959.36 2358.663 2084.589 2579.015 3089.402 1876.426 3102.085 4796.619
31 Jul-09 3412.06 2739.148 2088.145 2409.382 2732.674 2219.615 2865.352 2272.974
32 Aug-09 2667.74 2836.968 1753.818 2454.061 2871.431 2330.705 3011.47 3467.348
33 Sep-09 2779.43 3127.578 1555.223 2527.309 3156.292 2443.499 3025.843 4956.111

9000
Actual
BP
8000 BP1
BP2
BP3
7000 BP4
BP5
BP6
6000

5000

4000

3000

2000

1000

0
0 3 6 9 12 15 18 21 24 27 30 33

Fig. 15. The comparison chart of the actual value and the predicted value of the seven models.

Table 2
Comparing the three criteria of the seven models (the BPi (i = 1, 2, 3, 4, 5, 6) network and the BP network).

Errors BP BP1 BP2 BP3 BP4 BP5 BP6

MAE 930.2404 1620.6 1038.4 1340.3 675.9543 1067.1 1377.7
RMSE 1174.3 2006.2 1412.7 1984.4 847.5841 1349.7 1619.1
MAPE 0.2492 0.4364 0.2584 0.4831 0.1948 0.3056 0.4276
 J.-Z. Wang et al. / Expert Systems with Applications 38 (2011) 14346–14355 14355

not satisfactory. The lack of a good forecasting model motivates us Franses, P. H., & Ghijsels, H. (1999). Additive outliers, GARCH and forecasting
volatility. International Journal of Forecasting, 15(1), 1–9.
to find an improved method of making forecasts called WDBP
Goswami, J. C., & Chan, A. K. (1999). Fundamentals of Wavelets: Theory, Algorithms,
model. In this method, we first decompose the original data into and Applications. Wiley Publishers, pp. 149–152.
multiple layers by the wavelet transform, and every layer has a Hansen, J. V., & Nelson, R. D. (2002). Data mining of time series using stacked
low-frequency signal and high-frequency signal. Then we establish generalizers. Neurocomputing, 43(1–4), 173–184.
Kim, K. J., & Han, I. (2000). Genetic algorithms approach to feature discretization in
a BP neural network model based on the low-frequency signal artificial neural networks for the prediction of stock price index. Expert Systems
which is used to predict the future value respectively. Real data with Applications, 19(2), 125–132.
are used to illustrate the application of the WDBP neural network Li, S. T., & Kuo, S. C. (2008). Knowledge discovery in financial investment for
forecasting and trading strategy through wavelet-based SOM networks. Expert
and show the improved accuracy of using the WDBO model. Systems with Applications, 34, 935–951.
Wavelet transform is an effective data pre-processing tool Li, S. T., & Shue, L. Y. (2004). Data mining to aid policy making in air pollution
which may be combined with other forecasting methods, such as management. Expert Systems with Applications, 27(3), 331–340.
Mörchen. F. (2003). Time series feature extraction for data mining using DWT and DFT.
statistical and other AI models. Investigating these forecasting ap- Department of Mathematics and Computer Science Philips-University Marburg.
proaches could be a future research topic. Technical report, 33.
Oh, K. J., & Kim, K.-J. (2002). Analyzing stock market tick data using piecewise
nonlinear model. Expert System with Applications, 22(3), 249–255.
Acknowledgments Sarantis, N. (2001). Nonlinearities, cyclical behavior and predictability in stock
markets: International evidence. International Journal of Forecasting, 17(3),
The research was supported by the Ministry of Education over- 459–482.
Shen, L., & Loh, H. T. (2004). Applying rough sets to market timing decisions.
seas cooperation ‘‘Chunhui Projects’’ under Grant Z2007-1- 62012
Decision Support Systems, 37, 583–597.
and the Fundamental Research Fund for Physics and Mathematics Thawornwong, S., & Enke, D. (2004). The adaptive selection of financial and
of Lanzhou University under Grant (LZULL200910). The third and economic variables for use with artificial neural networks. Neurocomputing, 56,
corresponding author wishes to thank the support from NSERC 205–232.
Ture, M., & Kurt, I. (2006). Comparison of four different time series methods to
Grant RGPIN197319 of Canada. forecast hepatitis A virus infection. Expert Systems with Applications, 31, 41–46.
Vellido, A., Lisboa, P. J. G., & Meehan, K. (1999). Segmentation of the on-line
References shopping market using neural networks. Expert Systems with Applications, 17,
303–314.
Wang, Y. (2002a). Predicting stock price using fuzzy grey prediction system. Expert
Armano, G., Marchesi, M., & Murru, A. (2005). A hybrid genetic-neural architecture System with Applications, 22, 33–39.
for stock indexes forecasting. Information Sciences, 170(1), 3–33. Wang, Y. (2003). Mining stock prices using fuzzy rough set system. Expert System
Azadeh, A., Ghaderi, S. F., & Sohrabkhani, S. (2008). Annual electricity consumption
with Applications, 24(1), 13–23.
forecasting by neural network in high energy consuming industrial sectors. Wang, Y. F. (2002b). Predicting stock price using fuzzy grey prediction system.
Energy Conversion and Management, 49, 2272–2278.
Expert Systems with Applications, 22(1), 33–39.
Chen, A. S., Leung, M. T., & Daouk, H. (2003). Application of neural networks to an Wang, Y. H. (2009). Nonlinear neural network forecasting model for stock index
emerging financial market: Forecasting and trading the Taiwan stock index. option price: Hybrid GJR-GARCH approach. Expert Systems with Applications, 36,
Computers and Operations Research, 30, 901–923.
564–570.
Chun, S. H., & Kim, S. H. (2004). Data mining for financial prediction and trading: Zhang, G. (2003). Time series forecasting using a hybrid ARIMA and neural network
Application to single and multiple markets. Expert Systems with Applications, 26,
model. Neurocomputing, 50, 159–175.
131–139. Zhang, Y. D., & Wu, L. N. (2009). Stock market prediction of S&P 500 via combination
Coifman, R. R., & Wickerhauser, M. V. (1992). Entropy-based algorithms for best- of improved BCO approach and BP neural network. Expert Systems with
basis selection. IEEE Transactions on Information Theory, 38, 713–718. Applications, 36, 8849–8854.
Enke, D., & Thawornwong, S. (2005). The use of data mining and neural networks for
forecasting stock market returns. Expert Systems with Applications, 29, 927–940.