You are on page 1of 8

Applied Soft Computing 11 (2011) 13881395

Contents lists available at ScienceDirect

Applied Soft Computing


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

A hybrid ANFIS model based on AR and volatility for TAIEX forecasting


Jing-Rong Chang a , Liang-Ying Wei b , Ching-Hsue Cheng c,
a
Department of Information Management, Chaoyang University of Technology 168, Jifong East Road, Wufong Township, Taichung County 41349, Taiwan
b
Department of Information Management, Yuanpei Institute of Science and Technology, No. 306, Yuanpei Street, Hsin Chou 300, Taiwan
c
Department of Information Management, National Yunlin University of Science and Technology, 123, Section 3, University Road, Touliu, Yunlin 640, Taiwan

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

Article history: Time series models have been applied to forecast stock index movements and make reasonably accu-
Received 3 February 2009 rate predictions. There are, however, two major drawbacks of conventional time series models: (1) most
Received in revised form 6 January 2010 conventional time series models use only one variable to forecast; and (2) the rules that are mined from
Accepted 25 April 2010
articial neural networks (ANNs) are not easily understandable. To solve these problems and enhance
Available online 2 May 2010
the forecasting performance of fuzzy time series models, this paper proposes a hybrid adaptive network-
based fuzzy inference system (ANFIS) model that is based on AR and volatility to forecast stock price
Keywords:
problems of the Taiwan stock exchange capitalization weighted stock index (TAIEX). To evaluate fore-
Adaptive network-based fuzzy inference
system (ANFIS)
casting performance, the proposed model is compared with Chens model and Yus model. Our results
Fuzzy time series indicate that the proposed model is superior to other methods with regard to root mean squared error
TAIEX forecasting (RMSE).
Autoregressive 2010 Elsevier B.V. All rights reserved.

1. Introduction (momentum) plays an important role in affecting the volatility of


stock index. Yet, the integration with stock index momentum and
Individual investors, stock fund managers and nancial analysts price forecast is an undeveloped topic in the subject of fuzzy time
attempt to predict price activity in the stock market based on their series. Thus, the t days of momentum of a stock index (denoted
professional knowledge or with the assistance of stock analysis as Mo(t)) is used as an input variable for the proposed forecasting
tools. Higher accuracy is of paramount importance, because greater model.
prots are realized if more accurate predictions are made. There- In recent years, many time series models have been proposed
fore, stock analysts have perennially strived to discover methods and applied to the subject of economic forecasting. Engle [11] pro-
of predicting stock price accurately. According to the efcient mar- posed the ARCH (p) (autoregressive conditional heteroscedasticity)
kets theory, the stock price will represent the optimal forecast of model, which has been used by many nancial analysts; the GARCH
present value [20]. [3] (generalized ARCH) model is the generalized form of ARCH.
But, stock prices change dramatically, which complicates stock Box and Jenkins [4] proposed the autoregressive moving average
volatility forecasting. To address this problem, several technical (ARMA) model, which combines a moving average process with a
indicators have been adopted to forecast stock trends, such as trad- linear difference equation to obtain an autoregressive moving aver-
ing volume, the proportion of short sellers, and various options that age model, and the ARMA model forecasts under linear stationary
are contracted on a stock. Stock analysts have assumed that there conditions. Models that describe such homogeneous nonstationary
are recurring patterns (trends) in the stock market that will appear behavior can be obtained by assuming certain suitable differences
in the future. Therefore, investors use price uctuations, such as of the process to be stationary. Therefore, the autoregressive inte-
chart patterns, to predict outcomes. Therefore, investors use the grated moving average model (ARIMA) [4] with the assumption of
price motions such as chart patterns to predict the future. linearity among the variables was proposed to manage nonstation-
Recently, a technical indicator, momentum, has been used to ary behavior datasets. Based on these literatures, the autoregressive
forecast stock price [16,23]. In research by Tanaka-Yamawaki and (AR) model is the fundamental method in time series models.
Tokuoka [23], the experiment results showed that the momen- However, the traditional time series method requires more his-
tum indicator appeared in each optimum optimal combination of torical data, and the data must form a normal distribution to yield
technical indicators for each stock symbol, based on a genetic algo- a better forecast. Further, linguistic expressions are often used to
rithm. For these reasons, we nd that the volatility of a stock index describe daily observations. Hence, Song and Chissom [21] rst
proposed the original model of the fuzzy time series, and subse-
quent research has focused on the two major processes of the fuzzy
Corresponding author. Tel.: +886 5 5342601x5312; fax: +886 5 531 2077. time series model: (1) fuzzication and (2) establishment of fuzzy
E-mail address: chcheng@yuntech.edu.tw (C.-H. Cheng). relationships and forecasting.

1568-4946/$ see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2010.04.010
J.-R. Chang et al. / Applied Soft Computing 11 (2011) 13881395 1389

In the fuzzication process, the length of intervals for the reliant on expert knowledge [25]. ANFIS has been adopted to make
universe of discourse can affect the forecast, and Huarng [12] pro- implementation in many areas. For example, Ubeyli and Guler [25]
posed distribution-based and average-based lengths to address this presented an ANFIS-based system for the quasi-TEM characteristics
issue. In addition, Chen and Chung proposed a new method [7], in of microshield lines with practical cavity sidewall proles, which
which the length of linguistics intervals is tuned by genetic algo- provides an accurate computation of the characteristic impedance
rithms. In establishing fuzzy relationships and forecasting, Yu [27] and the cavity capacitance sensitivity of the microshield lines. In
argued that recurrent fuzzy relationships should be considered 2008, a novel ANFIS-based system is proposed for modeling where
during forecasts, recommending that different weights should be the number of data pairs employed for training is minimized by
assigned to various fuzzy relationships. Therefore, Yu [27] proposed application of an engineering statistical technique called full fac-
a weighted fuzzy time series method to forecast TAIEX prices. To torial design [5]. By employing this model the number of data
take advantage of neural networks (nonlinear capabilities), Huarng required for learning in the ANFIS network could be signicantly
and Yu [13] chose a neural network to establish fuzzy relation- reduced and thereby computation time as well as computation
ships in fuzzy time series, which are also nonlinear, but the process complexity is remarkably reduced. Besides, Avci [1] also developed
of mining fuzzy logical relationships is not easily understandable, a wavelet packet entropy adaptive network-based fuzzy inference
much like a black box [8]. Yet, these models have been limited to system (WPEANFIS) for classication of the twenty 512 512 tex-
one forecasting variable [9,25]. ture images obtained from Brodatz image album. The experimental
During recent years, many researchers have applied data min- result indicating efciency of WPEANFIS developed method was
ing techniques to nancial analysis. Kimoto et al. [17] developed tested and a mean 93.12% recognition success was obtained.
a system to predict stock market prices by using neural networks. It has been proved that the ANFIS model can be an effective tool
Nikolopoulos and Fellrath [18] combined genetic algorithms and for modeling OC relays. In this section, we briey review related
neural networks to develop a hybrid system for investment advis- studies of subtractive clustering (Subclust) and ANFIS, which are
ing. Kim and Han [15] proposed a genetic algorithmic approach the main relevant model for this study.
to feature discretization and determine connection weights for
articial neural networks (ANNs) to predict stock prices. Roh [19] 2.1. Subtractive clustering
integrated a neural network and a time series model to forecast the
volatility of stock indices. Chiu [10] developed the subtractive clustering, one of the fuzzy
Based on the discussion above, there are two major drawbacks clustering, to estimate both the number and initial locations of clus-
in these models. (1) Most conventional time series models use only ter centers. Consider a set T of N data points in a D-dimensional
one variable in forecasting. Yet, there is tremendous noise that is hyper-space, where each data point Wi (i = 1, 2, . . ., N) Wi = (xi, yi ),
caused by changes in market conditions and environments. There- where xi denotes the p input variables and yi is the output variable.
fore, nancial analysts should consider many market variables in The potential value pi of data point is calculated by Eq. (1):
forecasting. For this reason, forecasting models should use more
variables to improve forecasting accuracy. (2) The ANN is a black 
N
2
box method, and the rules that are mined from articial neural pi = e||Wi Wj || , (1)
networks (ANNs) are not easily understandable [8]. Nevertheless, j=1

the forecasting rules are useful in advising investors to buy or sell


where = 4/r2 , r is the radius dening a Wi neighborhood, and ||||
stocks in the future.
denotes the Euclidean distance.
To overcome these drawbacks, this paper considers that the
The data point with many neighboring data points is chosen as
volatility of the TAIEX plays an important role in inuencing TAIEX
the rst cluster center. To generate the other cluster centers, the
forecasting results, which is adopted into the AR model to build the
potential pi is revised of each data points Wi by Eq. (2):
primary model. Then, the results of the primary model are rened
and optimized by an adaptive network-based fuzzy inference sys- pi = pi p1 exp(||Wi W1 ||2 ), (2)
tem (ANFIS), which uses fuzzy if-then rules to model the qualitative
aspects of human knowledge and can be applicable for human use. where is a positive constant dening the neighborhood which
Based on these concepts, a hybrid ANFIS model is proposed to will have measurable reductions in potential. W1 is the rst cluster
forecast the Taiwan stock index. First, this paper calculates the center and p1 is its potential value.
volatility of the TAIEX and uses the ANFIS to forecast Taiwan stock From Eq. (2), the method selects the data point with the high-
index prices, which combines the AR model and the volatility of est remaining potential as the second cluster center. For general
the TAIEX (momentum) to forecast TAIEX prices (t + 1). Second, we equation, we can rewrite Eq. (2) as Eq. (3).
optimize fuzzy inference system (FIS) parameters using an adaptive pi = pi pk exp(||Wi Wk ||2 ), (3)
network, which can overcome the limitations of statistical methods
(the data need to obey a certain mathematical distribution). Thus, where Wk = (xk , yk ) is the location of the k th cluster center and
we expect that the proposed model will generate greater prots pk is its potential value.
for investors as a result of more accurate forecasting of the TAIEX, As the end of the clustering process, the method obtains q cluster
which is the same concept that has been mentioned by Soros [22]. centers and D corresponding spreads Si , i = (1, .., D). Then, we dene
The rest of this paper is organized as follows. Section 2 describes their membership functions. The spread is calculated according to
related studies; Section 3 briey presents the proposed model; and .
Section 4 describes the experiments and comparisons. Finally, the
conclusions of the study are drawn in Section 5. 2.2. ANFIS: adaptive-network-based fuzzy inference system

Jang [14] proposes adaptive-network-based fuzzy inference sys-


2. Related works tem (ANFIS), which is a fuzzy inference system, implemented in
the framework of adaptive networks. For illustrating the system,
The ANFIS is a fuzzy Sugeno model put in the framework of we assume the fuzzy inference system which consists of ve layers
adaptive systems to facilitate learning and adaptation [14]. Such of adaptive network with two inputs x and y and one output z. The
framework makes the ANFIS modeling more systematic and less architecture of ANFIS is shown as Fig. 1.
1390 J.-R. Chang et al. / Applied Soft Computing 11 (2011) 13881395

Fig. 1. The architecture of ANFIS network.

Then, we suppose that the system consists of two fuzzy if-then 2.6. Layer 4
rules based on Takagi and Sugenos type [23]:
Every node i in this layer is a square node with a node func-
Rule 1: If x is A1 and y is B1 , then f1 = p1 x + q1 y + r1 . tion (see Eq. (8)). Parameters in this layer will be referred to as
Rule 2: If x is A2 and y is B2 , then f2 = p2 x + q2 y + r2 . consequent parameters
O4,i = wi fi = wi (pi + qi + ri ), (8)
The node in the i-th position of the k-th layer is denoted as Ok,i , where pi , qi , ri are the parameters.
and the node functions in the same layer are of the same function
family as described below: 2.7. Layer 5

2.3. Layer 1 The single node in this layer is a circle node labeled that com-
putes the overall output as the summation of all incoming signals
This layer is the input layer and every node i in this layer is a (see Eq. (9)):
 
square node with a node function (see Eq. (4)). Ol,i is the member- wf
ship function of Ai , and it species the degree to which the given O5,i = wi fi = i=1 i = overall output (9)
i=1
wi
x satises the quantier Ai . We select the bell-shaped membership i
function as the input membership function (see Eq. (2)) with maxi-
mum equal to 1 and minimum equal to 0. This is because empirical 3. Proposed model
evidence shows that such distributions in nancial markets are rea-
sonable in times of relatively stationary periods only, but not in From the reviewed literature, there are two major drawbacks,
general, and it is suitable for this paper as mentioned in Section 1: (1) time series models use only one
variable [26]; and (2) the ANN is a black box method, and the rules
O1,i = Ai (x) for i = 1, 2 (4) that are mined from articial neural networks (ANNs) are not easily
understandable [8]. To overcome these drawbacks, the volatility of
1 TAIEX (momentum) will be incorporated into the AR model, and a
Ai (x) = , (5)
2 bi fuzzy inference system that uses fuzzy if-then rules is developed.
1 + [(x ci /ai ) ]
Based on these concepts, this paper proposes a hybrid ANFIS
where ai , bi , ci are the parameters and b is a positive value and c (adaptive network-based fuzzy inference system) model to forecast
denotes the center of the curve. Taiwan stock index prices. First, this paper calculates the volatil-
ity of TAIEX to forecast TAIEX prices (t + 1). Second, we optimize
the fuzzy inference system parameters using the adaptive network.
2.4. Layer 2 Then, the overall owchart of the proposed model is shown as Fig. 2.
This section uses numerical data in a step-by-step example to
Every node in this layer is a square node labeled which mul- demonstrate the core concept of the proposed algorithm.
tiplies the incoming signals and sends the product out by Eq. (6):
3.1. Step1: select dataset
O2,i = wi = Ai (x) Bi (y) for i = 1, 2 (6)
In this step, we choose TAIEX prices from 1997 to 2003 (seven
2.5. Layer 3 sub-datasets) to illustrate the proposed model (such as the year
2000 sub-datasets, which contain 271 transaction days). Each
Every node in this layer is a square node labeled N. The i-th node training dataset encompasses JanuaryOctober, and the remaining
calculates the ratio of the i-th rules ring strength to the sum of all dataset (from November to December) is used for testing.
rules ring strengths by Eq. (7). Output of this layer can be called
normalized ring strengths 3.2. Step 2: estimate TAIEX model

wi To estimate the TAIEX model, we use the E-Views software pack-


O3,i = wi = for i = 1, 2 (7)
w1 + w2 age to t the time series model for different orders of TAIEX over 7
J.-R. Chang et al. / Applied Soft Computing 11 (2011) 13881395 1391

Fig. 3. Testing the order of AR and momentum (Lag/Period test).

3.4.1. Step 4.1: dene and partition the universe of discourse for
input variables by subtractive clustering
First, we dene each universe of discourse for variables accord-
ing to the minimum and maximum values of each variable. Second,
we partition the universe of discourse by subtractive clustering [10]
Fig. 2. Flowchart of the proposed procedure. (Gaussian membership function).

years. The least squares method is used to build the model, and ve 3.4.2. Step 4.2: set the type of membership function for output
linear regression variables (close price (t 1) to close price (t 5)) variable
are selected to be estimated and tested for each year of the TAIEX. In this step, we set linear type membership functions for output
variable. For example, suppose that there are two inputs (TAIEX(t),
3.3. Step 3: set the order of AR and momentum and calculate Mo(t)) and three linguistic intervals that are partitioned by subtrac-
multi-order momentum tive clustering in each input variable. Therefore, a typical rule in a
Sugeno fuzzy model is described as follows:
From Step 2, the estimating TAIEX model is obtained, and its If x(TAIEX(t)) = Ai and y (Mo(t)) = Bi then fi = pi x + qi y +
results provide the reference resources to determine the order of ri ,where x(TAIEX(t)), y(Mo(t)), are linguistic variables, Ai and
AR. From the reviewed literature, Tanaka-Yamawaki and Tokuoka Bi , are the linguistic values (high, middle, low), fi denotes the i-th
[24] used one-order and two-order momentum to predict stock output value, pi , qi , ri , are the parameters (i = 1, 2, 3).
prices; thus, we also set the order of momentum from one to two.
By testing the order of AR and momentum (Lag/Period test, see
Fig. 3), the proposed models will be set. For example, if the order Table 1
for the lag periods of AR is changed from one to two and the order The one-order and two-order momentum of TAIEX.
of momentum is raised from one to two, then there are four types Date TAIEX(t + 1) Mo(t) Mo(t 1)
of forecasting models: (AR(1) Mo(1)), (AR(1) Mo(2)), (AR(2) Mo(1)),
2000/1/4 8756.55
and (AR(2) Mo(2)). To calculate multi-order momentum, we dene 2000/1/5 8849.87
two volatility variables (Mo(t) and Mo(t 1)) and calculate them 2000/1/6 8922.03 93.32
by Eqs. (10) and (11). Table 1 lists the values of Mo(t) and Mo(t 1) 2000/1/7 8845.47 72.16 93.32
between 2000/1/4 and 2000/1/31. 2000/1/10 9102.60 76.56 72.16
2000/1/11 8927.03 257.13 76.56
Mo(t) = TAIEX(t) TAIEX(t 1) (10) 2000/1/12 9144.65 175.57 257.13
2000/1/13 9107.19 217.62 175.57
Mo(t 1) = TAIEX(t 1) TAIEX(t 2), (11) 2000/1/14 9023.24 37.46 217.62
2000/1/15 9191.37 83.95 37.46
where Mo(t) is the momentum between time t and t 1, and 2000/1/17 9315.43 168.13 83.95
Mo(t 1) is the momentum between time t 1 and t 2. 2000/1/18 9250.19 124.06 168.13
2000/1/19 9151.44 65.24 124.06
2000/1/20 9136.95 98.75 65.24
3.4. Step 4: generate ANFIS forecast model 2000/1/21 9255.94 14.49 98.75
2000/1/24 9387.07 118.99 14.49
2000/1/25 9372.37 131.13 118.99
The ANFIS forecast model will be introduced in Steps 4.14.3,
2000/1/26 9581.96 14.70 131.13
and the ANFIS method uses subtractive clustering [10] mentioned 2000/1/27 9628.98 209.59 14.70
in Section 2.1 to partition the universe of discourse for input vari- 2000/1/28 996.91 47.02 209.59
ables and generate the fuzzy inference system. The sub-steps of 2000/1/29 9636.38 67.93 47.02
Step 4 are described as follows: 2000/1/31 9744.89 60.53 67.93
1392 J.-R. Chang et al. / Applied Soft Computing 11 (2011) 13881395

Table 2
The performance comparisons of different models (TAIEX).

Models Year

1997 1998 1999 2000 2001 2002 2003

Yus model [27] 165 164 145 191 167 75 66


Chens model [6] 154 134 120 176 148 101 74
Proposed model AR(1) Vo(1) 133a 117a 100 173 119 61a 53a
Proposed model AR(1) Vo(2) 136 118 97a 149 121 62 53a
Proposed model AR(2) Vo(1) 133a 119 103 206 125 62 53a
Proposed model AR(2) Vo(2) 135 119 102 146a 116a 62 57
a
The best performance among six models.

3.4.3. Step 4.3: generate fuzzy inference system 4.1. Building the TAIEX model
First, the linguistic intervals, as input membership functions, are
obtained in Step 4.1, and the output membership functions are set Based on statistical tests, the least squares method is used to
in Step 4.2. Then, the ANFIS model generates fuzzy if-then rules, build the model. Use the E-Views software package to t time series
in which the linguistic values (Ai , Bi , Ci ) from input membership models for different lags and orders of the TAIEX over 7 years. In
functions are used as the if-condition component and the output each year of the TAIEX, ve linear regression variables (i.e., clos-
membership function (fi ) is the then component. ing price (t 1) to closing price (t 5)) are selected to be estimated
and tested. If the p-value is less than the signicance level of 0.05,
3.5. Step 5: train fuzzy inference system parameters from training the null hypothesis is rejected. Take the TAIEX in 1997 as an exam-
datasets ple (see Fig. 3)the p-value (0.0000) for the closing price (t 1) is
less than the signicance level (0.05) with regard to the four vari-
In this step, we use the least squares method and the back propa- ables (close price (t 1) to close price (t 5)). Further, the variable
gation gradient descent method to train the forecasting model. This close price (t 1) is not equal to zero. Therefore, the order of AR
study sets the epoch to 50 (the process is executed for a predeter- is one.
mined xed number (50) of iterations unless it terminates while the
training error converges) as the criterion for stopping and obtains
the parameters for the selected output membership function. 4.2. Performance comparison

3.6. Step 6: forecast testing datasets by the trained model From Section 4.1, we see that the order of AR for the TAIEX
model is one. In this study, we set the order for the AR model
The FIS parameters of the forecasting models are determined from one to two, and the order for the momentum is changed from
when the criterion for stopping from Step 5 is reached; then, the one to two. Therefore, we obtain four types of forecasting mod-
training forecasting model is used to forecast TAIEX(t + 1) for the els: (1) the one-order AR with one-order momentum model (AR(1)
target datasets for testing. Mo(1)); (2) the one-order AR with two-order momentum model
(AR(1) Mo(2)); (3) the two-order AR with one-order momentum
3.7. Step 7: calculate RMSE and compare with the other models model (AR(2) Mo(1)); and (4) the two-order AR with two-order
momentum model (AR(2) Mo(2)). After the experiments, we gen-
Calculate RMSE values in the tested datasets by Eq. (12). Then, erate a forecast for the seven tested sub-datasets in each model.
the RMSE is used as an evaluation criterion to compare with other Then, this paper compares the performance of the proposed model
models. with the conventional fuzzy time series modelChens [6] model.
 Furthermore, to judge whether the proposed method is superior
n 2
(actual(t) forecast(t)) to the latest fuzzy time series modelthe weighted fuzzy time
t=1
RMSE = , (12) seriesYus model [27] is compared with the proposed procedure.
n
The forecasts of Chens model, Yus model, and the proposed
where actual(t) denotes the real TAIEX value, forecast(t) denotes model are listed in Table 2, wherein the proposed model outper-
the predicted TAIEX value, and n is the number of data. forms the other models. For advanced model validity, the yearly
predicted values of the TAIEX index (from 1997 to 2003), based on
4. Experiments and comparisons Chens, Yus, and the proposed models, are shown in Figs. 410 in
Appendix.
This section describes the contents of the estimating TAIEX On comparison of the four proposed models, the AR(1) Mo(1)
model and compares its performance. To verify the proposed model outperforms the other models, demonstrating that the pro-
model, we use TAIEX prices from 1997 to 2003 (including seven posed model reduces forecasting errors more effectively when the
sub-datasets) to execute the experiment. The sub-datasets for the order of AR and momentum are equal to one. Further, these results
rst 10 months in these 7 years are used for training, and those replicate the statistical test results in Section 4.1, in which the order
from November to December are selected for testing. of AR is one.

Table 3
The prots comparisons of different models (TAIEX).

Year Models

Yus model [27] Chens model [6] Proposed model Proposed model Proposed model Proposed model
AR(1) Mo(1) AR(1) Mo(2) AR(2) Mo(1) AR(2) Mo(2)

1997 0.02 107 127 306 247 352 752a


a
The best prots among three models.
J.-R. Chang et al. / Applied Soft Computing 11 (2011) 13881395 1393

For making some advanced simulation trades and showing the The optimal threshold parameter is obtained when the fore-
prots, we set two trade rules by the Taiwan Futures Exchange casting performance reaches best prots in the training dataset.
(TAIFEX) and use the two trade rules to calculate prots, and From the optimal threshold parameter and Eq. (13), the prots
assume that the prots unit is equal to one. So the prot formula is for different models are calculated and the prots results are shown
dened as Eq. (13). in Table 3.

Rule 1: sell rule 5. Conclusions

|forecast(t) actual(t)| A new fusion ANFIS model based on an AR model and volatility of
If
actual(t) the TAIEX (momentum) has been developed to forecast stock price
and forecast(t + 1) actual(t) > 0 then sell problems in Taiwan; furthermore, the proposed model is compared
with two recent models, Chens model [6] and Yus model [27], to
Rule 2: buy rule evaluate its performance. After verication and comparison, the
proposed method outperforms the other methods.
|forecast(t) actual(t)|
If
actual(t) (1) According to Table 2, it is evident that the proposed model
and forecast(t + 1) actual(t) < 0 then buy is superior to the other methods with regard to the RMSE,
because the proposed model takes the AR model and volatil-
where denotes threshold parameter (0 < < 0.07, the threshold ity of TAIEX with ANFIS learning into account to forecast TAIEX
parameter depends on daily uctuation of TAIEX). prices. Besides, in Table 3, the prots for different models also
Prot denition convince us that the proposed is outperforms the other meth-
ods.

p

Prot = (actual(t + 1) actual(t)) (2) From Table 2, the performance of the one-order AR model is
better than the two-order models. Clearly, the one-order AR
ts =1
model reduces forecasting errors more effectively than the two-

q order adaptive model. Also, these results are the same as the
+ (actual(t) actual(t + 1)), (13) statistical test results in Section 4.1, and the order of AR is one.
tb =1 However, this is consequence of the TAIEX data and not of the
stock market, so cannot generalize.
where p represents the total number of days for selling, q repre-
sents the total number of days for buying, ts represents the t-th The proposed model primarily uses input variables (different-
day for selling, and tb represents the t-th day for buying. order AR model and different-order momentum) to forecast the

Fig. 4. The original and predicted values of TAIEX index between 11/4/1997 and 12/31/1997.

Fig. 5. The original and predicted values of TAIEX index between 11/3/1998 and 12/31/1998.
1394 J.-R. Chang et al. / Applied Soft Computing 11 (2011) 13881395

TAIEX on the next trading day. To develop the proposed model, investors. In addition, traditional fuzzy time series models always
practical collected stock index datasetsfrom the TAIEX (between use the closing price of a stock index as input data for verication.
1997 and 2003)were used in this empirical experiment. The 7- In fact, different time horizons of forecasts, such as one minute and
year TAIEX datasets were spilt into seven sub-datasets based on one hour, are also an interesting future direction. The proposed
year. Datasets that comprise the rst 10 months (JanuaryOctober) model can be applied to make stock index forecasts in different
are used for training, and datasets that encompass the last 2 time horizons.
months, November and December, are used to test each sub-
datasets. Acknowledgements
In future work, the proposed model can be applied to other stock
markets, such as in China, Japan, and Hong Kong. Other impor- The authors would like to thank the anonymous referees for
tant technology indicators could be incorporated into the proposed providing very helpful comments and suggestions. Their insight
model to enhance accuracy, and different rule-based articial intel- and comments led to a better presentation of the ideas expressed
ligence techniques can be used to generate useful decision rules for in this paper.

Fig. 6. The original and predicted values of TAIEX index between 11/2/1999 and 12/28/1999.

Fig. 7. The original and predicted values of TAIEX index between 11/2/2000 and 12/30/2000.

Fig. 8. The original and predicted values of TAIEX index between 11/2/2001 and 12/31/2001.
J.-R. Chang et al. / Applied Soft Computing 11 (2011) 13881395 1395

Fig. 9. The original and predicted values of TAIEX index between 11/4/2002 and 12/31/2002.

Fig. 10. The original and predicted values of TAIEX index between 11/4/2003 and 12/31/2003.

Appendix A. [13] K.H. Huarng, T.H.K. Yu, The application of neural networks to forecast fuzzy
time series, Physica A 336 (2006) 481491.
[14] J.S. Jang, ANFIS: Adaptive-Network-based Fuzzy Inference Systems, IEEE Trans-
See Figs. 410. actions on Systems, Man, and Cybernetics 23 (3) (1993) 665685.
[15] K. Kim, I. Han, Genetic algorithms approach to feature discretization in articial
neural networks for prediction of stock index, Expert System with Applications
References 19 (2000) 125132.
[16] M.-J. Kim, S.-H. Min, I. Han, An evolutionary approach to the combination of
[1] E. Avci, Comparison of wavelet families for texture classication by using multiple classiers to predict a stock price index, Expert Systems with Appli-
wavelet packet entropy adaptive network based fuzzy inference system, cations 31 (2006) 241247.
Applied Sost Computing 8 (2008) 225231. [17] T. Kimoto, K. Asakawa, M. Yoda, M. Takeoka, Stock market prediction system
[3] T. Bollerslev, Generalized autoregressive conditional heteroscedasticity, Jour- with modular neural network, in: In Proceedings of the international joint
nal of Econometrics 31 (1986) 307327. conference on neural networks, San Diego, California, 1990, pp. 16.
[4] G. Box, G. Jenkins, Time series analysis: Forecasting and control, Holden-Day, [18] C. Nikolopoulos, P. Fellrath, A hybrid expert system for investment advising,
San Francisco, 1976. Expert Systems 11 (4) (1994) 245250.
[5] M. Buragohain, C. Mahanta, A novel approach for ANFIS modelling based on full [19] T.H. Roh, Forecasting the volatility of stock price index, Expert Systems with
factorial design, Applied Sost Computing 8 (2008) 609625. Applications 33 (2007) 916922.
[6] S.M. Chen, Forecasting enrollments based on fuzzy time-series, Fuzzy Sets Sys- [20] R.J. Shiller, From Efcient Markets Theory to Behavioral Finance, Journal of
tems 81 (1996) 311319. Economic Perspectives 17 (1) (2003) 83104.
[7] S.M. Chen, N.Y. Chung, Forecasting Enrollments Using High-Order Fuzzy Time [21] Q. Song, B.S. Chissom, Forecasting enrollments with fuzzy time-series Part I,
Series and Genetic Algorithms, International of Intelligent Systems 21 (2006) Fuzzy Sets and Systems 54 (1993) 110.
485501. [22] G. Soros, The Alchemy of Finance, John Wiley & Sons Inc., 2003.
[8] T.L. Chen, C.H. Cheng, H.J. Teoh, High-order fuzzy time-series based on multi- [23] T. Takagi, M. Sugeno, Derivation of fuzzy control rules from human operators
period adaptation model for forecasting stock markets, Physica A 387 (2008) control actions, in: in Proc. IFAC Symp. Fuzzy Inform., Knowledge Representa-
876888. tion and Decision Analysis, 1983, pp. 5560.
[9] C.H. Cheng, J.R. Chang, C.A. Yeh, Entropy-based and trapezoid fuzzication- [24] M. Tanaka-Yamawaki, S. Tokuoka, Adaptive use of technical indicators for the
based fuzzy time series approaches for forecasting IT project cost, Technological prediction of intra-day stock prices, Physica A 383 (2007) 125133.
Forecasting and Social Change 73 (2006) 524542. [25] E.D. Ubeyli, I. Guler, Adaptive neuro-fuzzy inference system to compute
[10] S.L. Chiu, Fuzzy model identication based on cluster estimation, Journal of quasi-TEM characteristic parameters of microshield lines with practical cavity
Intelligent and Fuzzy Systems 2 (1994) 267278. sidewall proles, Neurocomputing 70 (2006) 296304.
[11] R.F. Engle, Autoregressive conditional heteroscedasticity with estimator of the [26] H.K. Yu, K.H. Huarng, A bivariate fuzzy time series model to forecast the TAIEX,
variance of United Kingdom ination, Econometrica 50 (4) (1982) 9871008. Expert Systems with Applications 34 (2008) 29452952.
[12] K.H. Huarng, Effective lengths of intervals to improve forecasting in fuzzy time [27] H.K. Yu, Weighted fuzzy time-series models for TAIEX forecasting, Physica A
series, Fuzzy Sets and Systems 123 (2001) 155162. 349 (2005) 609624.

You might also like