Computers & Industrial Engineering 99 (2016) 153161

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

Taiwanese export trade forecasting using firefly algorithm based

K-means algorithm and SVR with wavelet transform
R.J. Kuo a,, P.S. Li b
a
Department of Industrial Management, National Taiwan University of Science and Technology, No. 43, Section 4, Kee-Lung Road, Taipei 106, Taiwan
b
AU Optronics Corporation, No. 1 Jhong-Ke Road, Central Taiwan Science Park, Taichung 407, Taiwan

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

Article history: In order to develop a prediction system for export trade value, this study proposes a three-stage forecast-
Received 8 March 2016 ing model which integrates wavelet transform, firefly algorithm-based K-means algorithms and firefly
Received in revised form 9 July 2016 algorithm-based support vector regression (SVR). First, wavelet transform is utilized to reduce the noise
Accepted 12 July 2016
in data preprocessing. Then, the firefly algorithm-based K-means algorithm is employed for cluster anal-
Available online 14 July 2016
ysis. Finally, a forecasting model is built for each cluster individually. For evaluation, this study compares
methods with and without clustering. In addition, both non-wavelet transform and wavelet transform for
Keywords:
data preprocessing are investigated. The experimental results indicate that the forecasting algorithm
Cluster analysis
Firefly algorithm
with both wavelet transform and clustering has better performance. Besides, firefly algorithm-based
Support vector regression SVR outperforms the other algorithms.
Forecasting 2016 Elsevier Ltd. All rights reserved.
Wavelet transform

1. Introduction
Because of Taiwans relatively small domestic market and the
nations lack of resources, Taiwans economic development counts
heavily on international trade. Thus, being able to make forecasts
about export trade is a very important issue. In recent years,
because of the innovation of forecasting techniques and improve-
ment in forecasting accuracy, forecasting methodology is neces-
sary for enhancing the decision-making process in both industry
and government. However, according to previous researches on
export trade forecasting, most forecasting methods usually con-
struct a traditional model or fuzzy time series model as a tool for
forecasting future value and data analysis related to export trade
(Tu, Hsien-Lun, & Chi-Chen, 2009; Wang, 2011; Wong, Tu, &
Wang, 2010). Therefore, other intelligent methods may be
employed in order to obtain greater accuracy in forecasting.
Currently, the support vector regression (SVR) model is being
widely used in many fields to analyze time series forecasting, for
example tourism demand (Chen, 2011; Pai, Hung, & Lin, 2014;
Shahrabi, Hadavandi, & Asadi, 2013), traffic flow (Castro-Neto,
Jeong, Jeong, & Han, 2009; Hong, Dong, Chen, & Wei, 2011; Hong,
Dong, Zheng, & Lai, 2011; Li, Hong, & Kang, 2013) and demand fore-
casting (Guanghui, 2012; Lu & Wang, 2010; Wu, 2010). Besides,
Corresponding author.
E-mail addresses: rjkuo@mail.ntust.edu.tw (R.J. Kuo), karenli127@gmail.com
(P.S. Li).
SVR usually derives more precise results than the other methods.
However, determining the SVR parameters is very important for
establishing the SVR model. Some metaheuristics have been
employed to determine SVR parameters, like genetic algorithm
(GA) (Cai, Sheng, & Xiao-bin, 2009; Chen & Wang, 2007; Fang,
Wang, Qi, & Zheng, 2008; Ju & Wu, 2010), particle swarm optimiza-
tion algorithm (PSO) (Chen & Liu, 2013; Jiansheng & Enhong, 2010;
Siamak, Bahrami Jovein, & Ramezanianpour, 2012), differential
algorithm (DE) (Cai, Qu, & Li, 2013; Li & Cai, 2008; Pan, Cheng, &
Ding, 2013; Wang, Li, Niu, & Tan, 2012) and firefly algorithm (FA)
(Kavousi-Fard, Samet, & Marzbani, 2014; Kazem, Sharifi, Hussain,
Saberi, & Hussain, 2013; Xiong, Bao, & Hu, 2014). In addition, a
two-stage forecasting model was developed based on SVM and
metaheuristic which brings a good predictive value than other
machine learning model (Olatomiwa et al., 2015).
This study intends to present a novel three-stage forecasting
model which integrates wavelet transform, FA-based K-means
algorithms and the FA-based SVR model. Wavelet transform is first
applied to reduce the noise in data preprocessing. The FA-based K-
means algorithm is then utilized for cluster analysis. Finally, the
FA- based SVR model is used to develop the forecasting model
for each cluster individually.
The remainder of this study is organized as follows. Section 2
presents the SVR model, while the proposed three-stage forecast-
ing model is proposed in Section 3. Section 4 shows the experimen-
tal results for export forecasting. Finally, the concluding remakes
are offered in Section 5.

http://dx.doi.org/10.1016/j.cie.2016.07.012
0360-8352/ 2016 Elsevier Ltd. All rights reserved.
154 R.J. Kuo, P.S. Li / Computers & Industrial Engineering 99 (2016) 153161

2. Support vector regression where e is a precision parameter representing the radius of the tube
located around the regression function f x. The region enclosed by
Support vector machine (SVM) is an artificial intelligence the tube is called the e-sensitive zone since the loss function
method which was first developed by Vapnik (1995). SVM is based assumes a zero value in this region and the prediction errors with
on structural risk minimization (SRM) principle which aims to a value smaller than e are not penalized (Lu & Wang, 2010).
minimize an upper bound of generation error; it consists of the Furthermore, w and b are estimated by minimizing the follow-
sum of training error and a confidence interval (Guo, Sun, Li, & ing optimization problem:
Wang, 2008). In Vapnik (1999, 1995) promoted the SVM method 1
to support vector regression (SVR) by using a new type of loss func- Minimize kwk2 3
2
tion called e-insensitive loss function which is used to penalize

yi  w; uxi b 6 e
errors as long as they are greater than e; it is assumed that e is Subject to
known beforehand (Guo et al., 2008). SVR is a non-linear kernel-
w; uxi b  yi 6 e
based regression method which seeks to locate the best regression In addition, to deal with feasibility issues and to make the
hyperplane with the smallest risk in high dimensional feature method stronger, points forming the e-sensitive band are not elim-
space (Yeh, Huang, & Lee, 2011). inated. Instead, we penalize these points by introducing slack vari-
The SVM model consists of both classification and regression ables ni and ni (Kazem et al., 2013). The SVR minimizes the overall
tasks, it has many advantages, including: a global optimal solution errors as follows:
can be found, the result is a general solution avoiding overtraining,
1 X
and nonlinear solutions can be calculated efficiently due to the Minimize kwk2 C ni ni 4
usage of inner products (Thissen, van Brakel, de Weijer, Melssen, 2 i1
8
& Buydens, 2003). The SVM technique has been used in a range > y
< i  w; uxi b 6 e ni
of applications, including financial stock market prediction (Ding,
Subjected to w; uxi b  yi 6 e ni
Song, & Zen, 2008; Huang, Nakamori, & Wang, 2005; Kim, 2003; >
:
Tay & Cao, 2001, 2002), electric load and price (Che & Wang, ni; ni P 0 for i 1; . . . ; l
2010; Pai & Hong, 2005; Yan & Chowdhury, 2014) and sales fore- The first term of Eq. (4), having the concept of maximizing the
casting (Wu, 2009). distance of two separated training data, is used to regularize
It is assumed that a set of training data which consists of data weight sizes, penalize large weights and maintain regression func-
points fx1 ; y1 ; . . . ; x ; y , i = 1, 2 , . . . , and xi denotes the space of tion flatness. The second term is to penalize training errors of f x
the input pattern, and yi is the corresponding target value. The and y by using the e-sensitive loss function (Hong, Dong, Chen,
input is first mapped onto an n-dimension feature space by ux, et al., 2011; Hong, Dong, Zheng, et al., 2011). C is a modifying coef-
which is a kernel function, to transform the non-linear input into ficient representing the trade-off between empirical risk and struc-
a linear mode in a high dimensional feature space (Lu, 2014). The ture risk. The optimum value of each parameter can be solved by
goal of SVR is to estimate a function f x that is as accurate as pos- Lagrange with an appropriate modifying coefficient C, band area
sible to get the target value yi for every xi (Guo et al., 2008). The width e and kernel function K (Kao, Chiu, Lu, & Chang, 2013). The
function f is represented using a linear function, as in Eq. (1): general form of the SVR-based regression function can be written
f x w  ux b; 1 as Eq. (5):

where f x denotes the forecasting value, w is a vector of weight X

f x; w ai  ai Kxi ; x b; 5
coefficients, b is a bias constant, ux is a nonlinear mapping from i1
the input space to the feature space and w  ux describes the
dot production in the feature space. In SVR, the problem of nonlin- where ai ; ai are nonzero Lagrangian multipliers and the solution
ear regression in the lower dimension input space (x) is trans- for the dual problem. Any function that meets Mercers condition
formed into a linear regression problem in a high dimension can be used as the kernel function. Although several options for
feature space (Fig. 1(a) and (b)) (Lu & Wang, 2010). the kernel function are available, the most widely used kernel
The robust e-sensitive loss function (Le ) (Fig. 1(c)) given below function is the radial basis function (RBF) which is defined as
is the most commonly used equation: follows:

!
jf x  yj  e if jf x  yj P e kxi  xj k2
Le f x; y 2 Kxi ; xj exp  ; 6
0 otherwise 2r 2

Fig. 1. Transformation process illustration of an SVR model.

 R.J. Kuo, P.S. Li / Computers & Industrial Engineering 99 (2016) 153161 155

where r denotes the width of the RBF. SVR has the best perfor-
mance in most forecasting programs as the value of r is set
between 0.1 and 0.5 (Cherkassky & Ma, 2004).
3. The proposed three-stage forecasting model
The framework of the proposed three-stage forecasting model is
illustrated in Fig. 2. Firstly, the raw data need to be preprocessed in
order to delete the missing value or detect the noise by implement-
ing wavelet transform and performing normalization. Secondly,
FA-based K-means algorithms are built to cluster the data with
similar features; therefore, the best one will be chosen. For the
next step, the best algorithm is applied to cluster the export trade
datasets. Furthermore, the data for each cluster are employed to
construct the forecasting models, including: GA-SVR, PSO-SVR,
FA-SVR and DE-SVR. Finally, this study uses MSE as the perfor-
mance indicator.
3.1. Data preprocessing using wavelet transform
Data preprocessing plays a crucial role in data mining. It can
transform the data into a specified format and delete noise or out-
liers; the data we get can thereby be more meaningful for analysis.
This study employs wavelet transform to transform the time series
data first; data normalization is then implemented.
Wavelet analysis indicates a time-scale representation of the
signal; it can provide information about both the time and fre-
quency domains (Tan & Pedersen, 2009). At different resolution
levels, time series data can be decomposed into some components
by using the wavelet function, which is called the mother wavelet
(Tiwari & Chatterjee, 2010). In addition, wavelet transform can
detect many properties of a time series that may not be revealed
by other signal analysis techniques, such as trends, change points,
self-similarity and discontinuities (Nalley, Adamowski, & Khalil,
2012). This step is very important at the beginning of implement-
ing forecasting system, since selecting proper input parameters is
the key issue to influence the forecasting accuracy (Mojumder,
Ong, Chong, Shamshirband, & Abdullah Al, 2016). In the other
way, different feature selection techniques for choosing more sig-
nificant data will also reach more accurate prediction easily (Hu,
Bao, Chiong, & Xiong, 2015; Chang, Wu, & Lin, 2016). Furthermore,
discrete wavelet transform (DWT) can detect and decompose sig-
nals on all scales, and requires less computation time (Partal &
Kk, 2006). The discrete wavelet transform is as follows (Qi &
Yan, 2009):
!
1 t  ks0 s0j
wj;k t q w ; 7
s0j s0j
dilation factor, k is an integer that determines the translation, s0 is a
specified fixed dilation step greater than 1 and s0 denotes the loca-
tion parameter. The wavelet coefficient W w j; k for DWT is defined
as follows:
 
1X N1
t
W w j; k x t w  k ; 8
2j
j
22 t0
where the wavelet coefficient is evaluated at scale s 2 j and loca-
tion c 2 j k to explain the variation of signals at different scales and
locations (Partal & Kk, 2006).
DWT operates two sets of function viewed as high-pass and
low-pass filters for wavelet decomposition. The original time series
are passed through high-pass and low-pass filters and separated at
different scales. In the decomposition phase, the low-pass filter
removes the higher frequency components of the signal, while
the high-pass filter picks the remaining parts. Then, the filtered sig-
nals are down-sampled by two and the results are called approxi-
mation coefficients (cA) and detail coefficients (cD). At level 1, the
cA1 contains the general trend or low frequency components of the
signal f t, and the cD1 is associated with the high frequency com-
ponents of the signal f t (Kao et al., 2013).
Given a signal X of length N, the DWT consists of log 2 N stages
at most. The length of each filter is equal to 2N. If n = length(s), the
signals F and G are of length n + 2N  1 and the coefficients cA1 and
 
cD1 are of length floor n1 2
N.
3.2. Firefly algorithm-based K-means clustering (FAK)
The Firefly Algorithm (FA) is a recent technique of natural inspi-
ration that has been utilized for solving nonlinear optimization
problems. In the FA, only the distance is necessary for the move-
ment (Senthilnath, Omkar, & Mani, 2011). In Senthilnaths study
(2011), the result showed that FA-based K-means algorithm
enhanced the performance with PSO-based K-means and artificial
bee colony-based K-means clustering. The detailed procedure of
FA-based K-means algorithm is as follows:
Step 1: Parameters setup
The parameters include the size of population (N), light absorp-
tion coefficient (c), attractiveness value (b0 ), trade-off constant (a)
and k clusters.
Step 2: Initialization
Generate N fireflies (X id ) as the clusters centroids, which can be
represented as:

X id zi1 ; . . . ; zij ; . . . zik ; 9

where w is the transforming function called the mother wavelet, t is
the time and s is the translation factor of the wavelet over the time where zij denotes the jth clusters centroids for the ith chromosome
series, s denotes the scale factor, j is an integer that determines the (1 5 i 5 N) and (1 5 j 5 k).

Forecasting
Cluster 1 (FA-SVR)

Wavelet Cluster
Database Normalization analysis
transform Forecasting
(FAK) Cluster 2 (FA-SVR)

Cluster n Forecasting
(FA-SVR)

Fig. 2. The framework of the proposed forecasting model.

156 R.J. Kuo, P.S. Li / Computers & Industrial Engineering 99 (2016) 153161

Step 3: Fitness calculation other soft computing techniques, FA-based machine learning shows
high estimated accuracy (Soltani, Moghaddam, Karim,
Calculate the fitness value for each particle by using Euclidean Shamshirband, & Sudheer, 2015). The FA is used to determine the
distance. best parameter set within the user-defined number of iterations
for training an optimal SVR forecasting model.
Step 4: Attractiveness FA-SVR is a hybrid method with an integrated firefly algorithm
and support vector regression. FA is able to assist machine learning
In the firefly algorithm, it is assumed that the attractiveness of a model in selection process effectively and can enhance the fore-
firefly is determined by its brightness, which is associated with the casting capability more accurately (Hoang & Pham, 2016). Further-
objective function of the optimization problem. The attractiveness more, FA is a simple, powerful global search technique so that
of a firefly can be formulated as follows: researchers can have promising setting parameters for SVR. The
cr 2ij
detailed FA-SVR procedure is as follows.
b b0 e ; 10
Step 1: Parameters initialization
where b0 is the attractiveness value at r = 0, c is the light absorption
coefficient at the source; and r ij is the Euclidean distance between xi
Initialize the parameters (N, c, b0 , a). The parameters are the
and xj :
size of the population of FA algorithm (N), light absorption coeffi-
v
u d cient (c), attractiveness value (b0 ), trade-off constant (a) and the
uX
r ij kxi  xj k t xi;k  xj;k 2 ; 11 number of decision variables (D). D is set as 2 due to C and r.
k1
Step 2: Population initialization
Step 5: Firefly movement
Create an N D matrix with uniform probability distribution
The movement of a firefly i, which is attracted to another more random values. The generation method is defined as:
attractive firefly j, is determined by:
  xi;j rand  highj  lowj lowj; 14
cr 2ij 1
xnew
i xold
i be xold
j  i
xold a rand  ; 12 where i = 1, 2, 3, . . . , N, rand is a random number with a uniform
2
probability distribution; high[j] and low[j] are the upper and lower
where a determines the random behavior of movement. bounds of jth column, respectively. The chromosome x is repre-
sented as x = p1 and p2, where p1 and p2 are C and r, respectively.
Step 6: Euclidean distance calculation
Step 3: Training phase
Calculate the Euclidean distance of all the x data vectors to all
clusters centroids for each firefly. In K-fold cross validation, the training data set is divided into K
subsets. The regression function is built with a given set of param-
Step 7: Assign new cluster eters (C, r), using the K-1 subset as the training set. The fitness
function is defined as the MSE of the 5-fold cross validation
Assign each data vector x to the closest cluster centroid. method on the training dataset, as in Eqs. (15) and (16):
Min f MSEcross validation ; 15
Step 8: New centroids Pl
jai  pi j2
MSEcross validation i1 ; 16
Recalculate the cluster centroids vector for each firefly by using: l
1X
Cennew
d xid ; 13 Table 1
nd Average MSE result for B1 dataset.

where Cennew
d is the new centroid vectors of cluster dth, and nd is the Methods
Original Wavelet Clustering
Wavelet-cl
size of cluster d, xid is ith data in cluster d. Models ustering
training 0.000360 0.000270 0.000293 0.000247
GA-SVR
Step 9: Check if the number of iterations is satisfied testing 0.000568 0.000388 0.000485 0.000289
training 0.000355 0.000265 0.000274 0.000248
Stop if the number of iterations is satisfied; otherwise, go back PSO-SVR
testing 0.000515 0.000364 0.000478 0.000293
to Step 3.
training 0.000349 0.000251 0.000273 0.000247
FA-SVR
testing 0.000462 0.000340 0.000424 0.000224
3.3. Forecasting stage
training 0.000468 0.000266 0.000289 0.000246
DE-SVR
This study uses the radial basis function (RBF) for the kernel testing 0.000655 0.000394 0.000449 0.000216
function of the FA- based SVR model. The kernel function is used training 0.000392 0.000294 0.000302 0.000292
to construct a nonlinear decision hyper-surface on the SVR input ARIMA
testing 0.000588 0.000448 0.000475 0.000355
space. It is popular as the accuracy of SVR depends on the choice of
training 0.002444 0.002356 0.000635 0.000572
parameters. The three parameters are regulation constant C, loss BPN
testing 0.002505 0.002153 0.000802 0.000663
function e, and r which is the parameter for the width of RBF. These
parameters are important in regard to the accuracy of SVR forecast- training 0.004415 0.001419 0.002733 0.000609
Regression
ing; this study fixes the e and combines with the metaheuristics testing 0.004509 0.004213 0.002273 0.001890
algorithms for tuning a better parameter set, C and r, to generate Dark gray color means that it is the best method for training data.
the minimum forecasting mean square error (MSE). Comparing with Light gray color means that it is the best method for testing data.
 R.J. Kuo, P.S. Li / Computers & Industrial Engineering 99 (2016) 153161 157

where l is the number of training data samples; ai is the actual value cr 2ij
b b0 e ; 17
and pi is the predicted value.
where b0 is the attractiveness value at r = 0, c is the light absorption
coefficient at the source; and rij is the Euclidean distance between xi
Step 4: Fitness evaluation
and xj :
v
u d
Input the generated populations. Calculate and record the fit- uX
ness values of all individuals by Eqs. (15) and (16). The solution
rij kxi  xj k t xi;k  xj;k 2 ; 18
k1
with a smaller MSE of the training dataset has a smaller fitness
value; therefore, it has a better chance to survive in the next Step 6: Firefly movement
generation.
The movement of a firefly i, which is attracted to another more
Step 5: Attractiveness attractive firefly j, is determined by:

In the firefly algorithm, it is assumed that the attractiveness of a

firefly is determined by its brightness, which is associated with the
objective function of the optimization problem. The attractiveness Table 4
Average MSE result for European dataset.
of a firefly can be formulated as follows:
Methods Wavelet-cl
Original Wavelet Clustering
Table 2 Models ustering
Average MSE result for B2 dataset.
training 0.006342 0.005961 0.003395 0.002245
Methods Wavelet-cl GA-SVR
Original Wavelet Clustering testing 0.033683 0.011631 0.031525 0.011596
Models ustering training 0.006166 0.006242 0.003365 0.002229
training 0.002107 0.001964 0.001288 0.000862 PSO-SVR
GA-SVR testing 0.030412 0.011521 0.028697 0.010904
testing 0.001913 0.001751 0.001456 0.000809 training 0.006075 0.005978 0.003271 0.002240
training 0.002124 0.001966 0.001247 0.000840 FA-SVR
PSO-SVR testing 0.030150 0.011575 0.028536 0.011457
testing 0.001947 0.001820 0.001345 0.000710 training 0.006074 0.005914 0.003284 0.002299
training 0.002057 0.001886 0.001275 0.000849 DE-SVR
FA-SVR testing 0.030172 0.011820 0.028550 0.011605
testing 0.001869 0.001686 0.001159 0.000720 training 0.008049 0.009717 0.004165 0.002953
training 0.002115 0.001950 0.001471 0.001123 ARIMA
DE-SVR testing 0.037869 0.015136 0.034170 0.012203
testing 0.001924 0.001746 0.001240 0.000768 training 0.009866 0.007785 0.004145 0.003046
training 0.002547 0.002089 0.002287 0.001794 BPN
ARIMA testing 0.038259 0.012917 0.033026 0.015436
testing 0.002023 0.002380 0.001682 0.001397 training 0.010067 0.008369 0.004858 0.003582
training 0.002495 0.002981 0.002257 0.002088 Regression
BPN testing 0.038403 0.012217 0.035361 0.015971
testing 0.003315 0.002572 0.002628 0.001948 Dark gray color means that it is the best method for training data.
training 0.004311 0.002233 0.002920 0.002456 Light gray color means that it is the best method for testing data.
Regression
testing 0.004123 0.002029 0.002559 0.001891
Dark gray color means that it is the best method for training data.
Light gray color means that it is the best method for testing data.
Table 5
Average MSE result for North American dataset.
Table 3
Average MSE result for B3 dataset. Methods
Wavelet-cl
Original Wavelet Clustering
Methods Wavelet-cl ustering
Original Wavelet Clustering Models
Models ustering
training 0.006319 0.006690 0.004601 0.002198
training 0.002290 0.002282 0.000360 0.000150 GA-SVR
GA-SVR testing 0.042741 0.013147 0.025964 0.01295
testing 0.001674 0.001584 0.001240 0.000388
training 0.006522 0.006860 0.004777 0.002162
training 0.002267 0.002233 0.000326 0.000139 PSO-SVR
PSO-SVR testing 0.042350 0.012923 0.025221 0.012554
testing 0.001680 0.001609 0.001136 0.000382
training 0.006124 0.006715 0.004458 0.002158
training 0.002162 0.002091 0.000352 0.000120 FA-SVR
FA-SVR testing 0.042226 0.012992 0.025052 0.01254
testing 0.001642 0.001555 0.001118 0.000355
training 0.002362 0.002073 0.000417 0.000143 training 0.006123 0.006610 0.004479 0.002538
DE-SVR DE-SVR
testing 0.001696 0.001539 0.001132 0.000338 testing 0.042619 0.012921 0.025063 0.01258

training 0.002659 0.002485 0.000487 0.000356 training 0.009868 0.011407 0.005815 0.003616
ARIMA ARIMA
testing 0.001944 0.001748 0.001441 0.000435 testing 0.044257 0.015092 0.027363 0.017894

training 0.002540 0.002493 0.001699 0.000458 training 0.008582 0.007524 0.004779 0.003305
BPN BPN
testing 0.002727 0.002552 0.002185 0.000510 testing 0.043799 0.013303 0.026387 0.015306

training 0.002418 0.002327 0.001805 0.000480 training 0.010589 0.012513 0.006434 0.003851
Regression Regression
testing 0.002445 0.001761 0.002573 0.000563 testing 0.045611 0.016093 0.028969 0.018115
Dark gray color means that it is the best method for training data. Dark gray color means that it is the best method for training data.
Light gray color means that it is the best method for testing data. Light gray color means that it is the best method for testing data.
158 R.J. Kuo, P.S. Li / Computers & Industrial Engineering 99 (2016) 153161

 
cr 2ij 1 Computation for Time Series Prediction Minisite website. The first
xnew
i xold
i be x old
j  x old
i a rand  ; 19
2 one is the exchange rates between US Dollars and UK Pounds (B1),
which is a time series dataset; it shows the mean value of the US
Step 7: Stop criteria
Dollar per British Pound rate, taken each month from January
1981 till July 2005. The second one is the exchange rates between
Check if the specific number of iteration is satisfied to stop; US Dollars and Euros (B2); it shows the mean value of the US Dol-
otherwise, go back to Step 3. lars per Euro rate, taken each month from January 1979 till July
2005. The third one is the Consumer Price Index of Spain (B3); it
Step 8: Testing phase shows the Consumer Price Index (CPI) per month in Spain from
January 1960 till June 2005. The data are managed to use the
The best-performing parameter set is applied to testing model. records of the first five months to predict the value of the sixth
Compute the predicted value for each testing sample according to month.
the trained metaheuristics-based SVR with the best parameter set Three datasets are based on the export trade from Taiwan to
obtained from the training phase in the same cluster. North America (US and Canada), Europe (France, German, Italy
and the UK) and China. Each dataset has monthly records from
4. Experimental results January 2000 to November 2014. Three export trade datasets are
from the Taiwan National Statistics website.
4.1. Data collection
4.2. Computational results
There are three time series benchmark datasets provided by
the Spanish Central Bank from the Evolutionary and Neural K-fold cross-validation (K = 5) is applied to confirm the statisti-
cal independent random process, and every experiment is imple-
Table 6 mented for 30 runs. Tables 16 show the average MSE values of
Average MSE result for Chinese dataset.
training and testing for methods including: original, wavelet, clus-
Methods Wavelet-cl tering and wavelet-clustering process by using FA-SVR. Besides,
Original Wavelet Clustering
Models ustering GA-SVR, PSO-SVR, DE-SVR, ARIMA, BPN and Regression are also
training 0.005013 0.005041 0.001328 0.001166 conducted for comparison purpose. Figs. 38 are the MSE values
GA-SVR of B1, B2, B3, European dataset, North American dataset and Chi-
testing 0.008086 0.006391 0.005277 0.003416
nese dataset, respectively. Besides, the MSE values are lower after
training 0.004973 0.005331 0.001358 0.001003
PSO-SVR the process of clustering and wavelet transform. Furthermore, for
testing 0.008050 0.006372 0.005210 0.003394 B1 dataset, FA-SVR has better results than the other models under
training 0.004697 0.004956 0.001283 0.001096 original, wavelet transform and the clustering processes. However,
FA-SVR
testing 0.007529 0.006363 0.005121 0.003299 ARIMA has a better result than DE-SVR under the original process.
training 0.004690 0.004975 0.001284 0.001283 DE-SVR performs better under wavelet transform with clustering.
DE-SVR For B2 dataset, FA-SVR performs better under the original, wavelet
testing 0.007622 0.006395 0.005149 0.003320
transform and clustering processes. PSO-SVR performs better
training 0.008606 0.007695 0.001849 0.001497 under wavelet transform with clustering. For the B3 dataset, FA-
ARIMA
testing 0.009980 0.008565 0.005880 0.003557 SVR has better results under the original and clustering processes.
training 0.009188 0.006778 0.002049 0.001555 PSO-SVR has better results under wavelet transform, and wavelet
BPN transform with clustering. For the European dataset, FA-SVR has
testing 0.010579 0.008260 0.006774 0.00368
better results under the original and clustering processes. PSO-
training 0.009761 0.007846 0.001868 0.001758
Regression SVR has better results under wavelet transform, and wavelet trans-
testing 0.011166 0.008945 0.005248 0.003781 form with clustering processes. For the American dataset, FA-SVR
Dark gray color means that it is the best method for training data. has better results under the original, clustering and wavelet with
Light gray color means that it is the best method for testing data. clustering processes. DE-SVR has better results under wavelet

Fig. 3. MSE value of each model for B1 dataset.

R.J. Kuo, P.S. Li / Computers & Industrial Engineering 99 (2016) 153161 159

Fig. 4. MSE value of each model for B2 dataset.

Fig. 5. MSE value of each model for B3 dataset.

Fig. 6. MSE value of each model for European dataset.

160 R.J. Kuo, P.S. Li / Computers & Industrial Engineering 99 (2016) 153161

Fig. 7. MSE value of each model for North American dataset.

Fig. 8. MSE value of each model for Chinese dataset.

transform. For the Chinese dataset, FA-SVR has better results under
the original, wavelet transform, clustering, and wavelet transform
with clustering processes.
In fact, the MSE values are very similar among the four meta-
heuristics algorithms. In general, the FA-SVR model has the lowest
MSE value compared to the other six models, and Regression has
the highest MSE value compared to the other six models: B1, B2,
B3, European, Northern American and Chinese datasets.
5. Conclusions
This study proposes an integrated model of cluster analysis and
metaheuristics- based SVR with wavelet transform to predict the
values of Taiwanese export trade. In order to enhance the accuracy
of forecasting capability, we not only utilize metaheuristics algo-
rithms to find the optimal solutions for the SVR model, but also
apply wavelet transform and clustering analysis for data prepro-
cessing. Wavelet transform can help to decrease the impact of
noise and outliers, and clustering analysis can cluster the data with
similar features; therefore, they can improve the performance. In
this study, four metaheuristics are applied (GA, PSO, FA and DE)
to optimize the SVRs parameters. The computational results indi-
cate that FA usually has better performance than the other algo-
rithms because its expansive searching technique helps to find
the optimal parameters. According to the experimental results,
FA-SVR also outperforms ARIMA, BPN and Regression. However,
some directions are required: simple K-means algorithm is com-
bined with metaheuristics algorithms to cluster data in this study.
In future research, it may be feasible to utilize the time series clus-
tering method for time series dataset. The number of input nodes
influences both the autoregressive structures of time series and
the forecasting performance. Therefore, a technique to systemi-
cally determine the number of input nodes in SVR will be a poten-
tial direction for future development. In this study, only time series
data are used in the forecasting models. In fact, export trade fore-
casting is a complex topic to analyze. There are many influential
effects, such as the price index of exports, exchange rate and some
other financial indices. Furthermore, obtaining more precise
results of forecasting by applying multivariate statistical analysis,
feature selection and measuring the variable weight to forecasting
models is promising.
References
Cai, L., Qu, S., & Li, X. (2013). Self-adapt evolution SVR in a traffic flow forecasting. In
2013 IEEE international conference on signal processing, communication and
computing (ICSPCC) (pp. 15).
 R.J. Kuo, P.S. Li / Computers & Industrial Engineering 99 (2016) 153161
Cai, Z.-J., Sheng, L., & Xiao-bin, Z. (2009). Tourism demand forecasting by support
vector regression and genetic algorithm. In 2nd IEEE international conference on
computer science and information technology (pp. 144146).
Castro-Neto, M., Jeong, Y.-S., Jeong, M.-K., & Han, L. D. (2009). Online-SVR for short-
term traffic flow prediction under typical and atypical traffic conditions. Expert
Systems with Applications, 36(3), 61646173. Part 2.
Chang, P.-C., Wu, J.-L., & Lin, J.-J. (2016). A Takagi-Sugeno fuzzy model combined
with a support vector regression for stock trading forecasting. Applied Soft
Computing, 38, 831842.
Che, J., & Wang, J. (2010). Short-term electricity prices forecasting based on support
vector regression and Auto-regressive integrated moving average modeling.
Energy Conversion and Management, 51(10), 19111917.
Chen, K.-Y. (2011). Combining linear and nonlinear model in forecasting tourism
demand. Expert Systems with Applications, 38(8), 1036810376.
Chen, P.-Y., & Liu, L. (2013). Study on coal logistics demand forecast based on PSO-
SVR. In 2013 10th international conference on service systems and service
management (pp. 130133).
Chen, K.-Y., & Wang, C.-H. (2007). Support vector regression with genetic
algorithms in forecasting tourism demand. Tourism Management, 28(1),
215226.
Cherkassky, V., & Ma, Y. (2004). Practical selection of SVM parameters and noise
estimation for SVM regression. Neural Networks, 17(Mo. 1), 113126.
Ding, Y., Song, X., & Zen, Y. (2008). Forecasting financial condition of Chinese listed
companies based on support vector machine. Expert Systems with Applications,
34(4), 30813089.
Fang, S. F., Wang, M. P., Qi, W. H., & Zheng, F. (2008). Hybrid genetic algorithms and
support vector regression in forecasting atmospheric corrosion of metallic
materials. Computational Materials Science, 44(2), 647655.
Guanghui, W. (2012). Demand forecasting of supply chain based on support vector
regression method. Procedia Engineering, 29, 280284.
Guo, X., Sun, L., Li, G., & Wang, S. (2008). A hybrid wavelet analysis and support
vector machines in forecasting development of manufacturing. Expert Systems
with Applications, 35(12), 415422.
Hoang, N.-D., & Pham, A.-D. (2016). Hybrid artificial intelligence approach based on
metaheuristic and machine learning for slope stability assessment: A
multinational data analysis. Expert Systems with Applications, 46, 6068.
Hong, W.-C., Dong, Y., Chen, L.-Y., & Wei, S.-Y. (2011). SVR with hybrid chaotic
genetic algorithms for tourism demand forecasting. Applied Soft Computing, 11
(2), 18811890.
Hong, W.-C., Dong, Y., Zheng, F., & Lai, C.-Y. (2011). Forecasting urban traffic flow by
SVR with continuous ACO. Applied Mathematical Modelling, 35(3), 12821291.
Hu, Z., Bao, Y., Chiong, R., & Xiong, T. (2015). Mid-term interval load forecasting
using multi-output support vector regression with a memetic algorithm for
feature selection. Energy, 84, 419431.
Huang, W., Nakamori, Y., & Wang, S.-Y. (2005). Forecasting stock market movement
direction with support vector machine. Computers & Operations Research, 32
(10), 25132522.
Jiansheng, W., & Enhong, C. (2010). A novel hybrid particle swarm optimization for
feature selection and kernel optimization in support vector regression. In 2010
International conference on computational intelligence and security (CIS)
(pp. 189194).
Ju, Y.-F., & Wu, S.-W. (2010). Village electrical load prediction by genetic algorithm
and SVR. In 2010 3rd IEEE international conference on computer science and
information technology (ICCSIT) (pp. 278281).
Kao, L.-J., Chiu, C.-C., Lu, C.-J., & Chang, C.-H. (2013). A hybrid approach by
integrating wavelet-based feature extraction with MARS and SVR for stock
index forecasting. Decision Support Systems, 54(3), 12281244.
Kavousi-Fard, A., Samet, H., & Marzbani, F. (2014). A new hybrid modified firefly
algorithm and support vector regression model for accurate short term load
forecasting. Expert Systems with Applications, 41(13), 60476056.
Kazem, A., Sharifi, E., Hussain, F. K., Saberi, M., & Hussain, O. K. (2013). Support
vector regression with chaos-based firefly algorithm for stock market price
forecasting. Applied Soft Computing, 13(2), 947958.
Kim, K.-J. (2003). Financial time series forecasting using support vector machines.
Neurocomputing, 55(12), 307319.
Li, J., & Cai, Z. (2008). A novel automatic parameters optimization approach based
on differential evolution for support vector regression. In L. Kang, Z. Cai, X. Yan,
& Y. Liu (Eds.). Advances in computation and intelligence (Vol. 5370,
pp. 510519). Berlin, Heidelberg: Springer.
Li, M.-W., Hong, W.-C., & Kang, H.-G. (2013). Urban traffic flow forecasting using
GaussSVR with cat mapping, cloud model and PSO hybrid algorithm.
Neurocomputing, 99, 230240.
Lu, C.-J. (2014). Sales forecasting of computer products based on variable selection
scheme and support vector regression. Neurocomputing, 128, 491499.
Lu, C.-J., & Wang, Y.-W. (2010). Combining independent component analysis and
growing hierarchical self-organizing maps with support vector regression in
product demand forecasting. International Journal of Production Economics, 128
(2), 603613.
Mojumder, J. C., Ong, H. C., Chong, W. T., Shamshirband, S., & Abdullah Al, M. (2016).
Application of support vector machine for prediction of electrical and thermal
performance in PV/T system. Energy and Buildings, 111, 267277.
161
Nalley, D., Adamowski, J., & Khalil, B. (2012). Using discrete wavelet transforms to
analyze trends in streamflow and precipitation in Quebec and Ontario (1954
2008). Journal of Hydrology, 475, 204228.
Olatomiwa, L., Mekhilef, S., Shamshirband, S., Mohammadi, K., Petkovic, D., &
Sudheer, C. (2015). A support vector machinefirefly algorithm-based model for
global solar radiation prediction. Solar Energy, 115, 632644.
Pai, P.-F., & Hong, W.-C. (2005). Forecasting regional electricity load based on
recurrent support vector machines with genetic algorithms. Electric Power
Systems Research, 74(3), 417425.
Pai, P.-F., Hung, K.-C., & Lin, K.-P. (2014). Tourism demand forecasting using novel
hybrid system. Expert Systems with Applications, 41(8), 36913702.
Pan, H., Cheng, G., & Ding, J. (2013). Drilling cost prediction based on self-adaptive
differential evolution and support vector regression. In H. Yin, K. Tang, Y. Gao, F.
Klawonn, M. Lee, & T. Weise, et al. (Eds.). Intelligent data engineering and
automated learning IDEAL (Vol. 8206, pp. 6775). Berlin, Heidelberg: Springer.
Partal, T., & Kk, M. (2006). Long-term trend analysis using discrete wavelet
components of annual precipitations measurements in Marmara region
(Turkey). Physics and Chemistry of the Earth, Parts A/B/C, 31(18), 11891200.
Qi, X., & Yan, X. (2009). Forecast of the total volume of import-export trade based on
grey modelling optimized by genetic algorithm. In Third international
symposium on intelligent information technology application (pp. 545547).
Senthilnath, J., Omkar, S. N., & Mani, V. (2011). Clustering using firefly algorithm:
Performance study. Swarm and Evolutionary Computation, 1(3), 164171.
Shahrabi, J., Hadavandi, E., & Asadi, S. (2013). Developing a hybrid intelligent model
for forecasting problems: Case study of tourism demand time series.
Knowledge-Based Systems, 43, 112122.
Siamak, S. G., Bahrami Jovein, H., & Ramezanianpour, A. A. (2012). Hybrid support
vector regression Particle swarm optimization for prediction of compressive
strength and RCPT of concretes containing metakaolin. Construction and Building
Materials, 34, 321329.
Soltani, M., Moghaddam, T. B., Karim, M. R., Shamshirband, S., & Sudheer, C. (2015).
Stiffness performance of polyethylene terephthalate modified asphalt mixtures
estimation using support vector machine-firefly algorithm. Measurement, 63,
232239.
Tan, C., & Pedersen, C. N. S. (2009). Financial time series forecasting using improved
wavelet neural network : . Aarhus Universitet, Datalogisk Institut.
Tay, F. E. H., & Cao, L. (2001). Application of support vector machines in financial
time series forecasting. Omega, 29(4), 309317.
Tay, F. E. H., & Cao, L. J. (2002). Modified support vector machines in financial time
series forecasting. Neurocomputing, 48(14), 847861.
Thissen, U., van Brakel, R., de Weijer, A. P., Melssen, W. J., & Buydens, L. M. C. (2003).
Using support vector machines for time series prediction. Chemometrics and
Intelligent Laboratory Systems, 69(12), 3549.
Tiwari, M. K., & Chatterjee, C. (2010). Development of an accurate and reliable
hourly flood forecasting model using waveletbootstrapANN (WBANN) hybrid
approach. Journal of Hydrology, 394(34), 458470.
Tu, Y.-H., Hsien-Lun, W., & Chi-Chen, W. (2009). An evaluation of comparison
between multivariate fuzzy time series with traditional time series model for
forecasting taiwan export. In 2009 WRI world congress on computer science and
information engineering (pp. 462467).
Vapnik, V. (1995). The nature of statistical learning theory. New York: Springer-
Verlag.
Vapnik, V. N. (1999). An overview of statistical learning theory. IEEE transactions on
neural networks, 10(5), 988999.
Wang, C.-C. (2011). A comparison study between fuzzy time series model and
ARIMA model for forecasting Taiwan export. Expert Systems with Applications, 38
(8), 92969304.
Wang, J., Li, L., Niu, D., & Tan, Z. (2012). An annual load forecasting model based on
support vector regression with differential evolution algorithm. Applied Energy,
94, 6570.
Wong, H.-L., Tu, Y.-H., & Wang, C.-C. (2010). Application of fuzzy time series models
for forecasting the amount of Taiwan export. Expert Systems with Applications,
37(2), 14651470.
Wu, Q. (2009). The forecasting model based on wavelet m-support vector machine.
Expert Systems with Applications, 36(4), 76047610.
Wu, Q. (2010). Product demand forecasts using wavelet kernel support vector
machine and particle swarm optimization in manufacture system. Journal of
Computational and Applied Mathematics, 233(10), 24812491.
Xiong, T., Bao, Y., & Hu, Z. (2014). Multiple-output support vector regression with a
firefly algorithm for interval-valued stock price index forecasting. Knowledge-
Based Systems, 55, 87100.
Yan, X., & Chowdhury, N. A. (2014). Mid-term electricity market clearing price
forecasting utilizing hybrid support vector machine and auto-regressive
moving average with external input. International Journal of Electrical Power &
Energy Systems, 63, 6470.
Yeh, C.-Y., Huang, C.-W., & Lee, S.-J. (2011). A multiple-kernel support vector
regression approach for stock market price forecasting. Expert Systems with
Applications, 38(3), 21772186.