Received: 18 July 2017 Revised: 30 October 2017 Accepted: 12 January 2018

DOI: 10.1111/exsy.12271

ORIGINAL ARTICLE

Fuzzy time series model based on weighted association rule for

financial market forecasting
Ching‐Hsue Cheng | Chung‐Hsi Chen

Department of Information Management,

National Yunlin University of Science and Abstract
Technology, Douliou, Yunlin, Taiwan Fuzzy time series have been used to forecast future problems based on historical data. However,
Correspondence previously fuzzy time series methods have some problems: (a) subjectively determined the length
Ching‐Hsue, Cheng, Department of of intervals, (b) only discusses the relations for previous period and next period in forecasting
Information Management, National Yunlin
University of Science and Technology, 123
stage, and (c) no appropriately consider the weights of fuzzy relations. Based on these reasons,
University Road, Section 3, Douliou, Yunlin this study proposes a novel fuzzy time series model with 3 focuses: (a) proposing a spread parti-
64002, Taiwan. tion algorithm to calculate the length of intervals under the given number of linguistic value and
Email: chcheng@yuntech.edu.tw
automatically determine the lower bound and upper bound of universe, (b) considering multi‐
order fuzzy relation, and (c) building weighted fuzzy association rules based on Apriori's associa-
tion rules and compute the cardinality of each fuzzy relation. That is, this study proposed a novel
method based on Apriori's large itemsets to build the weighted fuzzy time series model for solving
forecast problems. To verify the proposed model, the gold price data sets and exchange rates are
employed as experimental data sets. This study compared the proposed model with other
methods for forecasting accuracy, and the comparison results shown that the proposed method
is better than other methods in root mean square error and mean absolute percentage error
criteria.

KEY W ORDS

financial market forecasting, fuzzy time series, multi‐order fuzzy rule, weighted association rules

1 | I N T RO D U CT I O N

Forecasting activities in our daily life are frequent and play an important role. Forecasting tools can help people to make decisions, such as the
weather prediction, the financial prediction, the stock prediction, and the enrolments of a school. The traditional forecasting methods can predict
the sequential problem but cannot deal with the problem, which relates linguistic values of historical data. Therefore, based on Zadeh's fuzzy sets
theory (Zadeh, 1965), Song and Chissom (1993a, 1993b, 1994) proposed fuzzy time series forecast model to deal with classical time series prob-
lems. They used two kinds of fuzzy time series forecast models, which are time‐invariant fuzzy time series model and time‐variant fuzzy time series
model, to forecast the enrolments of students of University of Alabama (Song & Chissom, 1993b, 1994). Next, Chen (1996) proposed a new fuzzy
time series to forecast the students' enrolments of Alabama University.
In financial predictions, predicting the gold price and exchange rates are very important investments. Hence, investment fund managers and
financial analysts predicted price by professional knowledge and investment analysis methods, such as technical analysis, fundamental analysis,
and time series models (Su & Cheng, 2016). In recent years, this type of forecasting problem in statistics, econometrics, and mathematical finance
has been studied extensively. There are many existing methods to help the user for predicting the price and the future index, those methods include
the work of Faff, Brooks, and Kee (2002), Huarng and Yu (2005), Shin and Sohn (2004), Wang (2002), and Yu (2005). Fuzzy time series model has
been used in several forecasting problems, such as forecasting university enrolment (Chen, 1996, 2002; Chen & Chung, 2006; Chen & Hsu, 2004),
temperature forecasting (Chen & Hwang, 2000), and stock index forecasting (Huarng & Yu, 2005; Cheng, Chang, & Yeh, 2006; Huarng, 2001a,
2001b; Huarng & Yu, 2003, 2004, 2006; Yolcu & Lam, 2017). Therefore, this study proposed an approach based on Apriori association rules for
financial market forecasting.

This study is based on conference paper (Liu, et al. 2011) to extend the proposed methods for solving forecast problems.

Expert Systems. 2018;e12271. wileyonlinelibrary.com/journal/exsy Copyright © 2018 John Wiley & Sons, Ltd 1 of 15
https://doi.org/10.1111/exsy.12271
2 of 15 CHENG AND CHEN

The main problem of past fuzzy time series models is subjectively to determine the length of intervals. Therefore, how to come up with a
model to confirm the appropriate interval length is necessary. In recent years, many fuzzy time series methods have been presented to solve
forecasting problems (Chen, 2002; Cheng, Chen, Teoh, & Chiang, 2008; Cheng, Cheng, & Wang, 2008; Chen & Chung, 2006; Chen & Hsu,
2004; Chen & Hwang, 2000; Huarng, 2001a, 2001b; Huarng & Yu, 2005; Huarng & Yu, 2003, 2004, 2006; Hwang, Chen, & Lee, 1998;
Lee, Wang, & Chen, 2007, 2008). Many of them had better forecasting performance than the traditional method, and those studies obtained
an appropriate interval length of linguistic value to improve the performance of forecasting. However, no literature reasonably reveal how to
determine the appropriate lower bound and upper bound of universe. If the upper bound of testing data is greater than the last language or
the lower bound of testing, data is less than the first language, the data will cause error for data fuzzification. Therefore, it is important to build
a more reasonable lower bound and upper bound of universe.
In weighted problem, Chen proposed the fuzzy time series method of fuzzy relations with equal weights (Chen, 1996). In fact, each fuzzy rela-
tion has different effects on prediction in the real world. The fuzzy relationship occurs more frequently, the degree of impact will be higher. Then,
Yu (2005) proposed the weighted fuzzy time series models, Yu's weighted models considered the recent fuzzy logistic rules, which are more impor-
tant than the past rules. Many previous studies have proposed different fuzzy time series models to deal with uncertain and vague data. A draw-
back of these models is that they do not appropriately consider the weights of fuzzy relations. Therefore, in order to solve the problem of the
weight of fuzzy relations, it is necessary to assign the appropriate weight.
Based on previous reasons mentioned, this study proposed a fuzzy time series model based on weighted Apriori's association rules, which has
three focuses:

1. Utilize objective method to determine automatically the length of interval and proposed spread partition algorithm to decide the linguistic
intervals.

2. Propose the weighted fuzzy association rules.

3. Build the forecasting model for one‐order fuzzy rules and multi‐order fuzzy rules.

The remaining content of this paper is organized as follows. Section 2 reviews the related literatures. In Section 3, the proposed model is intro-
duced by using the proposed algorithm for one‐order and multi‐order weighted fuzzy rules. Section 4 evaluates the performance of proposed
model and compares with the other models. In Section 5, the profits of data sets are shown. Section 6 is the conclusions and future work.

2 | R E LA T E D WO R K

In this section, the related literature included fuzzy sets theory, fuzzy time series, and association rule are briefly introduced.

2.1 | Fuzzy sets theory

Fuzzy sets theory was introduced by Zadeh (1965), who considered many problems with uncertainty and imprecision in the real world. It is difficult
to use the classical logic to solve this type of problem. Therefore, he proposed the fuzzy sets theory to deal with the vague and imprecise data
(Zadeh, 1965, 1975a, 1975b, 1976, 1988). Nowadays, fuzzy theory has been used in different areas, such as expert systems, databases, and infor-
mation retrieval systems.
Zadeh (1975a, 1975b) pointed out that a language variable is “a variable whose values are words or sentences in natural or artificial language.”
For instance, linguistic variable is described by linguistic terms in nature, such as “good,” “medium,” and “bad.” Each linguistic variable could be
assigned one or more linguistic values, which are connected to a numeric value through the membership function. For example, the linguistic var-
iable S includes five fuzzy linguistic terms by triangular fuzzy numbers to express unquantified matters.
Defuzzification is the convention of transforming the fuzzy relations into crisp quantity. The defuzzification methods have been proposed and
popular defuzzified fuzzy output functions. There are four kinds of defuzzification methods (Ross, 2000):

1. Max‐membership principle
This scheme is limited to peak output function, the method is given by the algebraic expression


μc z* ≥μc ðzÞ for all z∈Z: (1)

2. Centroid method
This method is the most popular defuzzification method, it also called centre of area or centre of gravity. It is given by the algebraic expression

∫μC ðzÞ⋅zdz
z* ¼ : (2)
∫μC ðzÞ dz
CHENG AND CHEN 3 of 15

3. Weighted average method

This method is only applicable to symmetrical output membership functions, it is given by the algebraic expression

∑μC ðzÞ⋅z
z* ¼ : (3)
∑μC ðzÞ

4. Mean‐max membership
This method also called middle‐of maxima, it is closely related to max‐membership principle, except that the locations of the maximum mem-
bership can be non‐unique. It is given by the algebraic expression

aþb
z* ¼ : (4)
2

2.2 | Fuzzy time series

Song and Chissom (1993b) proposed fuzzy time series model to deal with forecasting problems involving historical data with linguistic
expression (Zadeh, 1975a, 1975b, 1976), many fuzzy time‐series models have been proposed based on Song and Chissom's definitions.
Chen's model (1996) is very common, because of easy computation and good forecasting performance. Singh and Borah (2013) proposed
an efficient time series forecasting model based on fuzzy time series. Next, some definitions of fuzzy time series are introduced in the
following.
In general, a fuzzy set is a category with fuzzy boundaries. Let U be the universe of discourse, U = {u1, u2, …, un}. A fuzzy Set A on U is defined by

A ¼ f A ðu1 Þ=u1 þ f A ðu2 Þ=u2 þ … þ f A ðun Þ=un ;

where fA is the membership function of fuzzy Set A, fA : U → [0, 1], ui is an element of fuzzy Set A, fA(ui) indicates the degree of membership of ui in
A, fA(ui) ∈ [0, 1] and 1 ≤ i ≤ n.

Definition 1. Fuzzy time series

Let Y(t) (t = …, 0, 1, 2, …) be a subset of real numbers. Y(t) be the universe of discourse defined by fuzzy set fi(t). Let F(t) be a collection of
fi(t) (i = 1, 2, …), and defined as a fuzzy time‐series on Y(t) (t = …, 0, 1, 2, …).

Definition 2. Fuzzy time series relationship

If there exists a fuzzy relationship R(t − 1, t), the relationship can be expressed as F(t) = F(t − 1) × R(t − 1, t), where R(t − 1, t) rep-
resents a fuzzy relationship between F(t) and F(t−1), × is an operator, then F(t) is said to be caused by F(t − 1). When F(t − 1) = Ai and
F(t) = Aj, the FLR between F(t) and F(t − 1) can be denoted as Ai → Aj, where Ai and Aj are the left‐hand side and right‐hand side of
the fuzzy logical relationship (FLR).
Fuzzy time series model contains establishing fuzzy relationships, forecasting, and defuzzifying, these computation steps could briefly use
Chen's model (Chen, 1996) to introduce as follows:

Step 1. Define the universe of discourse and intervals.

Based on the domain, the universe of discourse can be defined as U = [Dlow, Dup], where Dlow and Dup mean the lower value and the upper value
of U. As the length of interval is determined, U can be partitioned into several equal length intervals.
Step 2. Define the fuzzy sets.
Define the fuzzy sets based on the universe of discourse and fuzzify the historical data.

Step 3. Fuzzify the rules.

For instance, if the maximal degree of membership of that datum is in Aj, a datum is fuzzified to Aj.
Step 4. Establish FLRs.
Establish FLRs and group them based on the current states of the data of the FLRs. For example, A1 → A2, A1 → A1, A1 → A3, and they can be
grouped as A1 → A1, A2, A3.
Step 5. Forecast.
Let F(t − 1) = Ai. There are two cases as following.

Case 1. In the FLR sequence, there is only one FLR. If Ai → Aj, then the forecast value F(t) is equal to Aj.

Case 2. There has more than one FLR in the FLR sequence. If Ai → Ai, Aj,⋯, Ak, then F(t), forecast value, is equal to Ai, Aj, …, Ak.
Step 6. Defuzzify.
Apply the centre‐of‐gravity method to get the forecasting results. This method is the most often adopted method of defuzzification.
4 of 15 CHENG AND CHEN

2.3 | Association rule

In the Information Technology industry, data mining is a rapidly growing field. Techniques of data mining include clustering, association rules,
decision trees, artificial neural networks, and genetic algorithms. Association rules in unsupervised learning system are the most common form
of local pattern discovery. Besides, Apriori algorithm (Agrawal & Srikant, 1994) is the most traditional and well known association rules meth-
odologies. The basic constructs of association rules are that a transaction consists of a set of items, and a database consists of a set of trans-
actions. An association rule can be denoted as X → Y, where X and Y are sets of items, also called an itemset. The ratio of the number of
transactions including itemset X and the total number of transactions in the database is called support of itemset X. The confidence of an asso-
ciation rule X → Y is a ratio of support of X∪Y to X. Furthermore, if support of an itemset is greater than minimum support, which is predefined
by user, is called large itemset.
Agrawal and Srikant (1994) proposed the algorithm of Apriori in 1994. The important concept of Apriori is that generates the candidate
itemsets from only the large itemsets found in the previous pass. The algorithm through several iterations computes the large (frequent) itemsets
in the database. Iteration i computes all large i‐itemsets (itemsets with i elements), and each iteration has two steps, which are candidate generation
and candidate counting and selection. In the first step, the created set of candidate itemsets contains all i‐itemsets, and the algorithm calculates
support of association rules searching again through the whole database. Then, based on knowledge about infrequent itemsets obtained from pre-
vious iterations, the Apriori algorithm reduces the set of candidate itemsets by pruning‐apriori‐those candidate itemsets that are infrequent in the
second phase (Kantardzic, 2002).

3 | PROPOSED MODEL

The traditional time‐series models (i.e., statistical methods) could solve the sequence data problem but could not handle the uncertain and vague
information. To improve the problems, Song and Chissom (1993b) proposed fuzzy time series model to deal with uncertain and vague problems,
and then many researchers have proposed different fuzzy time series models (Chang, Lee, Liao, & Cheng, 2007; Chen, 1996; Cheng et al., 2006;
Hwang et al., 1998; Lee et al., 2008; Yu, 2005). According to these researches, this paper summarized the influence factors of fuzzy time series
models as follows.

1. Length of interval: The length of interval must influence the forecasting accuracy, furthermore, previous fuzzy time series models subjectively
defined the linguistic intervals, which will result in forecasting shortcomings, such as some linguistic intervals do not include any data and some
intervals include many data (Cheng et al., 2006; Chen & Hwang, 2000; Huarng, 2001b; Yu, 2005).
2. Lag period: In the past researches, the forecasting of stock price is based on lag periods data, and the lag periods is an important factor.

3. Weight of fuzzy rule: Chen's model (Chen, 1996) uses the centre of gravity method to defuzzify with the equal weights regards each fuzzy
rules as the same priority; however, each fuzzy rule has the different influence on the forecasting result. In the past, fuzzy time series
model concerned about the FLRs and ignored the correlation between the recent period prices (Cheng, Chen, et al., 2008). The investment
decisions are usually related on the recent stock prices; consequently, the forecasting stock price is strongly associated to the pervious
stock price.

Based on mentioned above, this study is based on conference paper (Liu et al., 2011) to extend the proposed methods for solving the
problems:

1. Proposed a reasonable algorithm to determine the appropriate lower bound and upper bound of universe, then the data could be adequately
spread in the linguistic intervals;
2. Based on the work of Teoh, Chen, Cheng, and Chu (2009), who have revealed that the estimated time lags for the two empirical databases are
one‐lagged period, however, this study will consider multi‐order fuzzy relation; and
3. Build fuzzy rule based on Apriori‐AR and calculate the weights of fuzzy relations for forecasting financial market problems.

3.1 | Algorithm
To compute the proposed model easily, this section proposes an algorithm with seven steps. The detailed description of each step is introduced as
follows:

Step 1. Define the universe of discourse and determine the length of intervals automatically.
First, define the universe of discourse U. Let U = [Dlow, Dup], where Dlow and Dup denote the lower value and the upper value of U. Partition U
into n equal intervals and the length is l of each interval as
CHENG AND CHEN 5 of 15

2   3
Dmax −Dmin Dmax −Dmin
þ
6 n n−1 7
l¼6
4
7;
5 (5)
2

where Dmin and Dmax are the minimum value and the maximum value of observation. Then, set n as the number what intervals will be
partitioned, and the length of intervals l can be calculated by Equation 5. According to the Miller's concept of limitation of humans on the capac-
ity of processing information, that is, the magical number seven, plus or minus two, this step partitions the universe of discourse into seven
equal‐length intervals referring linguistic expression to represent the observation conditions (Miller, 1956). After determining the length of inter-
vals, then use the proposed “A” (automatically determining lower bound and upper bound) to calculate the lower value Dlow and the upper value
Dup of the universe of discourse U.

Step 2. Assign the linguistic value (fuzzify the data).

Using the length l to define the linguistic value L1, L2, …,Li, …, Ln. Then, according to each record in which interval and determined the
corresponding linguistic value. The membership value (grade) of the historical data is calculated by the membership function. The first L1
and last Ln membership functions are modified as Equations 6 and 8, and the middle Li membership functions can be defined as
Equation 7.

8
> 1; x<a
>
>
>
< 1; a≤x<b
μL1 ðxÞ ¼ c−x : (6)
>
> ; b≤x<c
>
> c−b
:
0; x>c

8
>
> 0;
>
> x−a x<a
>
>
< ; a≤x<b
μLi ðxÞ ¼ b−a : (7)
>
> c−x ; b≤x<c
>
>
>
> c−b
: x>c
0;

8
> 0; x<a
>
>
>
< x−a; a≤x<b
μLn ðxÞ ¼ b−a : (8)
>
>
> 1; b≤x<c
>
:
1; x>c

Algorithm A: Automatically determining lower bound and upper bound

Input: length l of intervals, minimum observation Dmin, maximum observation Dmax

Output: lower value Dlow of U, higher value Dup of U
Method:
   
Dmin Dmin −0:5l
1. if ¼ then
l l
  
Dmin
2. Dlow ¼ þ 0:5l
l
3. else
 
Dmin
4. Dlow ¼ l
l
5. if (Dlow = 0) then
Dmin
6. Dlow ¼
2
   
Dmax Dmax þ 0:5l
7. if ¼ then
l l
  
Dmax
8. Dup ¼ þ 0:5
l
6 of 15 CHENG AND CHEN

9. else
  
Dmax
10. Dup ¼ þl
l

where μLi ðxÞ denotes that the membership value of crisp data x belongs to fuzzy sets Li, and a, b, c denote the lower bound, midpoint, and the upper
bond of intervals of Li, respectively. If the data meet two membership functions, the highest membership value is chosen and linguistic value is
determined.

Step 3. Find the FLRs.

This study finds the FLRs according to the training data. FLRs include one‐order fuzzy relations and k orders fuzzy relations and are
described as:
Step 3.1. Find the one‐order fuzzy relations.
Find the one‐order fuzzy relations from the historical data. The one‐order fuzzy relations are defined as

 o1  o1
o1 þ1
o1 þ1 
Li ; μ Li → Lj ; μ Lj
 o2  o2
o2 þ1
o2 þ1 
Li ; μ Li → Lj ; μ Lj
⋮
 ov  ov
ov þ1
ov þ1  (9)
Li ; μ Li → Lj ; μ Lj
⋮
 on  on
on þ1
on þ1 
Lm ; μ Lm → Lj ; μ Lj

where 1 ≤ i ≤ m, 1 ≤ j ≤ m, m is the total number of the linguistic values, and 1 ≤ v ≤ n, n is the total occurrence times in the same relation. Ov denotes
v times for occurrence of the same relation. Loi 1 is the antecedent of the rule Li → Lj for the first time occurrence, and Ljo1 þ1 is the next linguistic value
of Loi 1 . In this study, Li and Lj, respectively, represent the antecedent and consequent linguistic values of the relation rules.

Step 3.2. Find the k‐orders fuzzy relations.

The k orders fuzzy relations (k = 2, 3, …) are established from (k‐1) orders fuzzy rules. The k‐orders fuzzy relations can be defined as









Loi 1 −kþ1 ; μ Lio1 −kþ1 ; …; Lio1 −1 ; μ Lio1 −1 ; Loi 1 ; μ L0i 1 → Ljo1 þ1 ; μ Ljo1 þ1




  


Loi 2 −kþ1 ; μ Lio2 −kþ1 ; …; Lio2 −1 ; μ Lio2 −1 ; Loi 2 ; μ Loi 2 → Loj 2 þ1 ; μ Ljo2 þ1








 (10)
Loi v −kþ1 ; μ Loi v −kþ1 ; …; Liov −1 ; μ Loi v −1 ; Loi v ; μ L0i v → Ljov þ1 ; μ Ljov þ1




  


Loi n −kþ1 ; μ Loi n −kþ1 ; …; Loi n −1 ; μ Loi n −1 ; Lomn ; μ Lomn → Loj n þ1 ; μ Loj n þ1 :

All the definitions of variables are the same with one‐order fuzzy relations.

Step 4. Calculate the weight by cardinality of fuzzy relations.

The following Equation 11 is derived to calculate the cardinality of k‐orders fuzzy relation. If k = 1 means the one‐order, k = 2 means two
orders, and so on. The cardinality of k‐orders fuzzy relation could be defined as

 n


 

o −kþ1 o −1 o þ1
W Loi s −kþ1 ; …; Lios þ1 ; Lti s ; Ljos þ1 ¼ ∑ min μ Li p
o
; …; μ Li p ; μ Li p ; μ Lj p (11)
p¼1

where n denotes the total number of the same fuzzy relations, k denotes k‐orders fuzzy relation, and s denotes s‐th relation. Loi s −kþ1 is the linguistic
value of | − k + 1|th antecedent part of s‐th relation. On the whole, the weight of fuzzy relation is obtained by computing the cardinality of k‐orders
fuzzy relation, and W does not limit to [0,1].
After computing the confidence of each fuzzy relation, select the one‐order fuzzy rules and k‐orders fuzzy rules. The confidence of fuzzy rela-
tion is calculated by Equation 12.

W Li →Lj
CLi →Lj ¼ ; (12)
nðLÞ
∑ W ðLi →Lk Þ
k¼1

where CLi →Lj denotes the confidence of fuzzy rule Li → Lj, and n(L) denotes the number of linguistic values. The threshold value for confidence
(denoted as α) can be automatically adapted by minimal root mean square error (RMSE), the selected fuzzy rules with weights are greater than
CHENG AND CHEN 7 of 15

α. In this study, use training data to build the fuzzified relations and named it as fuzzy relations. Applying the threshold value selects fuzzy relations,
all selected fuzzy relations are called fuzzy rules.

Step 5. Fit the fuzzy rules for linguistic value.

From Steps 3–4, the forecasting rules can be found by proposed B (Rule Finding Algorithm), and then make forecast. First, select the fitting
fuzzy rules to calculate the forecasted values. The one‐order forecasting results represent that forecasting only use one‐order fuzzy rules,
and the k‐order forecasting results mean use k‐orders fuzzy rules.
Step 5.1. Fit one‐order fuzzy rules for linguistic value.
For all one‐order fuzzy rules, if LðtÞ ¼ Lti s , the fuzzy rule which has the same antecedent with Lti s is defined as the forecasting rule, where L(t) is
denoted as the previous linguistic value of forecasting value.
Step 5.2. Fit k‐order fuzzy rules for linguistic value
If Lðt−k þ 1Þ ¼ Lits −kþ1 ; Lðt−k þ 2Þ ¼ Lits −kþ2 ; …; Lðt−1Þ ¼ Lti s , and LðtÞ ¼ Lti s , the fuzzy rule which has the same antecedent with
ts −kþ1 ts −kþ2
Li ; Li ; …; Lti s is defined as the forecasting rule for k‐order fitting fuzzy rules.

Algorithm B: Rule Finding Algorithm

 
Input: the dataset of linguistic value L, the antecedent of the rule Li → Lj for the v time occurrence and the membership Loi v ; μ Loi v , the next lin-



guistic value of Loi v and the membership Loj v þ1 ; μ Ljov þ1 , total occurrence times in the same relation n, fuzzy relations orders k, train number

trainnum, the confidence thresholdα

Output: the forecasting rule FRule of dataset
Method:
1 do
 


2 Loi v ; μ Loi v → Ljov þ1 ; μ Ljov þ1









3 Liov −kþ1 ; μ Liov −kþ1 ; …; Liov −1 ; μ Loi v −1 ; Loi v ; μ L0i v → Ljov þ1 ; μ Ljov þ1

4 for all rule in dataset Rule

5 t = cout++
6 until t = trainnum
7 do

 n


 

o −kþ1 o −1 o þ1
8 W Lios −kþ1 ; …; Lios þ1 ; Lti s ; Ljos þ1 ¼ ∑ min μ Li p
o
; …; μ Li p ; μ Li p ; μ Lj p
p¼1

9 until dataset Rule has no item

10 do

W Li →Lj
11 CLi →Lj ¼
nðLÞ
∑ W ðLi →Lk Þ
k¼1

12 if>α then
13 Li → Lj in dataset FRule
14 until dataset Rule has no item

Step 6. Forecast by using the established fuzzy rules

Two kinds of fuzzy rule sets has been obtained from Step 5, then calculate the forecasting value by the established fuzzy rules.
Step 6.1. Forecast by using one‐order fuzzy rules
The one‐order forecasting results only consider one‐order fitting fuzzy rules. When the antecedent of the forecasting data is corresponding
with the fuzzy rules, the one‐order forecasting value is calculated by Equation 13.

n1 
∑ W p *D Lj
p¼1 Loi s ;Loj s þ1
F ðt þ 1Þ ¼ n1 ; (13)
∑ W p o þ1

p¼1 Loi s ;Lj s

where F(t + 1) denotes forecasted value, n1 denotes the total number of one period fitting fuzzy rules, and Wis the weights of the fitting fuzzy rules.
D(Lj) is a defuzzified value of Lj and the value is calculated by Equation 14.

 aL þ bLj þ cLj
D Lj ¼ j ; (14)
3
8 of 15 CHENG AND CHEN

where aLj , bLj , and cLj are the lower bound, the midpoint, and the upper bound of the linguistic interval. If the antecedent of the forecasting data have
no fitting fuzzy rules, and the linguistic value is in the range of linguistic value for training rule sets, find the fuzzy rules which antecedent linguistic
value plus and minus one linguistic value, then use the fuzzy rules (Loi−1
v
,Loiþ1
v
) to forecast by Equation 13. If the antecedent of the forecasting data has
no fitting fuzzy rules, and the linguistic value (Loi v ) is out of all training rules, the one period forecasting value is gained by calculating the average of
the interval of linguistic value (Loi v ) by Equation 15.

ILui þ ILl i
Fðt þ 1Þ ¼ ; (15)
2

where ILui is upper value of the interval of linguistic value and ILl i is lower value of the interval of linguistic value.

Step 6.2. Forecast by using high‐order fuzzy rules

The high‐order forecasting value is similar to one‐order forecasting, the detailed computation was shown in Algorithm C (High‐Order Forecast-
ing Algorithm), when all orders fitting fuzzy rules, the k‐order forecasting is represented as:

n1 
∑ W r1 os *D Lj Þ
r1 ¼1 Li ;Ljos þ1
F ðt þ 1Þ ¼ n1 n2 nk
∑ W r1 os þ ∑ W r2 o −1 þ … þ ∑ W rk o −kþ1
r 1 ¼1 Li ;Loj s þ1 r 2 ¼1
s Li ;Loi s ;Ljos þ1 s
r k ¼1 Li ;Loi s −kþ2 ;…;Loi s ;Loj s þ1

n2 
∑ W r2 os−1 *D Lj
r 2 ¼1 Li ;Loi s ;Loj s þ1
þ n1 n2 nk þ …… (16)
∑ W r1 os þ ∑ W r2 o −1 þ … þ ∑ W rk o −kþ1
r 1 ¼1 Li ;Loj s þ1 r 2 ¼1
s Li ;Loi s ;Ljos þ1 s
r k ¼1 Li ;Loi s −kþ2 ;…;Loi s ;Loj s þ1

nk 
∑ W r *D Lj
r¼1 Loi s −nþ1 ;Lios −nþ2 ;…;Loi s ;Loj s þ1
þ n1 n2 nk
∑ W r1 os þ ∑ W r2 o −1 þ … þ ∑ W rk o −kþ1
r 1 ¼1 Li ;Loj s þ1 r 2 ¼1
s Li ;Loi s ;Ljos þ1 s
r k ¼1 Li ;Loi s −kþ2 ;…;Loi s ;Ljos þ1

where n1, n2, and nk mean the total number of one, two, and k orders fitting fuzzy rules. If there does not exist any fitted fuzzy rules, F(t + 1) is
calculated by the same method of one‐order fuzzy rules forecast.

Step 7. Forecast and evaluation

This paper employed mean absolute percentage error (MAPE) and RMSE as the performance criteria and compared with the other fuzzy time‐
series models. The MAPE and RMSE are defined as

n
F −A

i i

MAPE ¼ i¼1 i
(17)
n

and
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
un
u
u ∑ ðAi −F i Þ2
t
RMSE ¼ i¼1 ; (18)
n

where Fi is the forecasting value of year i, Ai is the actual value of year i, and n is the total number of years. Use these criteria to evaluate the fore-
casting results of the proposed method with the existing methods.

Algorithm C: High‐Order Forecasting Algorithm

Input: the dataset of the forecasting rule FRule, the weights of the fitting fuzzy rules W, D(Lj) defuzzified value of Lj, upper value of the interval of
linguistic value ILui , lower value of the interval of linguistic value ILl i
Output: forecasted value F(t + 1)
Method:
1. i←2
2. do
3. if ( i-1 = t )
CHENG AND CHEN 9 of 15

n1 
∑ W p *D Lj
p¼1 Loi s ;Loj s þ1
4 F ðt þ 1Þ ¼ n1
∑ W p
p¼1 Loi s ;Ljos þ1

5. else
ILui þ ILl i
6. F ðt þ 1Þ ¼
2
7. end if
8. if ( i-3 = t-2 and i-2 = t-1 and i-1 = t )
n1 
Fðt þ 1Þ ¼ ∑ W r1 os *D Lj Þ
r 1 ¼1 Li ;Ljos þ1
9. n1 n2 nk

∑ W
þ ∑ W
 þ…þ ∑ W
 þ⋯
r
1 2 r k r

r 1 ¼1 o o þ1 r 2 ¼1 o −1 o o þ1 rk ¼1 o −kþ1 os −kþ2 o o þ1
L s ;L s L s ;L s ;L s L s ;L ;…;L s ;L s
i j i i j i i i j

10. else
ILui þ ILl i
11. F ðt þ 1Þ ¼
2
12. end if
13. until i = 13

4 | EXPERIMENT AND COMPARISON

This section provides model verification, performance evaluation, and comparison for the proposed model. The daily gold price data set is collected
from the Central Trust of China, and the exchange rates (U.S. Dollar vs. Taiwan Dollar) are got from the Central Bank of the Republic of China
(Taiwan) as verification data sets. Moreover, MAPE and RMSE are used as evaluation criteria and compare the proposed model with other fuzzy
time‐series models that include ARIMA (Box & Jenkins, 1976) and GRNN (Specht, 1991) and those was proposed by Chen (1996) and Yu (2005).

4.1 | Forecasting for gold price

In this study, the daily gold price data set, from 2004/4/1 to 2005/09/29, is selected to test the proposed model. This data set includes 359
records, and partition it into two parts: training data set and testing data set. Training data set has 241 records and testing data set has 118 records.
Using the B partitioned, the universe of discourse, then, automatically determines the length of intervals to fuzzify the data and calculate the
membership value. Tables 1 and 2 respectively show the fuzzified results of training data set and testing data set for gold price. Next, the FLRs are
constructed based on the AR. The one‐order, two orders, and three orders fuzzy rules and the weights of rules are presented in Tables 3–5, respec-
tively. Then, employ the one‐order and high‐order fuzzy rules to obtain the fuzzy forecast and defuzzify the fuzzy forecast to get the forecasting
value. Table 6 shows the one‐order and multi‐order forecasting results. Finally, this paper uses Chen's model (1996), Yu's model (2005), Chang et al.'
s model (2007), ARIMA (Box & Jenkins, 1976), GRNN (Specht, 1991), and the proposed model to forecast testing data set for performance mea-
surement. The comparison results are shown in Table 7 for MAPE and RMSE criteria.
From Table 7, the MAPE of the proposed Models A and B are 0.48 and 0.39, respectively. The proposed models outperform Chen's model
(1996), Yu's model (2005), Chang et al.'s model (2007), ARIMA (Box & Jenkins, 1976), and GRNN (Specht, 1991). The RMSE of proposed model

TABLE 1 Fuzzified result for training data set

Date Actual value Linguistic value Membership value

2004/4/1 16,843 L8 0.038760

2004/4/4 16,735 L6 0.914729
2004/4/7 16,837 L8 0.364341
2004/4/8 16,791 L7 0.860465
2004/4/11 16,869 L9 0.627907
2004/4/12 16,874 L9 0.356589
2004/4/13 16,862 L9 0.992248
2004/4/14 16,879 L9 0.085271
2004/4/15 16,779 L7 0.488372
⋮ ⋮ ⋮ ⋮
2005/3/30 16,917 L10 0.023256
2005/3/31 16,937 L11 0.937984

Note: The vertical dots denote that there are many similar data values in this Table.
10 of 15 CHENG AND CHEN

TABLE 2 Fuzzified result of testing data set

Date Actual value Linguistic value Membership value

2005/4/1 16,843 L8 0.038760

2005/4/4 16,735 L6 0.914729
2005/4/7 16,837 L8 0.364341
2005/4/8 16,791 L7 0.860465
2005/4/11 16,869 L9 0.627907
2005/4/12 16,874 L9 0.356589
2005/4/13 16,862 L9 0.992248
2005/4/14 16,879 L9 0.085271
2005/4/15 16,779 L7 0.488372
⋮ ⋮ ⋮ ⋮
2005/9/28 19,192 L33 0.775194
2005/9/29 19,467 L34 0.906977

Note: The vertical dots denote that there are many similar data values in this Table.

TABLE 3 One‐order fuzzy rules with weight

One‐order fuzzy rules Weight

L1 → L1 0.465116
L1 → L2 0.356589
L2 → L1 0.480620
L2 → L2 3.007752
L3 → L2 0.232558
L3 → L4 0.015504
⋮ ⋮
L33 → L34 1.565891
L34 → L34 0.906977

Note: The vertical dots denote that there are many similar data values in this Table.

TABLE 4 Two orders fuzzy rules with weight

Two orders fuzzy rules Weight

L1, L1 → L4 0.015504
L1, L2 → L2 0.356589
L1, L4 → L3 0.015504
L2, L1 → L1 0.465116
L2, L1 → L2 0.015504
⋮ ⋮
L34, L33 → L33 0.519380
L34, L34 → L23 0.604651

Note: The vertical dots denote that there are many similar data values in this Table.

TABLE 5 Three orders fuzzy rules with weight

Three orders fuzzy rules Weight

L1, L1, L4 → L3 0.015504

L1, L2, L2 → L2 0.356589
L1, L4, L3 → L4 0.015504
L2, L1, L1 → L4 0.015504
L2, L1, L2 → L2 0.015504
L2, L2, L1 → L1 0.465116
⋮ ⋮
L33, L34, L34 → L23 0.604651
L34, L34, L23 → L24 0.604651

Note: The vertical dots denote that there are many similar data values in this Table.
CHENG AND CHEN 11 of 15

TABLE 6 Forecasting results

Date Actual value One‐order forecast High‐order forecast

2005/4/1 16,843 ─ ─
2005/4/4 16,735 16,747 16,750
2005/4/7 16,837 16,794 16,793
2005/4/8 16,791 16,808 16,835
2005/4/11 16,869 16,903 16,895
2005/4/12 16,874 16,903 16,890
2005/4/13 16,862 16,903 16,890
2005/4/14 16,879 16,903 16,890
2005/4/15 16,779 16,808 16,806
2005/4/18 16,759 16,747 16,739
⋮ ⋮ ⋮ ⋮
2005/9/28 19,192 19,368 19,366
2005/9/29 19,467 19,368 19,366

Note: The vertical dots denote that there are many similar data values in this Table, and the symbol “─” denote no available data.

TABLE 7 Comparison results for testing dataset of the gold price

Models MAPE (%) RMSE

Chen (1996) 1.03 218.88

Yu (2005) 1.01 229.19
Chang et al. (2007)—One period 0.91 207.71
Chang et al. (2007)—Multiperiods 0.90 205.92
ARIMA (2,1,1; Box & Jenkins, 1976) 0.91 150.42
GRNN(Specht, 1991) 0.72 452.81
Proposed A 0.48 108.61
Proposed B 0.39 88.74

Note. (1) “Proposed A” denotes proposed one‐order model; “Proposed B” is proposed multi‐order model. (2) The ARIMA (2,1,1) is optimized lag‐period by
Eviews lag tested. (3) The highlighted entries denote the best performance in eight models for each year dataset.

by one‐order and multi‐order are 108.61 and 88.74, respectively; from the results, the proposed model gets the smaller RMSE than other models.
Therefore, the proposed method has made a great improvement in forecasting the gold price. The performance of three orders is better than one
order, because the previous three orders gold price all influence the forecast price. It can be said the proposed three‐order method is suitable for
gold price.

4.2 | Forecasting for exchange rates (USD vs. TWD)

In this experiment, the exchange rates (USD vs. TWD) for 2004–2008 are used for forecasting. Each year of experimental data sets is split into two
subsets, a training data set and a testing data set, where the two‐thirds are used for training, and the one‐third is used for testing. The comparison
results with different models are listed in Table 8. From Table 8, the proposed model outperforms the other models which are Chen's model (1996),
Yu's model (2005), Chang et al.'s model (2007), ARIMA (Box & Jenkins, 1976), GRNN (Specht, 1991). Because the proposed Method A wins four in
five subdata for RMSE, and MAPE with three win in five sub‐data. In the proposed model, one‐order performance is better than three orders,
because of the previous one‐lag period affect the forecast price more than the previous three‐lag period.

5 | PROFIT EVALUATIONS

In order to demonstrate the profit of the proposed model, the profit computation method proposed by Cheng and Wei (2009) are employed. The
profit computation method is defined as following:

Rule 1. Sell Rule

jforecastðtÞ−actualðtÞj
if ≤α and forecast(t + 1) − actual(t) > 0,then sell.
actualðtÞ

Rule 2. Buy Rule

12 of 15 CHENG AND CHEN

TABLE 8 Comparison results for testing dataset of the exchange rates

2004 2005 2006 2007 2008
Models MAPE RMSE MAPE RMSE MAPE RMSE MAPE RMSE MAPE RMSE

Chen (1996) 1.20 0.493 1.39 0.503 1.20 0.459 1.15 0.383 1.24 0.425
Yu (2005) 0.41 0.176 0.27 0.269 0.60 0.239 0.30 0.132 0.78 0.300
Chang et al. (2007)—1‐order 1.31 0.659 0.87 0.334 0.11 0.053 0.27 0.095 0.86 0.345
Chang et al. (2007)—n‐order 0.56 0.243 0.54 0.199 0.11 0.053 0.27 0.095 0.86 0.345
ARIMA 0.50 0.565 0.15 0.233 0.22 0.261 1.75 2.653 0.30 0.343
GRNN (Specht, 1991) 1.14 0.999 1.12 0.663 1.73 0.460 0.69 0.531 1.15 0.821
Proposed A 0.22 0.094 0.26 0.107 0.18 0.079 0.08 0.040 0.34 0. 127
Proposed B 0.30 0. 144 0.31 0.127 0.21 0.089 0.15 0.071 0.44 0.187

Note. “Proposed A” denotes proposed one‐order model; “Proposed B” is proposed 3‐order model; and the highlighted entries denote the best performance
in eight models for each year dataset.

TABLE 9 Profits comparisons of the gold price

Models Profits

Chen (1996) 7,030

Yu (2005) 7,438
Proposed A 11,429
Proposed B 12,882

Note: The highlighted entries denote the best performance in eight models for each year dataset.

TABLE 10 Profits comparisons of the exchange rate

Years
Models 2004 2005 2006 2007 2008

Chen (1996) 1.983 −1.004 0.256 −0.036 1.700

Yu (2005) 0.233 3.144 1.100 0.036 3.440
Proposed A 10.275 11.560 11.350s 6.269 14.486
Proposed B 10.631 10.812 10.951 6.180 14.077

Note: The highlighted entries denote the best performance in eight models for each year dataset.

jforecastðtÞ−actualðtÞj
If ≤α and forecast(t + 1) − actual(t) < 0,
actualðtÞ
then buy,
where α is threshold parameter (the threshold parameter depends on the data set). For this reason, the definition of profit is defined as
Equation 19.

p q
profit ¼ ∑ ðactualðt þ 1Þ−actualðtÞÞ þ ∑ ðactualðtÞ−actualðt þ 1ÞÞ; (19)
ts ¼1 tb ¼1

where p is the total number of days for selling, q is the total number of days for buying, and ts and tb are the t‐th day for selling and buying.
Therefore, this paper uses one and half years (2004/04–2005/09) gold price data set and the 5 years (2004–2008) exchange rate data set to
compute the profit and compare the proposed model with Chen's model (1996) and Yu's model (2005) as Tables 9 and 10. From the com-
pared profit results, the proposed Models A and B have more profits than Chen's model (1996) and Yu's model (2005). From profit evaluation
of this study, we can find that the better forecast accuracy could get the better profit as Table 7–10, and the timing of sell/buy depends on
the optimal threshold parameter α in training data set. However, in stock market, the main return factor is investment strategy (threshold
value) for the better profit (Ma, Xiong, He, & Zhang, 2017). Ma et al. (2017) built a basic trading strategy using directional change approach,
to select the optimal threshold θ and study the market performance for different portfolios that consist of different estimated period and
investment period.

6 | C O N CL U S I O N S

This study has proposed a novel fuzzy time series forecasting model, the proposed model can be more objective and reasonable to fuzzify historical
data, and based on Apriori's AR theory to build the weighted high‐order fuzzy rules. From the experiments of the gold price and the exchange rate
CHENG AND CHEN 13 of 15

data sets, the results shown that the proposed Models A and B outperform the listing models under MAPE, RMSE, and profit criteria. The reason
why the proposed model is better than the listing models, there are four facets as following:

1. Objectively determining lower bound and upper bound

After objectively calculated the length of interval by Equation 5, the proposed A automatically decides the lower/upper bound, and the dis-
course of universe is more reasonable than the past methods. That is, the proposed A objectively handles linguistic intervals, because it partitions
the linguistic intervals based on the distribution of observations. Therefore, the proposed A is more objective than Chen's model (Chen, 1996) and
Yu's model (Yu, 2005).

2. Fitting the number of order by association rules

Association rules can find the relationship between the previous orders and next order. In financial markets, the price not only influenced by
previous order but also the change of the previous price is one of the most important factor for forecasting next order (Cheng & Wei, 2009). From
the experimental results, three previous orders price influence the next price in the gold market. In the exchange rate data set, the next price had
affected by a previous orders price. Therefore, this paper utilizes association rules to forecast with better performance than other methods in
financial market.

3. Data‐derived weighted association rule

The weight of each fuzzy rule is obtained by calculating the cardinality with the same fuzzy relations. Nevertheless, Chen (1996) uses the cen-
tre of gravity method to defuzzify with the equal weights. For this reason, the weights of association rules are more reasonable than Chen's model
(Chen, 1996).

4. Overcome nonlinear problem

ARIMA is linearly time series models. However, in real worlds, there are many uncertainty and nonlinear interaction. The proposed model uses
the nonlinear rule‐based method (fuzzy sets theory) to deal this problem; therefore, the proposed method can get better performance than ARIMA.
Although the proposed method has made great improvement, still has room to improve. In future work, we can do more improvements: (a) In
order to prove its generality, the proposed model can be applied in other area, such as energy (Sadaei, Guimarães, da Silva, Lee, & Eslami, 2017) and
medical (Zhang et al., 2016). (b) This study, the high‐order forecasting, is combined the previous three orders data to forecast next order. In future
work, we can utilize more than three orders to forecast.

CONF LICT S OF INT ER E ST

None.

ORCID

Ching‐Hsue Cheng http://orcid.org/0000-0002-5509-6965

Chung‐Hsi Chen http://orcid.org/0000-0002-1367-8529

RE FE R ENC E S
Agrawal, R., & Srikant, R. (1994). Fast algorithm for mining association rules. Proceedings of the VLDB conference, 487‐499.
Box, P., & Jenkins, G. M. (1976). Time series analysis: Forecasting and control. San Francisco, CA: Holdenday Inc.
Chang, J. R., Lee, Y. T., Liao, S. Y., & Cheng, C. H. (2007). Cardinality‐based fuzzy time series for forecasting enrollments. Lecture Notes in Computer Science,
4570, 735–744.
Chen, S. M. (1996). Forecasting enrollments based on fuzzy time series. Fuzzy Sets and Systems, 81, 311–319.
Chen, S. M. (2002). Forecasting enrollments based on high‐order fuzzy time series. Cybernetics and Systems, 33, 1–16.
Chen, S. M., & Chung, N. Y. (2006). Forecasting enrollments using high‐order fuzzy time series and genetic algorithms. International Journal of Intelligent Sys-
tems, 21, 485–501. Ku
Chen, S. M., & Hsu, C. C. (2004). A new method to forecast enrollments using fuzzy time series. Applied Science and Engineering, 2, 234–244.
Chen, S. M., & Hwang, J. R. (2000). Temperature prediction using fuzzy time series. IEEE Transactions on Systems Man and Cybernetics Part B, 30(2), 263–275.
Cheng, C. H., Chang, J. R., & Yeh, C. A. (2006). Entropy‐based and trapezoid fuzzification‐based fuzzy time series approaches for forecasting IT project cost.
Technological Forecasting and Social Change, 73(5), 524–542.
Cheng, C. H., Chen, T. L., Teoh, H. J., & Chiang, C. H. (2008). Fuzzy time‐series based on adaptive expectation model for TAIEX forecasting. Expert Systems
with Applications, 34, 1126–1132.
14 of 15 CHENG AND CHEN

Cheng, C. H., Cheng, G. W., & Wang, J. W. (2008). Multi‐attribute fuzzy time series method based on fuzzy clustering. Expert Systems with Application, 34(2),
1235–1242.
Cheng, C. H., & Wei, L. Y. (2009). Fusion ANFIS models based on multi‐stock volatility causality for TAIEX forecasting. Neurocomputing, 72(16–18),
3462–3468.
Faff, R. W., Brooks, R. D., & Kee, H. Y. (2002). New evidence on the impact of financial leverage on beta risk: A time‐series approach. North American Journal
of Economics and Finance, 13, 1–20.
Huarng, K. (2001a). Effective lengths of intervals to improve forecasting in fuzzy time series. Fuzzy Sets and Systems, 123, 387–394.
Huarng, K. (2001b). Heuristic models of fuzzy time series for forecasting. Fuzzy Sets and Systems, 123(3), 369–386.
Huarng, K., & Yu, H. K. (2003). An N‐th order heuristic fuzzy time series model for TAIEX forecasting. International Journal of Fuzzy Systems, 5(4), 247–253.
Huarng, K., & Yu, H. K. (2004). Type 2 fuzzy time series for TAIEX forecasting. Paper presented at the Taiwan–Japan symposium 2004 on fuzzy system and
innovational computing.
Huarng, K., & Yu, H. K. (2005). A type 2 fuzzy time‐series model for stock index forecasting. Physica A, 353, 445–462.
Huarng, K. H., & Yu, H. K. (2006). The application of neural networks to forecast fuzzy time series. Physica A, 336, 481–491.
Hwang, J. R., Chen, S. M., & Lee, C. H. (1998). Handling forecasting problems using fuzzy time series. Fuzzy Sets and Systems, 100, 217–228.
Kantardzic, M. (Ed.) (2002, October). Data mining: concepts, models, methods, and algorithms. New Jersey: Wiley‐IEEE Press.
Lee, L. W., Wang, L. H., & Chen, S. M. (2007). Temperature prediction and TAIFEX forecasting based on fuzzy logical relationships and genetic algorithms.
Expert Systems with Applications, 33(3), 539–550.
Lee, L. W., Wang, L. H., & Chen, S. M. (2008). Temperature prediction and TAIFEX forecasting based on high‐order fuzzy logical relationships and genetic
simulated annealing techniques. Expert Systems with Applications, 34, 328–336.
Liu J. W., Cheng C. H., Su C. H., Tsai M. C., “Fuzzy multiple‐periods time‐series model based on GSP and AR for forecasting gold price”, Advanced Materials
Research, Vols. 211–212, pp. 1124–1128, 2011 (2011 INTERNATIONAL CONFERENCE ON MECHATRONICS AND INTELLIGENT MATERIALS,
Guilin, China, 2011).
Ma, J., Xiong, X., He, F., & Zhang, W. (2017). Volatility measurement with directional change in Chinese stock market: Statistical property and investment
strategy. Physica A: Statistical Mechanics and its Applications, 471, 169–180.
Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits on our capacity of processing information. Psychological Review, 63, 81–97.
Ross, T. J. (Ed.) (2000). Fuzzy logic with engineering applications (International ed.). New York: McGraw‐Hill.
Sadaei, H. J., Guimarães, F. G., da Silva, C. J., Lee, M. H., & Eslami, T. (2017). Short‐term load forecasting method based on fuzzy time series, seasonality and
long memory process. International Journal of Approximate Reasoning, 83, 196–217.
Shin, H. W., & Sohn, S. Y. (2004). Segmentation of stock trading customers according to potential value. Expert Systems with Applications, 27, 27–33.
Singh, P., & Borah, B. (2013). An efficient time series forecasting model based on fuzzy time series. Engineering Applications of Artificial Intelligence, 26(10),
2443–2457.
Song, Q., & Chissom, B. S. (1993a). Forecasting enrollments with fuzzy time series—Part I. Fuzzy Sets and Systems, 54(1), 10.
Song, Q., & Chissom, B. S. (1993b). Fuzzy time series and its models. Fuzzy Sets and Systems, 54(3), 269–277.
Song, Q., & Chissom, B. S. (1994). Forecasting enrollments with fuzzy time‐series—Part II. Fuzzy Sets and Systems, 62, 1–8.
Specht, D. (1991). A general regression neural network. IEEE Transactions on Neural Networks., 2, 568–576.
Su, C. H., & Cheng, C. H. (2016). A hybrid fuzzy time series model based on ANFIS and integrated nonlinear feature selection method for forecasting stock.
Neurocomputing, 205, 264–273.
Teoh, H. J., Chen, T. L., Cheng, C. H., & Chu, H. H. (2009). A hybrid multi‐order fuzzy time series for forecasting stock markets. Expert Systems with Applica-
tions, 36, 7888–7897.
Wang, Y. F. (2002). Predicting stock price using fuzzy grey prediction system. Experts Systems with Applications, 22, 33–39.
Yolcu, O. C., & Lam, H. K. (2017). A combined robust fuzzy time series method for prediction of time series. Neurocomputing, 247, 87–101.
Yu, H. K. (2005). Weighted fuzzy time series models for TAIEX forecasting. Physica A, 349, 609–624.
Zadeh, L. A. (1965). Inform. Control. Fuzzy Sets and Systems, 8, 338–353.
Zadeh, L. A. (1975a). The concept of a linguistic variable and its application to approximate reasoning—Part I. Information Science, 8, 199–249.
Zadeh, L. A. (1975b). The concept of a linguistic variable and its application to approximate reasoning II. Information Science, 8, 301–357.
Zadeh, L. A. (1976). The concept of a linguistic variable and its application to approximate reasoning III. Information Science, 9, 43–80.
Zadeh, L. A. (1988). Fuzzy Logic. IEEE Computer, 21, 83–93.
Zhang, T., Zhang, X., Liu, Y., Luo, Y., Zhou, T., & Li, X. (2016). The analysis of infectious disease surveillance data based on fuzzy time series method. Inter-
national Journal of Infectious Diseases, 45, 309–310.

Ching‐Hsue Cheng received the bachelor's degree in Mathematics from Chinese Military Academy in 1982, the master's degree in Applied
Mathematics from Chung‐Yuan Christian University in 1988, and the PhD degree in System Engineering and management from National
Defense University in 1994. He is now a professor of information management department in National Yunlin University of Science and Tech-
nology. His research is mainly in the field of fuzzy logic, fuzzy time series, soft computing, reliability, medical information management, and data
mining. He has published more than 200 papers (include 195 significant journal papers).
CHENG AND CHEN 15 of 15

Chung‐Hsi Chen was born in Tainan, Taiwan. He got his master's degree in Management Information System from National Yunlin University of
Science and Technology, Taiwan. Currently he is a PhD student at the National Yunlin University of Science and Technology, Taiwan, and now
his is a teacher of at Nanzih elementary school, Tainan.

How to cite this article: Cheng C‐H, Chen C‐H. Fuzzy time series model based on weighted association rule for financial market forecast-
ing. Expert Systems. 2018;e12271. https://doi.org/10.1111/exsy.12271