IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 40, NO.
5, OCTOBER 2010 1255
A Stochastic HMM-Based Forecasting
Model for Fuzzy Time Series
Sheng-Tun Li, Member, IEEE, and Yi-Chung Cheng
AbstractRecently, fuzzy time series have attracted more aca-
demic attention than traditional time series due to their capability
of dealing with the uncertainty and vagueness inherent in the
data collected. The formulation of fuzzy relations is one of the
key issues affecting forecasting results. Most of the present works
adopt IFTHEN rules for relationship representation, which leads
to higher computational overhead and rule redundancy. Sullivan
and Woodall proposed a Markov-based formulation and a fore-
casting model to reduce computational overhead; however, its
applicability is limited to handling one-factor problems. In this
paper, we propose a novel forecasting model based on the hidden
Markov model by enhancing Sullivan and Woodalls work to allow
handling of two-factor forecasting problems. Moreover, in order
to make the nature of conjecture and randomness of forecasting
more realistic, the Monte Carlo method is adopted to estimate the
outcome. To test the effectiveness of the resulting stochastic model,
we conduct two experiments and compare the results with those
from other models. The first experiment consists of forecasting the
daily average temperature and cloud density in Taipei, Taiwan,
and the second experiment is based on the Taiwan Weighted
Stock Index by forecasting the exchange rate of the New Taiwan
dollar against the U.S. dollar. In addition to improving forecasting
accuracy, the proposed model adheres to the central limit theorem,
and thus, the result statistically approximates to the real mean of
the target value being forecast.
Index TermsForecasting, fuzzy time series, hidden Markov
model (HMM), Monte Carlo method.
I. I NTRODUCTION
T HE FORECASTING problem of time series data plays an
important role in various domains, such as air pollution,
population growth, rainfall prediction, and stock forecasting.
It deals with forecasting future outcomes from a temporally
ordered sequence of past observed data points, whose values are
usually real numbers. However, traditional time series analysis
cannot handle the vagueness and uncertainty inherent in certain
data due to inaccuracies in measurements, incomplete sets of
observations, or difficulties in obtaining the measurements [1].
Manuscript received December 9, 2007; revised June 24, 2008 and
April 20, 2009; accepted November 1, 2009. Date of publication December 18,
2009; date of current version September 15, 2010. This work was supported
by the National Science Council, Taiwan, under Contracts NSC96-2416-H-
006-011-MY2 and NSC97-2410-H-165-003. This paper was recommended by
Associate Editor A. Nuernberger.
S.-T. Li is with the Institute of Information Management and the Department
of Industrial and Information Management, National Cheng Kung University,
Tainan 701, Taiwan (e-mail: stli@mail.ncku.edu.tw).
Y.-C. Cheng is with the Department of International Business Management,
Tainan University of Technology, Yongkang, Tainan County 710, Taiwan
(e-mail: t20042@mail.tut.edu.tw).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSMCB.2009.2036860
1083-4419/$26.00 2009 IEEE
Recently, the theory of fuzzy logic has been widely recognized
as a successful approach for dealing with data uncertainty.
For time series, the uncertain values can be modeled as fuzzy
variables, resulting in so-called fuzzy time series [1]. The term
fuzzy time series has been used with several different meanings
[2]: 1) time series with uncertain single data (fuzzy data) at each
point in time [3], [4]; 2) time series with fuzzified real-valued
single data at each point in time [5]; and 3) fuzzy time series
based on a set of elementary finite time series and composed
of several significant representative courses [6]. It is the second
definition that will be used for the rest of this paper.
In 1993, Song and Chissom introduced the theory of fuzzy
logic into forecasting time series problems and proposed a
new paradigm known as fuzzy time series, capable of dealing
with vague and incomplete data represented as linguistic values
under uncertain circumstances [5], [7], [8]. They established
a four-step framework to manipulate the forecasting problem:
1) determine and partition the universe of discourse into inter-
vals; 2) define fuzzy sets from the universe of discourse and
fuzzify the time series; 3) derive fuzzy relationships existing
in the fuzzified time series; and 4) forecast and defuzzify the
forecasting outputs. Finally, they validated their model using
enrollment data from the University of Alabama.
Since Song and Chissoms pioneering work, a number of
related research works have been reported that follow their
framework and aim to improve forecasting accuracy and/or
reduce the computational overhead. These works include order-
1 [9][15], high-order [16][21], single-factor [9], [10], [12]
[14], [18], [21], two-factor [11], [17], [19], [20], and multi-
factor models [15]. All of them rely on constructing groups
of fuzzy relationships in advance represented as IFTHEN
rules, which come with higher computational overhead and
rule redundancy [22]. In order to alleviate the computational
overhead in deriving the fuzzy relationship rules, Sullivan and
Woodall proposed a so-called Markov-based model by using
conventional matrix multiplication [13]. In their model, the
fuzzy relationships are represented by a time-invariant Markov
process, in which the relationship matrix is determined by the
probability distributions of linguistic values and each row of
the matrix is unit normalized. However, their model does not
take into consideration the frequency of state transitions on
the estimation of the relationship matrix, resulting in missing
information and causing the forecasting to be unsatisfactory
[23]. In addition, it is only applicable to one-factor forecasting
problems. In the real world, an event may be affected by inter-
related factors, and thus, a forecasting model that can handle
multiple ones will be more realistic; therefore, the results that it
produces will be more useful [19], [24].
1256 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 40, NO. 5, OCTOBER 2010
In this paper, we propose a novel forecasting model by
enhancing Sullivan and Woodalls Markov-based forecasting
one to allow handling two-factor forecasting problems. This
model is built on the basis of the hidden Markov model (HMM),
a probabilistic model that is commonly applied to time series
[25]. Moreover, by applying the Monte Carlo method when
estimating the forecasting outcome, the nature of conjecture
and randomness of the forecasting are made more realistic
[1]. To test the effectiveness of our model, we conduct two
experiments in forecasting daily average temperature and the
Taiwan Weighted Stock Index and compare the results with
those from other models.
The remainder of this paper is organized as follows. In
Section II, the basic concept of fuzzy time series is briefly
introduced, and in Section III, the new forecasting model based
on HMM is proposed. Section IV presents a performance
evaluation of the model and a comparison of the results. The
last section describes our conclusions and directions for future
work.
II. F UZZY T IME S ERIES
In this section, we briefly describe the concept of fuzzy time
series and its forecasting framework.
The definition of fuzzy time series used in this paper was first
proposed by Song and Chissom [7].
Definition 1: Let Y (t) (t = . . . , 0, 1, 2, . . .), a subset of R,
be the universe of discourse on which fuzzy sets fi (t) (i =
1, 2, . . .) are defined, and let F (t) be a collection of fi (t). Then,
F (t) is called a fuzzy time series on Y (t) (t = . . . , 0, 1, 2, . . .).
Song and Chissom employed a fuzzy relational equation to
develop their forecasting model under the assumption that the
observations at time t are dependent only upon the accumulated
results of the observations at previous times, which is defined
as follows.
Definition 2: If, for any fj (t) F (t), where j J, there
exist an fi (t 1) F (t 1), where i I, and a fuzzy relation
Rij (t, t 1), such that fj (t) = fi (t 1) Rij (t, t 1), let
R(t, t 1) = i,j Rij (t, t 1), where is the union oper-
ator and is the composition. R(t, t 1) is called the fuzzy
relation between F (t) and F (t 1), which can be represented
using the following fuzzy relational equation:
F (t) = F (t 1) R(t, t 1).
Definition 3: If we suppose that F (t) is caused by F (t 1),
F (t 2), . . ., or F (t m) (m > 0), then the first-order model
of F (t) can be expressed as
F (t) = F (t 1) R(t, t 1)
or
F (t) = (F (t 1) F (t 2) F (t m)) Ro (t, t m)
(2)
where is the union operator and is the composition.
R(t, t 1) is called the fuzzy relation between F (t) and F (t
1), and Ro (t, t k) is the fuzzy relation that joins F (t) with
F (t 1), F (t 2), . . ., or F (t k), where the subscript o
denotes the relationship or.
In the literature, the fuzzy relation Rij (t, t 1) is usually
represented by a fuzzy logical relationship rule (IFTHEN
rule), as in [5], [8][10], [16], [17], [19][22]. In this paper, the
fuzzy relation is realized by an HMM, which will be discussed
in the following section.
III. S TOCHASTIC HMM-BASED F ORECASTING M ODEL
Under realistic circumstances, there are usually multiple
related factors that influence the behavior and outcome of any
event. For example, when trying to predict todays temperature,
we could easily look up and observe the clouds in the sky. If
there are dense clouds, it can be intuitively inferred that the
temperature will be low. However, temperature depends on not
only cloud density but also temperature values in previous days.
We thus might obtain a better forecast for todays temperature
by combining knowledge about what happened in previous
days with the observed cloud state. These kinds of problems
are constantly encountered in the real world, which is why
our paper focuses on targeting them. Of course, the state of
the temperature is not merely controlled by both factors, as
elements such as winds and air pressure are also likely to have
an impact. However, in this paper, we limit ourselves to prob-
lems concerning two factors, in which both are probabilistically
related. This can be formally represented as follows.
Given two fuzzy time series F (t) = {fi (t)|t =
1, 2, . . . , T, i = 1, 2, . . . , n} and G(t) = {gi (t)|t =
1, 2, . . . , T, i = 1, 2, . . . , m}, where fi (t) and gi (t) are
the respective states at time t, the fuzzy relation among F (t),
G(t), and F (t 1) can be formulated as a fuzzy relational
equation
F (t) = (F (t 1), G(t)) R(t, t 1).
To solve the forecasting problem of fi (t), which is dependent
on fi (t 1) and gi (t), the theory of HMM is applied, in which
F (t) and G(t) are the hidden and observed state sequences,
respectively.
A. HMM
HMM is a statistical model to deal with symbols or signal
sequences that are assumed to be a Markov process [25],
[26]. The hidden Markov process is based on two essential
assumptions: 1) The next state is dependent only upon the
current state, and 2) each state-transition probability does not
vary in time, i.e., it is a time-invariant model.
(1) An HMM consists of two state sets and three
matrices of probabilities. The two sets are the hidden
state set S = {s1 , s2 , . . . , sn } and the observable state
set O = {o1 , o2 , . . . , om }, where the hidden states are
probabilistically related to the observable states and n and m
are the number of hidden and observable states, respectively.
The three relational probability matrices are defined between
the observable and hidden states: , A, and B. A triple
compact notation = (, A, B) is given to indicate the
LI AND CHENG: STOCHASTIC HMM-BASED FORECASTING MODEL FOR FUZZY TIME SERIES 1257

complete parameter set of an HMM. is a 1 n into m equal intervals, v1 , v2 , . . . , vm , with length lo being
initial state vector, denoted as = [1 , 2 , . . . , n ] = defined as lo = (1/m)[(Dmax o
+ D2o ) (Dmin
o
D1o )].
[Pr(s1,t=1 ), Pr(s2,t=1 ), . . . , Pr(sn,t=1 )], where i is the Step 2Defining the fuzzy sets on the universe of dis-
probability of each state occurring at initial time step t = 1. course and fuzzifying the time series: Given a traditional crisp
A is an n n state-transition matrix A = {aij }, in which time series, one needs a fuzzification procedure to obtain
aij is the state-transition probability from states si to sj , i.e., the corresponding fuzzy one. For hidden states, n fuzzy sets
aij = Pr(sj,t |si,t1 ) = Pr(si,t1 sj,t ). s1 , s2 , . . . , sn can be defined on U s using general membership
B is an n m confusion matrix B = {bij }, where bij is the functions, as expressed as follows:
probability of observing a state oj , given the hidden state si ,

n
i.e., bij = Pr(oj,t |si,t ). si = ij /uj (5)
The following are three major problems that the HMM j=1
solution has been successfully applied to:
1) evaluation: finding the probability of an observed se- where ij is the membership degree of si belonging to uj and
quence, given an HMM; is defined by
2) decoding: finding the sequence of hidden states that most 
1, if j = i
probably generated an observed sequence; ij = 0.5, if j = i 1 or i + 1
3) learning: generating an HMM, given a sequence of obser- 0, otherwise.
vations.
For problem 2), the Viterbi algorithm [27], [28] provides Then, for a given historical datum Yt , its membership degree
an effective way of finding the sequence of hidden states that belonging to interval ui is determined by the following heuristic
most probably generated an observed sequence, i.e., to find the rules.
single best state sequence Q = q1 q2 , . . . , qt , . . . , qT , for a given Rule 1) If Yt is located at u1 , the membership degrees are 1
observation sequence P = p1 p2 , . . . , pt , . . . , pT , where qt is a for u1 , 0.5 for u2 , and 0 otherwise.
hidden state at time t, qt S, and pt is an observable state at Rule 2) If Yt belongs to ui , 1 < i < n, then the degrees are
time t, pt O. Let t (i) be the best probability along a single 1, 0.5, and 0.5 for ui , ui1 , and ui+1 , respectively,
path at time t, which accounts for the first t observations and and 0 otherwise.
ends in state si Rule 3) If Yt is located at un , the membership degrees are 1
for un , 0.5 for un1 , and 0 otherwise.
t (i) = max Pr(q1 , q2 , . . . , qt = si , p1 , p2 , . . . , pt |). Then, Yt is fuzzified as sj , where the membership degree in
q1 ,q2 ,...,qt1
(3) interval j is maximal.
For observable states, m fuzzy sets o1 , o2 , . . . , om can be
To correctly retrieve the state sequence, the next best state is defined on U o , as expressed as follows:
obtained by
  
m

t+1 (j) = max t (i)aij bj (pt+1 ). (4) oi = ij /vj (6)

i j=1

In this paper, our objective is to perform a short-term forecast where ij is the membership degree of oi belonging to vj and
in a fuzzy time series, which fits problem 2), as described. is defined by

1, if j = i
ij = 0.5, if j = i 1 or i + 1
B. Forecasting Model
0, otherwise.
The proposed forecasting model expands Sullivan and
Woodalls model by combining HMM and Monte Carlos sim- The observation variable can then be fuzzified in the same way
ulation and consists of the following five steps. Please note that as the hidden variable.
if the given time series is a fuzzy time series, steps 1 and 2 are Step 3Modeling fuzzy logical relationships using HMM:
unnecessary. The objective of forecasting is to estimate the probabil-
Step 1Partitioning the universe of discourse into ity of hidden state si,t at time t, given the condition
several intervals of equal length: Let U s and U o be the that observable state ok,t is obtained at the same time,
discourse universes of hidden and observation variables, re- i.e., Pr(si,t |ok,t ). Following Bayes theorem, we have the
spectively. In general, U s and U o are defined as U s = [Dmin s
following:
D1 , Dmax + D2 ] and U = [Dmin D1 , Dmax + D2 ], where
s s s o o o o o
Pr(si,t ) Pr(ok,t |si,t )
s
Dmin s
, Dmax o
, Dmin o
, and Dmax are the respective minimal Pr(si,t |ok,t ) = (7)
Pr(ok,t )
and maximal values of the historical data of hidden and 
observation variables, and D1s , D2s , D1o , and D2o are proper Pr(si,t |sh,t1 ) Pr(sh,t1 ) Pr(ok,t |si,t )
positive numbers. U s is then partitioned into n equal = h .
intervals, u1 , u2 , . . . , un , with length ls being defined as Pr(ok,t )
ls = (1/n)[(Dmax s
+ D2s ) (Dmins
D1s )]. U o is partitioned (8)
1258 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 40, NO. 5, OCTOBER 2010

Therefore, a particular HMM can be characterized by the not exist; therefore, Pr(si,1 |ok,1 ) in (7) is derived as
following three matrices:
Pr(si,1 ) Pr(ok,1 |si,1 ) Pr(si,1 , ok,1 )
A = {aij }, where aij = Pr(sj,t |si,t1 ) Pr(si,1 |ok,1 ) = = .
Pr(ok,1 ) Pr(ok,1 )
B = {bij }, where bij = Pr(oj,t |si,t ) (13)

= {i }, where i = Pr(si,1 ). In such a way, the probability of hidden state si at t = 1 can

be determined by a function t (ok,1 ) defined as
The state-transition matrix A provides information about
the relation of two contiguous hidden states. The relation of 1 (ok,1 ) = B (:, [k])T
observation and hidden states is characterized as the confusion = [Pr(s1,1 ), Pr(s2,1 ), . . . , Pr(sn,1 )]
matrix B. is a vector with the probability of the initial state.
Now, we construct an HMM = (, A, B) whose parame- [Pr(ok,1 |s1,1 ), Pr(ok,1 |s2,1 ), . . . , Pr(ok,1 |sn,1 )]
ters are to be estimated by relative frequencies as follows. = [Pr(s1,1 ) Pr(ok,1 |s1,1 ), Pr(s2,1 ) Pr(ok,1 |s2,1 )
First, the initial state vector is set to be a 1 n matrix and . . . , Pr(sn,1 ) Pr(ok,1 |sn,1 )]
is defined as
= [Pr(ok,1 , s1,1 ), Pr(ok,1 , s2,1 ), . . . , Pr(ok,1 , sn,1 )] .
n(si,1 ) (14)
i = Pr(si,1 ) = (9)
N1
In general, the forecasting result is located at the state with
where n(si,1
) is the number of the initial state si in the data set the highest probability. However, according to the probability
and N1 = ni=1 n(si,1 ).
theorem, an event with a higher probability only has a greater
Next, the state-transition matrix A = {aij } is an n n ma-
chance of occurring but will not necessarily occur. Due to the
trix defined as follows:
nature of conjecture and randomness, a stochastic simulation, in
n(sj,t , si,t1 ) particular the Monte Carlo method, is adopted to estimate the
aij = Pr(sj,t |si,t1 ) = (10) outcome. The Monte Carlo method provides approximate solu-
n(si,t1 )
tions by stochastic sampling experiments and solves problems
 based on random numbers and probability statistics. The fore-
where, aij 0 and nj=1 aij = 1, 1 i n.
Finally, the confusion matrix B = {bij } is an n m matrix casting process consists of two subsequent tasks: normalization
represented as follows: and Monte Carlo simulation.
First, in order to compute the distribution of all possible fore-
n(oj,t , si,t ) casting outcomes, normalization is needed for the probability
bij = Pr(oj,t |si,t ) = (11) vectors of functions t (sh,t1 , ok,t ) and 1 (ok,1 ), which are
n(si,t )
expressed, respectively, as

where, bij 0 and m j=1 bij = 1, 1 i n.
Step 4Forecasting by stochastic simulation: As mentioned N t (sh,t1 , ok,t ) = [ts1 , ts2 , . . . , tsn ] (15)
earlier, the probability of si,t is determined by Pr(si,t |sh,t1 ) N 1 (ok,1 ) = [1s1 , 1s2 , . . . , 1sn ] (16)
and Pr(ok,t |si,t ), as derived in (8). The probabilities of all
possible hidden states occurring at time t (t 2) with the
where
transition influence of the previous hidden state sh,t1 and the
observation state ok,t are then computed. Such an influential
Pr(ok,t |si,t ) Pr(si,t |sh,t1 )
relation can be represented by the function t (sh,t1 , ok,t ) tsi = 
n (17)
defined as follows: Pr(ok,t |si,t ) Pr(si,t |sh,t1 )
i=1

t (sh,t1 , ok,t ) Pr(ok,1 , si,1 )

1si = 
n . (18)
= B (:, [k])T A ([h], :) Pr(ok,1 , si,1 )
i=1
= [(Pr(ok |s1 )Pr(s1,t |sh,t1 )), (Pr(ok |s2 )Pr(s2,t |sh,t1)) ,
. . . , (Pr(ok |sn )Pr(sn,t |sh,t1 ))] (12) Second, a stochastic experiment consisting of l Monte Carlo
simulations is conducted to determine the forecasting result si
where A([h], :) is the hth row of state-transition matrix A from N t (sh,t1 , ok,t ) or N 1 (ok,1 ). Such forecasting results
and B(:, [k]) is the kth column of confusion matrix B. The can be represented with a vector C
symbol of operator is an array multiplication, and thus,
A B means the element-by-element vector multiplication of C = [c1 , c2 , . . . , cn ] (19)
A and B.
For the special case when t = 1, where the observation state where ci 
is the number of forecasting hidden states belonging
is the only information available, the previous hidden state does to si and ni=1 ci = l.
LI AND CHENG: STOCHASTIC HMM-BASED FORECASTING MODEL FOR FUZZY TIME SERIES 1259

TABLE I TABLE II
T RAINING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR T RAINING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR
J UNE AND J ULY 1993 IN TAIPEI , TAIWAN AUGUST AND S EPTEMBER 1993 IN TAIPEI , TAIWAN

Step 5Defuzzifying the forecasting outputs: There are sev- TABLE III
eral defuzzification methods that can be chosen. For simplicity, T RAINING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR
J UNE AND J ULY 1994 IN TAIPEI , TAIWAN
we use the most popular one, namely, center of gravity, which
is expressed as

n
ci ti
i=1

n (20)
ci
i=1

where
1

(m1 + 0.5 m2 ), i=1
1.5
1
(0.5 mi1 + mi
ti = 2 (21)

+0.5 mi+1 ), i = 2, 3, . . . , n 1
1
1.5 (0.5 m n1 + mn ), i=n

and mi is the middle point of interval ui .

IV. E XPERIMENT AND E VALUATION

In order to demonstrate the effectiveness of the proposed
forecasting model, we conducted two experiments using real-
world data. The first experiment consisted of forecasting tem-
perature with observable cloud density in Taiwan, while the
second experiment was conducted to solve the forecasting
problem of Taiwans stock index with the observable exchange
rate of the New Taiwan dollar against the U.S. dollar. set (Tables IVI) to establish the parameter of , and that for
1996 (Tables VIIX) was used as the testing set to evaluate
forecasting performance.
A. Experiments Forecasting Temperature
Following the proposed forecasting model, the first step
With Cloud Density Data
was to determine and partition the universe of discourse. We
The time span of the data used is from June to September for partitioned temperature and cloud density into n and m even
1993, 1994, 1995, and 1996. The data were divided into two intervals u1 , u2 , . . . , un and v1 , v2 , . . . , vm , respectively. For
parts, i.e., that for 1993, 1994, and 1995 was used as the training the instance when n = 6 and m = 5, the temperature intervals
1260 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 40, NO. 5, OCTOBER 2010
TABLE IV
T RAINING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR
AUGUST AND S EPTEMBER 1994 IN TAIPEI , TAIWAN
TABLE V
T RAINING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR
J UNE AND J ULY 1995 IN TAIPEI , TAIWAN
were u1 = [21, 23), u2 = [23, 25), u3 = [25, 27), u4 = [27, 29),
u5 = [29, 31), and u6 = [31, 33], and the cloud density in-
tervals were v1 = [0, 20), v2 = [20, 40), v3 = [40, 60), v4 =
[60, 80), and v5 = [80, 100].
Next, the fuzzy sets were defined, and the time series was
fuzzified using (5). The linguistic hidden fuzzy sets were
defined as s1 = (freezing), s2 = (cold), s3 = (cool), s4 =
TABLE VI
T RAINING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR
AUGUST AND S EPTEMBER 1995 IN TAIPEI , TAIWAN
TABLE VII
T ESTING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR
J UNE 1996 IN TAIPEI , TAIWAN , AND F ORECASTED T EMPERATURE
(mild), s5 = (warm), and s6 = (hot), and the linguistic ob-
servable fuzzy sets were o1 = (very low), o2 = (low), o3 =
(medium), o4 = (high), o5 = (very high). With the fuzzy sets
being properly defined, the raw time series could be fuzzified.
For example, the temperature for June 2, 1993, was 25.3 C.
It was located at interval u3 , and the membership degree
vector was June 2,1993,s = [0, 0.5, 1, 0.5, 0, 0]. The maximum
membership degree appeared at u3 , with the corresponding
fuzzy set being s3 . On the other hand, the cloud density for
LI AND CHENG: STOCHASTIC HMM-BASED FORECASTING MODEL FOR FUZZY TIME SERIES 1261

TABLE VIII TABLE X

T ESTING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR T ESTING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR
J ULY 1996 IN TAIPEI , TAIWAN , AND F ORECASTED T EMPERATURE S EPTEMBER 1996 IN TAIPEI , TAIWAN , AND F ORECASTED T EMPERATURE

TABLE IX The state-transition matrix A = {aij } was a 6 6 matrix

T ESTING DATA OF AVERAGE T EMPERATURE AND C LOUD D ENSITY FOR resulting from (10)
AUGUST 1996 IN TAIPEI , TAIWAN , AND F ORECASTED T EMPERATURE

0.2500 0.2500 0.5000 0 0 0
0.1053 0.5263 0.3684 0 0 0

0 0.1379 0.5345 0.3103 0.0172 0
A= .
0 0.0174 0.1478 0.5217 0.3130 0

0 0 0 0.2230 0.7095 0.0676
0 0 0.0526 0.1579 0.3158 0.4737

The confusion matrix B = {bij } was a 6 5 matrix and was

obtained from (11)

0 0 0 0 1.0000
0.1429 0.0952 0.0952 0.0476 0.6190

0.0690 0.0690 0.0690 0.1724 0.6207
B= .
0 0.0259 0.1034 0.4138 0.4569

0.0270 0.1014 0.3514 0.3649 0.1554
0.1053 0.1579 0.4211 0.2105 0.1053

To forecast the temperature of June 1, 1996, on which the

cloud density at the same day was o5 , we obtained from (14)

June 1,1996 (o5,June 1,1996 )

June 2, 1993, was 96 and belonged to interval v5 , and the = B (:, [5])T
membership degree vector was June 2,1993,o = [0, 0, 0, 0.5, 1].
The fuzzy set was o5 because the maximum membership degree = [0.3333, 0, 0, 0.6667, 0, 0]
was at v5 . All the fuzzified results of the temperature and cloud [1, 0.6190, 0.6207, 0.4569, 0.1554, 0.1053]
density from 1993 to 1996 are shown in Tables IX.
In the third step, the fuzzy logical relationships existing in the = [0.3333, 0, 0, 0.304752, 0, 0]
fuzzy time series were modeled by = (, A, B) by counting
the frequency that hidden states occurred, as shown in (9) and the normalized vector was

= [0.3333, 0, 0, 0.6667, 0, 0]. N June 1,1996 (o5,June 1,1996 ) = [0.522371, 0, 0, 0.477629, 0, 0].
1262 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 40, NO. 5, OCTOBER 2010
We executed the Monte Carlo simulation 100 (l = 100) times
on the probability vector N June 1,1996 (o5,June 1,1996 ) and ob-
tained 100 stochastic forecasting hidden states, represented by
vector C
C = [53, 0, 0, 47, 0, 0]
of which 53 and 47 hidden states belonged to s1 and s4 ,
respectively. Finally, the crisp forecasting output was calculated
by carrying out defuzzification, which is defined in (20) (see the
equation shown at the bottom of the page).
Taking forecasting the temperature for June 4, 1996,
as another example, the cloud density on that day
was o2 , and the temperature state for June 3 was s5 .
Because the second column of confusion matrix B was
B(:, [2])T = [0, 0.0952, 0.0690, 0.0259, 0.1014, 0.1579] and
the fifth row of state-transition matrix A was A([5], :) = [0, 0,
0, 0.2230, 0.7095, 0.0676], the probability vector of hidden
states for June 4, 1996, could be calculated from (12)
June 4,1996 (s5,June 3,1996 , o2,June 4,1996 )
T
= B (:, [2]) . A ([5], :)
= [0, 0, 0, 0.000598, 0.025279, 0.0045].
This was further normalized to be
N June 4,1996 (s5,June 3,1996 , o2,June 4,1996 )
= [0, 0, 0, 0.019695, 0.832164, 0.148141].
It should be noted that the temperature for June 4, 1996,
had a greater chance of being s5 (warm), compared to s4 and
s6 , when the cloud density was high (o2 ) on that day and the
temperature for the previous day was warm (s5 ).
Next, we repeated the sampling experiments 100
(l = 100) times with the Monte Carlo simulation with
N June 4,1996 (s5,June 3,1996 , o2,June 4,1996 ) and obtained 100
stochastic forecasting hidden states, represented by vector C
C = [0, 0, 0, 5, 84, 11]

n
ci ti
i=1 53 22.6667 + 0 24 + 0 26 + 47 28 + 0 30 + 0 31.333

n =
100
ci
i=1

n
ci ti
i=1 0 22.6667 + 0 24 + 0 26 + 5 28 + 84 30 + 11 31.3333

n =
ci
i=1
= 30.0467
of which hidden states 5, 84, and 11 belonged to s4 , s5 , and s6 ,
respectively.
Finally, the crisp forecasting output was calculated by car-
rying out defuzzification, as defined in (20) (see the equation
shown at the bottom of the page).
All the forecasting results from June to September 1996 are
illustrated in Tables VIIX.
With the forecasting results, we then evaluated the perfor-
mance of the proposed fuzzy time series model by predicting
the temperature and comparing it with previous models. The
forecasting accuracy was measured in terms of the mean square
error (mse)
N
i=1 (F orecasting_V aluei Actual_V aluei )2
mse = .
N
The performance comparison was further conducted using
the average forecasting error percentage (AFEP), which is
defined as follows:
AF EP
1  |F orecasting_V alue Actual_V alue|
N
= 100%.
N i=1 Actual_V alue
In order to test the superiority of the proposed model, we
compared its performance with that of Lees model [19], which
is reported as having the best performance in the current lit-
erature. To make the comparison fair, the same intervals and
factors were used as that in Lees model. For Lees model,
the previous temperature state [F (t 1)] and the previous
cloud density state [G(t 1)] are the main and second factors,
respectively. In the proposed stochastic HMM-based model,
F (t 1) and G(t) are considered as the hidden and observable
states, respectively. The current temperature state [F (t)] is the
forecasting target of both models. Tables XI and XII illustrate
the results of Lees model with orders one to eight and that
of our model using 100 Monte Carlo simulations, in terms
of mse and AFEP based on two different interval partitions,
namely, n = 6 and m = 5, and n = 12 and m = 10. The results
= 25.1733
100
LI AND CHENG: STOCHASTIC HMM-BASED FORECASTING MODEL FOR FUZZY TIME SERIES
TABLE XI
P ERFORMANCE C OMPARISON B ETWEEN THE P ROPOSED M ODEL AND
L EE S M ODEL (n = 6, m = 5)
TABLE XII
P ERFORMANCE C OMPARISON B ETWEEN THE P ROPOSED M ODEL AND
L EE S M ODEL (n = 12, m = 10)
TABLE XIII
I MPACT OF THE N UMBER OF I TERATIONS OF M ONTE C ARLO S IMULATION
ON F ORECASTING P ERFORMANCE (n = 6, m = 5)
TABLE XIV
C OMPARISON OF THE F ORECASTING ACCURACY OF THE P ROPOSED
M ODEL W ITH VARIOUS M ONTE C ARLO S IMULATION T IMES
(n = 12, m = 10)
show that the proposed model outperformed Lees in all cases
and that it was also better than the classic HMM, which uses
maximum probability to decide the forecasting result (denoted
as HMM-based in the table). This demonstrates that our
approach is not only better at reflecting real-life situations but
also achieves superior performance results.
In order to investigate the sensitiveness of the forecasting
result to the number of iterations of the Monte Carlo simulation,
we tested several simulation runs, from 30 to 300 times, with
two different interval partitioning sets. The outcomes shown in
Tables XIII and XIV confirm that our model achieves better
results, compared to Lees model, in all numbers of simulation
runs. In addition, our results adhere to the central limit theorem,
and thus, the sampling distribution will have the same mean
as that of the population for large sample sizes (N 30),
confirming that the proposed model is indeed more accurate
than others in the literature.
1263
Fig. 1. Trend of the Taiwan Weighted Stock Index for 2005, 2006, and 2007.
Fig. 2. Trend of the exchange rate of the New Taiwan dollar against the U.S.
dollar for 2005, 2006, and 2007.
TABLE XV
P ERFORMANCE C OMPARISON OF THE TAIWAN
W EIGHTED S TOCK I NDEX (n = 6, m = 6)
B. Experiments Forecasting the Taiwan Weighted Stock Index
With Exchange Rate Data
In the second experiment, we applied the proposed model
to forecast the Taiwan Weighted Stock Index (hidden state)
using data from the exchange rate of the New Taiwan dollar
against the U.S. dollar (observable state) collected from 2005 to
2007, as shown in Figs. 1 and 2. The data set was divided into
the training set covering each January to October from 2005
to 2007 to build up the stochastic HMM model, with the test
set including each November to December from 2005 to 2007
for forecasting. Table XV illustrates the yearly comparison of
the proposed stochastic HMM model with Lees and traditional
HMM models, in which six intervals (n = 6, m = 6) were used
for both factors and all models. It shows that the proposed
model obtained better performance in terms of both mse and
AFEP. It can be seen that the forecasting results of all the
models were comparatively worse in 2007, and this is because
the yearly variation (range = 2465.32) was much lager than
those in 2005 (942.56) and 2006 (1569.92). Fig. 3 shows the
scatter diagrams of the predicted and actual indexes using
the proposed stochastic HMM and traditional HMM models,
showing a strong positive relationship between the two fore-
casting indexes. However, our model demonstrated a stronger
1264 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 40, NO. 5, OCTOBER 2010

Fig. 3. Correlation among predicted and actual stock indexes. (a) November to December of 2005. (b) November to December of 2006. (c) November to
December of 2007.
LI AND CHENG: STOCHASTIC HMM-BASED FORECASTING MODEL FOR FUZZY TIME SERIES
TABLE XVI
I MPACT OF THE N UMBER OF I TERATIONS OF M ONTE C ARLO S IMULATION
ON F ORECASTING P ERFORMANCE (n = 6, m = 6) FOR THE TAIWAN
W EIGHTED S TOCK I NDEX
relationship with the correlation coefficients of r = 0.968 (p <
0.01), r = 0.971 (p < 0.01), and r = 0.939 (p < 0.01) for the
years 2005, 2006, and 2007, respectively, as compared to the
traditional HMM model, with correlation coefficients of r =
0.962 (p < 0.01), r = 0.959 (p < 0.01), and r = 0.939 (p <
0.01). Table XVI shows the forecasting performance according
to different number of iterations of the Monte Carlo simulation
from 30 to 300, which all confirmed the superiority of the
proposed model over Lees and traditional HMM models.
V. C ONCLUSION AND D IRECTIONS FOR F UTURE W ORK
In this paper, we have proposed a novel stochastic fore-
casting model for fuzzy time series by extending Sullivan and
Woodalls Markov-based model. It was built upon an HMM in
which the fuzzy relationships were formulated as state transi-
tions so that it can handle two-factor forecasting problems. In
order to allow the model to more effectively reflect the real-
world situation and randomness of forecasting, Monte Carlo
simulation was applied to estimate the stochastic outcome.
The computations involved in forecasting were simple matrix
operations, and thus, our model is more efficient than other
IFTHEN-based models. We used the forecasting problem of
daily average temperature and cloud density in Taipei, Taiwan,
as the benchmark and conducted performance comparisons
with other models. The results demonstrated the superiority of
our model in forecasting. Moreover, the proposed probabilistic
forecasting model adheres to the central limit theorem, proved
by an experiment of sensitiveness, and thus, the forecasting
results statistically approximate to the real mean of the target
value being forecast. However, there are some limitations with
our model. First, it is developed based on HMM, in which the
observed and hidden states are probabilistically related, which,
of course, does not hold for all problems. The other shortcom-
ing is that when more zeros occur in state-transition matrix
A and confusion matrix B due to the less fuzzy relationships
existing in the historical temporal data, this tends to lead to the
hidden state being forecast as a zero probability vector. Dealing
with these limitations is of crucial importance for future work.
Other interesting directions include enhancing the proposed
model so that it is high order and/or time variant to improve
its forecasting power.
ACKNOWLEDGMENT
The authors would like to thank J.-Y. Li for his help with the
experimentation.
1265
R EFERENCES
[1] B. Mller and U. Reuter, Uncertainly Forecasting in Engineering.
Berlin, Germany: Springer-Verlag, 2007.
[2] B. Mller and U. Reuter, Prediction of uncertain structural responses
using fuzzy time series, Comput. Struct., vol. 86, no. 10, pp. 11231139,
May 2008.
[3] D. Hareter, Time series analysis with non-precise dataPart I, in Proc.
9th Spec. Conf. Probabilistic Mech. Struct. Reliability, 2004. [Online].
Available: http://www.cfd.sandia.gov.PMCpostconf.html
[4] D. Hareter, Time series analysis with non-precise dataPart II, in Proc.
9th Spec. Conf. Probabilistic Mech. Struct. Reliability, 2004. [Online].
Available: http://www.cfd.sandia.gov.PMCpostconf.html
[5] Q. Song and B. S. Chissom, Forecasting enrollments with fuzzy time
seriesPart I, Fuzzy Sets Syst., vol. 54, no. 1, pp. 19, Feb. 1993.
[6] S. F. Bocklisch and M. Pssler, Fuzzy time series analysis, in Advances
in Soft ComputingFuzzy Control, R. Hampel, M. Wagenknecht, and
N. Chaker, Eds. Berlin, Germany: Physica-Verlag, 2000, pp. 331345.
[7] Q. Song and B. S. Chissom, Fuzzy time series and its models, Fuzzy
Sets Syst., vol. 54, no. 3, pp. 269277, Mar. 1993.
[8] Q. Song and B. S. Chissom, Forecasting enrollments with fuzzy time
seriesPart II, Fuzzy Sets Syst., vol. 62, no. 1, pp. 18, Feb. 1994.
[9] S.-M. Chen, Forecasting enrollments based on fuzzy time series, Fuzzy
Sets Syst., vol. 81, no. 3, pp. 311319, Aug. 1996.
[10] S.-M. Chen and C.-C. Hsu, A new method to forecast enrollments using
fuzzy time series, Int. J. Appl. Sci. Eng., vol. 2, no. 3, pp. 234244,
2004.
[11] Y.-Y. Hsu, S.-M. Tse, and B. Wu, A new approach of bivariate fuzzy
time series analysis to the forecasting of a stock index, Int. J. Un-
certain. Fuzziness Knowl.-Based Syst., vol. 11, no. 6, pp. 671690,
2003.
[12] K. Huarng, Heuristic models of fuzzy time series for forecasting, Fuzzy
Sets Syst., vol. 123, no. 3, pp. 369386, Nov. 2001.
[13] J. Sullivan and W. H. Woodall, A comparison of fuzzy forecasting
and Markov modeling, Fuzzy Sets Syst., vol. 64, no. 3, pp. 279293,
Jun. 1994.
[14] R.-C. Tsaur, J.-C. Yang, and H.-F. Wang, Fuzzy relation analysis in fuzzy
time series model, Comput. Math. Appl., vol. 49, no. 4, pp. 539548,
Feb. 2005.
[15] K.-H. Huarng, T. H.-K. Yu, and Y. W. Hsu, A multivariate heuristic model
for fuzzy time-series forecasting, IEEE Trans. Syst., Man, Cybern. B,
Cybern., vol. 37, no. 4, pp. 836846, Aug. 2007.
[16] S.-M. Chen, Forecasting enrollments based on high-order fuzzy time
series, Cybern. Syst., Int. J., vol. 33, no. 1, pp. 116, Jan. 2002.
[17] S.-M. Chen and J.-R. Hwang, Temperature prediction using fuzzy time
series, IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 30, no. 2,
pp. 263275, Apr. 2000.
[18] J.-R. Hwang, S.-M. Chen, and C.-H. Lee, Handling forecasting problems
using fuzzy time series, Fuzzy Sets Syst., vol. 100, no. 13, pp. 217228,
Nov. 1998.
[19] L.-W. Lee, L.-H. Wang, S.-M. Chen, and Y.-H. Leu, Handling forecasting
problems based on two-factors high-order fuzzy time series, IEEE Trans.
Fuzzy Syst., vol. 14, no. 3, pp. 468477, Jun. 2006.
[20] L.-W. Lee, L.-H. Wang, and S.-M. Chen, Temperature prediction
and TAIFEX forecasting based on fuzzy logical relationships and ge-
netic algorithms, Expert Syst. Appl., vol. 33, no. 3, pp. 539550,
Oct. 2007.
[21] S.-T. Li and Y.-C. Cheng, Deterministic fuzzy time series model for
forecasting enrollments, Comput. Math. Appl., vol. 53, no. 12, pp. 1904
1920, Jun. 2007.
[22] S.-T. Li and Y.-C. Cheng, An enhanced deterministic fuzzy time series
forecasting model, Cybern. Syst., Int. J., vol. 40, no. 3, pp. 211235,
Apr. 2009.
[23] H.-K. Yu, A refined fuzzy time-series model for forecasting, Phys. A,
vol. 346, no. 3/4, pp. 657681, Feb. 2005.
[24] T. A. Jilani, S. M. Aqil Burney, and C. Ardil, Multivariate high order
fuzzy time series forecasting for car road accidents, in Proc. World Acad.
Sci. Eng. Technol., Jan. 2007, vol. 21, pp. 288293.
[25] L. R. Rabiner, A tutorial on hidden Markov models and selected appli-
cations in speech recognition, Proc. IEEE, vol. 77, no. 2, pp. 257286,
Feb. 1989.
[26] Y. Li, Hidden Markov models with states depending on observations,
Pattern Recognit. Lett., vol. 26, no. 7, pp. 977984, May 2005.
[27] G. D. Forney, The Viterbi algorithm, Proc. IEEE, vol. 61, no. 3,
pp. 268278, Mar. 1973.
[28] A. J. Viterbi, Error bounds for convolutional codes and an asymptotically
optimal decoding algorithm, IEEE Trans. Inf. Theory, vol. IT-13, no. 2,
pp. 260269, Apr. 1967.
1266 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 40, NO. 5, OCTOBER 2010
Sheng-Tun Li (M94) received the B.S. and M.S.
degrees in computer engineering from Tamkang
University, Tamsui, Taipei County, Taiwan, and the
Ph.D. degree in computer science from the Univer-
sity of Houston, University Park, TX.
He is currently a Professor with the Institute of
Information Management and the Department of
Industrial and Information Management, National
Cheng Kung University, Tainan, Taiwan. He is an
author or coauthor of five books, over 50 journal
articles, and numerous conference papers. He is a
holder of one patent. His research interests include knowledge engineering, data
mining, knowledge management system, and soft computing.
Dr. Li is a Member of IEEE.
Yi-Chung Cheng received the B.S. degree in busi-
ness mathematics from Soochow University, Taipei,
Taiwan, the M.S. degree in statistics from National
Central University, Jhongli City, Taiwan, and the
Ph.D. degree in industrial and information manage-
ment from National Cheng Kung University, Tainan,
Taiwan, in 1985, 1989, and 2008, respectively.
She is currently an Associate Professor with the
Department of International Business Management,
Tainan University of Technology, Yongkang, Tainan,
Taiwan. Her research interests include time series
forecasting and data mining.