Soft Comput (2008) 12:1105–1113

DOI 10.1007/s00500-008-0278-z

ORIGINAL PAPER

Forecasting analysis by using fuzzy grey regression model

for solving limited time series data
Ruey-Chyn Tsaur

Published online: 5 February 2008

© Springer-Verlag 2008

Abstract The grey model GM(1,1) is a popular forecasting require at least 50 and preferably 100 observations. How-
method when using limited time series data and is success- ever, it is sometimes impossible to collect 50 data or more for
fully applied to management and engineering applications. forecasting a new product demand or a new system develop-
On the other hand, the reliability and validity of the grey ing in today’s rapidly changing socio-economic situations.
model GM(1,1) have never been discussed. First, without Therefore, other forecasting models have been developed
considering other causes when using limited time series data, to cope with the problem when collected data is limited
the forecasting of the grey model GM(1,1) is unreliable, and and violates the basic assumption of normal distribution of
provide insufficient information to a decision maker. There- standard statistical models. The grey model with first-order
fore, for the sake of reliability, the fuzzy set theory was differential equation and one dependent variable model is
hybridized into the grey model GM(1,1). This resulted in the referred to as the grey model GM(1,1) (Deng 1986, 1989)
fuzzy grey regression model, which granulates a concept into and was introduced in management and engineering appli-
a set with membership function, thereby obtaining a possible cations for solving limited time series data. For example,
interval extrapolation. Second, for a newly developed product Hsu and Wen (2000) hybridized the grey model GM(1,1) and
or a newly developed system, the data collected are limited applied it to a multi-objective programming model to fore-
and rather vague with the result that the grey model GM(1,1) cast airline city-pair passenger traffic; Lin and Yang (2003)
is useless for solving its problem with vague or fuzzy-input applied the grey model GM(1,1) to forecast the output value
values. In this paper the fuzzy grey regression model is veri- of Taiwan’s optoelectronics industry; Hsu and Chen (2003)
fied to show its validity in solving crisp-input data and fuzzy- improved the grey model GM(1,1) to forecast the power
input data with limited time series data. Finally, two examples demand of Taiwan; Chiao and Wang (2002) provided a
for the LCD TV demand are illustrated using the proposed practical method to improve the lifetime of fluorescent
models. lamps; and Hsu (2003) applied three types of residual
modification models to revise the grey model GM(1,1) and
Keywords Grey model GM(1,1) · Forecasting · Fuzzy forecast the short-term demand in the integrated circuit indus-
regression model · Fuzzy grey regression model try. It is obvious that the grey model GM(1,1) is a good
method to use in the forecasting with limited time series
data. However, without considering other causes, the relia-
1 Introduction bility of the forecasting point-estimations obtained from the
grey model GM(1,1) are not stable when we take another
The time series model is a popular method for the forecast- sampling, and it does not provide sufficient information to
ing of economic, marketing, as well as social problems that a decision maker. In addition, when it comes to validity, a
newly developed product or a newly developed system, the
data collected are usually limited and rather vague, with the
R.-C. Tsaur (B)
result that the validity of the grey model GM(1,1) becomes
Department of Finance, Hsuan Chuang University,
48, Hsuan Chuang Road, Hsinchu 300, Taiwan useless for solving a problem using vague or fuzzy-input
e-mail: [email protected] values.

123
1106 R.-C. Tsaur

Zadeh (1965) introduced the concept of the fuzzy set with membership function, thereby obtaining the possibly
theory, rather then use the point estimation in the conven- interval extrapolation. Next, the fuzzy grey regression model
tional probability theory. Also, he used the fuzzy set theory is verified to show its validity in solving crisp-input data
to granulate a concept into a set with membership function or fuzzy-input data with limited time series data. Finally,
thereby decreasing the amount of data required. Tanaka et al. two examples for LCD TV demand are illustrated by our
(1982) extended this idea, and introduced the fuzzy regres- proposed models. The rest of this paper is organized as fol-
sion in his proposal of a non-parameter approach for evaluat- lows. Section 2 reviews the grey model GM(1,1) and the
ing the relation between independent variables and dependent fuzzy regression model. In Sect. 3, the fuzzy grey regression
variables. Kim et al. (1996) showed that its forecasting error model for solving limited data is constructed. Two examples
was better than that of statistical regression if the collected are illustrated in Sect. 4, and finally, conclusions are drawn
data were smaller by simulation and in comparison with those in Sect. 5.
two models. In addition, fuzzy regression models have been
applied to various problems such as forecasting and engi-
neering. In forecasting, Watada (1992) used fuzzy regres- 2 Reviewing the grey model GM(1,1) and the fuzzy
sion for time series analysis; Chang (1997) showed that the regression model
fuzzy regression model could be better explained in seasonal
analysis; Tseng et al. (2001) stated that Watada’s method 2.1 Grey model GM(1,1)
had a large forecasting error, and proposed the fuzzy ARIMA
(Auto-Regressive Integrated Moving Average) method to If an original time series set f0 is defined as
obtain a reliable forecasting interval. The lags of the differ-



enced series appearing in the forecasting equation are called f0 = { f t0
t ∈ 1, 2, . . . , n} (1)
“auto-regressive” terms, lags of the forecast errors are called
“moving average” terms, and a time series which needs to where t denotes the number of data observed in period t, then
be differenced in order to be made stationary is said to be an the AGO value f t1 of the original time series f t0 is obtained
“integrated” version of a stationary series. Tsaur et al. (2002) as
used two independent variables of preceding periodical data  t 

and index of time to show the pattern of the seasonal vari- ft =
1
fk0
t = 1, 2, . . ., n. (2)
ation; Tsaur (2003) showed that the fuzzy regression model k=1
could be used to forecast the demand of internet users under
The grey model GM(k, N ) (Deng 1986, 1989) is defined as
different product life cycles. In engineering, Chang et al.
Eq. (3) where k stands for the kth-order derivative of the
(1996) used the fuzzy linear regression model to analyze a
dependent variables Ft1 , and N stands for N variables (i.e.
typical ergonomics problem, and showed that these types of
one dependent variable Ft1 and N − 1 independent variables
analyses could accurately represent the experimental data;
X 11 (t), X 21 (t), . . . , X 1N −1 (t)).
Funga et al. (2006) used an asymmetric fuzzy linear regres-
sion to estimate the functional relationships for product plan- d k Ft1 d k−1 Ft1 d Ft1
ning based on quality function deployment by integrating the + a 1 + · · · + a k−1 + ak Ft1
dt k dt k−1 dt
least-squares regression into fuzzy linear regression, and pro-
= b1 X 11 (t) + b2 X 21 (t) + · · · + b N −1 X 1N −1 (t) (3)
posed a hybrid linear programming model in order to cope
with the typical vagueness or imprecision of functional rela- where a1 , a2 , . . ., ak and b1 , b2 , . . ., b N −1 are unknown para-
tionships in a product under a fuzzy environment; Abdalla meters. If k = 1 and N = 1, then the grey model GM(1,1)
and Buckley (2007) apply fuzzy Monte Carlo method to a with first-order differential equation and one dependent vari-
certain fuzzy linear regression problem to estimate the best able model can be constructed as
solution, and find that the best solution is a vector of trian-
gular fuzzy numbers for the fuzzy coefficients in the model. d Ft1
+ a Ft1 = b, t = 1, 2, . . ., n (4)
In addition, they use a quasi-random number generator to dt
produce random sequences of these fuzzy vectors which uni- where a represents the unknown developed parameter, b rep-
formly fill the search space. resents the unknown grey controlled parameter, and Ft1 is the
However, for the fuzzy ARIMA model it is still necessary dependent variable with AGO input value f t1 . For solving
to collect a lot of data to derive the parameters of ARIMA d Ft1
model (4), the derivative dt for the dependent variable is
(p,d,q), otherwise the model cannot work. In this study, for
represented as
reliability under limited time series, the fuzzy set theory is
hybridized into the grey model GM(1,1) to obtain the fuzzy d Ft1 F 1 − Ft1
grey regression model, and to granulate a concept into a set = lim t+h , ∀t ≥ 1. (5)
dt h→0 h

123
Forecasting analysis by using fuzzy grey regression model for solving limited time series data 1107

Because the collected data is a set of time-series, we tionship between independent variables and the dependent
assume the sampling time interval between period t and t+1 variable. The basic model assumes a fuzzy regression equa-
d F1 tion as follows
to be one unit. Then, the derivative dtt can be approxi-
mated to be the difference between two successive periods of Ỹi = Ã0 + Ã1 X i1 + · · · + Ã N X i N = ÃX i (12)
1 , defined as an inverse accumulated generating
Ft1 and Ft+1
operation (IAGO) variable Ft+10 as where X i = [1, X i1 , . . . , Xi N ]T is a vector of independent
variables for the i-th data; Ã0 , Ã1 , . . . , Ã N is a vector of
d Ft1 F 1 − Ft1
≈ t+1 = Ft+1
1
− Ft1 = Ft+1
0
, ∀t ≥ 1 (6) the fuzzy parameters presented in the form of symmetric
dt 1 triangular fuzzy numbers denoted by à j = (α j , c j ), j =
for the original (t + 1)-th time series data f t+1
0 , ∀t ≥ 1. In 0, 1, . . ., N , with its membership function described as (13)
order to have a more steady value for the dependent variable below where α j is its central value and c j is its spread value.
Ft1 , ∀t ≥ 1, the second part of model (4) is suggested as the ⎧
⎪ |α −a |
average of two successive periods of Ft1 and Ft+1 1 , ∀t ≥ 1 ⎨ 1 − jc j j , α j − c j ≤ a j ≤ α j + c j ,
(Deng 1986, 1989). Then, we can rewrite model (4) as µ Ã j (a j ) = ∀ j = 0, 1, . . . , N (13)
⎪
⎩
  0, otherwise
1 1
0
Ft+1 = a − (Ft+1 + Ft1 ) + b, ∀t ≥ 1. (7)
2 By applying the Extension Principle (Zadeh 1965), the
derived membership function of fuzzy number Ỹi is shown
If t = 1, 2, . . . , n −1, then (7) can be rewritten into matrix
as (14)
form as ⎧
⎡ 0⎤ ⎡ 1 1 ⎤ ⎪ |Y −X α|
⎨ 1− ciT |Xii | , X i = 0,
F2 − 2 (F2 + F11 ) 1
⎢ F0 ⎥ ⎢ ⎥  u(Yi ) = 1, X i = 0, Yi = 0 ∀i=1, 2, . . . , M
⎢ 3⎥=⎢ − 21 (F31 + F21 ) 1⎥ a
. ⎪
⎩ 0,
⎣...⎦ ⎢ ⎣ ...
⎥
.⎦ b
(8) X i = 0, Yi = 0
Fn0 − 21 (Fn1 + Fn−1 1 ) 1 (14)

By applying the least square method with input data sets f1 The aim of a fuzzy regression model is to find the narrow-
and f0 , the parameters of a and b in matrix â can be solved est fuzzy regression interval as described in (15) by requiring
as that the membership degree of each observation Yi is at least
  equal to the value h as shown in (16) below.
a  M 
â = = (BT B)−1 BT F0 (9)
b N 


MIN cj
Xi j
(15)
⎡ 0⎤ ⎡ 1 1 ⎤
f2 − 2 ( f 2 + f 11 ) 1 j=0 i=1
⎢ f0⎥ ⎢ 1 1


3 ⎥ , B = ⎢ − 2 ( f3 + f2 )
1 1⎥
Yi −X T α

where matrices F0 = ⎢ ⎣...⎦ ⎣
⎥ i
≥ h, ∀i = 1, 2, . . . , M
... .⎦ 1−
c T |X i |
(16)
f n0 − 21 ( f n1 + f n−1
1 ) 1
T Therefore, the following linear programming model can
and B are the transpose of matrix B. Then, the differential
be obtained for solving the fuzzy regression equation:
equation of model (4) can be solved to obtain the estimated
 M 
value fˆt+1
1 for the dependent variable F 1 , ∀t ≥ 1 as
t  N 


 MIN cj
Xi j

fˆt+1
1
= f 10 − (b/a) e−at + (b/a), ∀t ≥ 1. (10) j=0 i=1

N 
N



Finally, the estimated value fˆt+1
0 for the IAGO variable F 0 s.t. α j X i j +(1−h) c j
X i j
≥ Yi , ∀i = 1, 2, . . . , M
t+1
is obtained as j=0 j=0

N 
N



fˆt+1
0
= fˆt+1
1
− fˆt1 , ∀t = 1, 2, . . . , n. (11) α j X i j − (1 − h) c j
X i j
≤ Yi , ∀i = 1, 2, . . . , M
j=0 j=0
Therefore, by inputting the time series f 10 , f 20 , . . . , f n0 c ≥ 0, a ∈ , X i0 = 1, 0 ≤ h < 1; ∀i = 1, 2, . . . , M
into the grey model GM(1,1), it can obtain the extrapolative (17)
value of fˆ20 , fˆ30 , . . . , fˆn0 , and fˆn+1
0 .

Then, formula (12) can be rewritten as

2.2 The fuzzy regression model
Ỹi = (α0 , c0 ) + (α1 , c1 )X i1 + · · · + (α N , c N )X i N (18)
The fuzzy regression was first introduced by Tanaka et al. Each value of the dependent variable can be estimated as a
(1982). It is an alternative approach to evaluating the rela- fuzzy number Ỹi = (YiL , Yih=1 , YiU ), i = 1, 2, . . . , M where

123
1108 R.-C. Tsaur

the lower bound of Ỹi is YiL = (α − c)T X i ; the center value u A~ ( x)

i

of Ỹi is Yih=1 = α T X i ; the upper of Ỹi is YiU = (α + c)T X i

1
and c T = (c0 , c1 , . . . , c N ), α T = (α0 , α1 , . . . , α N ).
The degree of fitness of the estimated fuzzy regression
equation Ỹi = Ã0 + Ãi X i to the given data Yi is measured
by index h with Yih = {Yi | u Ỹi (Yi ) ≥ h}. The value of h
is a membership degree which requires that the collected
data are included in the derived fuzzy regression interval
at least to the degree h. Moskowitz and Kim (1993) pro-
posed that if one is confident with the collected data, then
x
a smaller value h is assigned, otherwise, a larger value h ai-ci ai ai+ci
should be given. Besides, Moskowitze and Kim also sug-
Fig. 1 The membership function of fuzzy parameter Ãi
gested that “If the solution for a fuzzy regression model is
obtained as à j,h 1 = (α ∗j , c∗j ), then the solution is changed
into à j,h 2 = (α ∗j , 1−h
1−h 1 ∗
c ) when the confidence value h is
2 j
adjusted from h 1 to h 2 ”.
For the sake of simplicity, we define the fuzzy parameters
Ã0 = (a0 , c0 ) and Ã1 = (a1 , c1 ) in the fuzzy grey regres-
sion model as symmetrical triangular fuzzy numbers with
3 Fuzzy grey regression model
central values a0 , a1 , and spread values c0 , c1 , respectively.
Besides, their figure and membership functions are formu-
In this section, for the purpose of reliability, we first hybridize
lated in Fig. 1 and Eq. (19)
the fuzzy set theory into the grey model GM(1,1) in Sect. 3.1
⎧
to evaluate the fuzzy relationship between dependent and ⎪
⎪ 1 − (aic−x) , ai − ci ≤ x ≤ ai
independent variables, and granulate a concept into a set with ⎪
⎨ i
(x−a )
membership function, thereby obtaining the possible interval u Ãi (x) = 1 − ci i , ai ≤ x ≤ ai + ci , ∀i = 0, 1
⎪
⎪
extrapolation. Second, for a newly developed product or a ⎪
⎩ 0, otherwise.
newly developed system the data collected are limited and
rather vague such that the grey model GM(1,1) is useless for (19)
solving a problem with vague or fuzzy-input values.
By applying the Extension Principle, the derived membership
Therefore, the fuzzy grey regression model is verified in 0 is shown as (20)
function of the fuzzy-output value of F̃t+1
Sect. 3.2 to show its validity in solving fuzzy-input data under
⎧
0 

f − a0 +a1 f 1



   
limited time series data. ⎪
⎪ 1− t+1 t
, a0 + a1 f t1 − c0 + c1 f t1
⎪
⎪ c0 +c1 f t1
⎪
⎪  
⎨ ≤ f t+1
0 ≤ a +a f1
3.1 Fuzzy grey regression model with crisp-input 0 1 t
0
u( f t+1 )=  
⎪
⎪
and fuzzy-output model ⎪
⎪ + c0 + c1 f t1
⎪
⎪
⎩

0, o.w.
For a limited time series set f
0 = { f t0
t ∈ 1, 2, . . . , n}, an (20)
AGO time series set f1 = { f t1
t ∈ 1, 2, . . . , n} is obtained as

f t1 = tk=1 f k0 . In this study, in order to forecast with limited The aim of a fuzzy grey regression model is to find the
time series data, we first hybridize the fuzzy set theory into narrowest regression interval by requiring that the member-
the grey model GM(1,1) to obtain the fuzzy grey regression 0 is at least equal to the
ship degree of each observation f t+1
model using limited data. Then, the fuzzy grey regression value h as shown in (21) below with 0 ≤ h < 1.
model with crisp-input and fuzzy-output value is constructed
0  


f 1

as per Definition 1 by fuzzyfying the grey model GM(1,1) of t+1 − a0 + a1 f t
1− ≥ h , ∀t = 1, 2, 3,…, n − 1.
Eq. (7). c0 + c1 f t1
(21)
Definition 1 The fuzzy grey regression model with crisp-
input and fuzzy-output value is defined as F̃t+10 = Ã0 + Finally, in order to obtain the minimum fuzzy relation in
Ã1 Ft , ∀t ≥ 1, where the fuzzy parameters Ã0 and Ã1 are
1 the fuzzy grey regression model, it is necessary to require the
fuzzy sets on the product space of the parameter for mapping spread values of the fuzzy-output value F̃t+1 0 , ∀t =
the input AGO time series f t1 of the independent variable 1, 2, 3, . . . , n − 1 to be as small as possible. Based on this
Ft1 to the fuzzy-output value of the fuzzy dependent variable objective, we can obtain the following linear programming
0 .
F̃t+1 model

123
Forecasting analysis by using fuzzy grey regression model for solving limited time series data 1109


n−1 ~
Membership function of f t +01
Min (c0 + c1 f t1 )
t=1
0 ≥ a − (1 − h)c + (a − (1 − h)c ) ∗ ( f 1 )
f t+1 (22) 1
0 0 1 1 t
0 ≤ a + (1 − h)c + (a + (1 − h)c ) ∗ ( f 1 )
f t+1 0 0 1 1 t ~
Membership function of Ft 0+1
a0 , a1 ∈ R, c0 , c1 ≥ 0, 0 ≤ h < 1, ∀t = 1, 2, . . . , n − 1

By solving the linear programming model (22), the fuzzy

parameters Ã0 = (a0 , c0 ) and Ã1 = (a1 , c1 ) can be solved.
ht
The equation of the fuzzy grey regression model is obtained
0
as F̃t+1 = (a0 , c0 ) + (a1 , c1 )Ft1 . This equation provides a
forecasting interval constructed byeach estimated fuzzy out-
0 = F 0,L , F 0,h t =1 , F 0,U
put F̃t+1 t = 1, 2, . . . , n where
Ft 0+1L f t 0+1L Ft 0+1 f t 0+1 f t 0+U1 Ft 0+U1
t+1 t+1 t+1
0,L
Ft+1 = (a0 − c0 ) + (a1 − c1 ) · f t1 is the lower bound of Fig. 2 The relation between membership functions of fuzzy -input and
0 ; F 0,h t =1 = a + a · f 1 is the center value of F̃ 0 ,
F̃t+1 fuzzy-output values
t+1 0 1 t t+1
0,U
and Ft+1 = (a0 + c0 ) + (a1 + c1 ) · f t1 is the upper bound of
0 .
F̃t+1 0 ≤ h < 1, ∀t = 1, 2, 3, . . ., n − 1. Then, we can derive
Therefore, by the crisp-input values f 10 , f 20 , . . . , f n0 , the
0
− (1 − h)et0 ≥ a0 − (1 − h)c0 + (a1 − (1 − h)c1 )
extrapolative fuzzy-output values F̃ˆ 0 , F̃ˆ 0 , . . . , F̃ˆ 0 , and
f t+1
2 3 n
∗( f t1 − (1 − h)et1 )
F̃ˆn+1
0 (23)
can be obtained to granulate a concept into a set with
membership function.
0
f t+1 + (1 − h)et0 ≤ a0 + (1 − h)c0 + (a1 + (1 − h)c1 )
∗( f t1 + (1 − h)et1 ). (24)
3.2 Fuzzy grey regression model with fuzzy-input
and fuzzy-output model Finally, in order to obtain the minimum fuzzy relation in
the fuzzy grey model GM(1,1), it is necessary to require the

0 , ∀t = 1, 2, 3,
0
spread values of the fuzzy-output value F̃t+1
If an original fuzzy time series set f̃ = { f˜t0
t ∈ 1, 2, . . . , n}
. . . , n − 1 to be as small as possible. Based on this objective,
with central value f t0 , and spread value et0 , ∀t ∈ 1, 2, . .
. , n we can obtain the following linear programming model:
1

is collected, then an AGO fuzzy time series set f̃ = { f˜t1
t ∈
 −1
n
1, 2, . . . , n} is obtained with central value f t1 = tk=1 f k0 , Min (c0 + a1 et1 + c1 f t1 )

and spread value et1 = tk=1 ek0 ∀t ∈ 1, 2, . . . , n, by the t=1
0 − (1 − h)e0 ≥ a − (1 − h)c
f t+1
concept of fuzzy addition. Then, the fuzzy grey regression t 0 0
model with fuzzy-input and fuzzy-output value is constructed +(a1 − (1 − h)c1 ) ∗ ( f t1 − (1 − h ) et1 ) (25)
0 + (1 − h)e0 ≤ a + (1 − h)c + (a + (1 − h)c )
f t+1
as per Definition 2 by fuzzyfying the grey model GM(1,1) of t 0 0 1 1
Eq. (7). ∗( f t1 + (1 − h)et1 )
a0 , a1 ∈ R, c0 , c1 ≥ 0, 0 ≤ h < 1, ∀t = 1, 2, . . . , n − 1
Definition 2 The fuzzy grey regression model with fuzzy-
input and fuzzy-output value is defined as F̃t+1 0 = Ã0 + By solving the linear programming model (25), the fuzzy
parameters Ã0 = (a0 , c0 ) and Ã1 = (a1 , c1 ) can be solved.
Ã1 F̃t1 ,∀t ≥ 1, where the fuzzy parameters Ã0 and Ã1 are
The equation of the fuzzy grey regression model is obtained
fuzzy sets on the product space of the parameter for mapping 0
as F̃t+1 = (a0 , c0 ) + (a1 , c1 ) F̃t1 . This equation provides
the AGO fuzzy time series f˜t1 of the fuzzy independent vari-
a forecasting interval constructed by each estimated
 fuzzy-
able F̃t1 to the fuzzy-output value of fuzzy dependent variable 0,L 0,h t =1 0,U
0 .
F̃t+1 output value F̃t+1 = Ft+1 , Ft+1 , Ft+1 , ∀t = 1, 2,
0

0,L
. . . , n where Ft+1 = (a0 − c0 ) + (a1 − c1 ) · ( f t1 − et1 ) is the
For the sake of simplicity, we define the fuzzy parame- lower bound of F̃t+1 0 ; F 0,h t =1 = a + a · f 1 is the center
t+1 0 1 t
ters Ã0 = (a0 , c0 ) and Ã1 = (a1 , c1 ) in the fuzzy grey 0,U
value of F̃t+1 , and Ft+1 = (a0 + c0 ) + (a1 + c1 ) · ( f t1 +
0
regression model as symmetrical triangular fuzzy numbers
with central values a0 , a1 , and spread values c0 ,c1 , respec- et1 ) is the upper bound of F̃t+1 0 . Finally, by the fuzzy-input

tively. By interval arithmetic, as in Fig. 2, a requirement values f˜10 , f˜20 , . . . , f˜n0 , the extrapolative fuzzy-output values
of the maximum degree of fit h between the fuzzy-output F̃ˆ 0 , F̃ˆ 0 , . . . , F̃ˆ 0 , and F̃ˆ 0 can be obtained. Therefore, the
2 3 n n+1
0
value F̃t+1 and the fuzzy-input value f˜t+10 = ( f t0 , et0 ) is fuzzy grey regression model verified its validity for solving
˜
presented as [ f t+1 ] ⊂ [ F̃t+1 ] (Tanaka et al. 1982) with
0 h 0 h fuzzy-input data under limited time series data.

123
1110 R.-C. Tsaur

Furthermore, the value h is referred to as the degree of fit Table 1 The sales of LCD TVs per year (unit: ten thousand)
between the fuzzy-input values and the extrapolative fuzzy- Year 2001 2002 2003 2004
output values of the fuzzy grey regression model, which thus
determines the range of the possible distributions of the pro- LCD TV Sales 81 150 393 870
posed model. That is, the value h is subjectively selected by a
decision maker as a membership degree which requires that
the collected data are included in the derived fuzzy interval Table 2 The solutions for the variables
to at least the degree h. Moskowitz and Kim (1993) proposed Variable a0 a1 c0 c1
that if one is confident with the collected data, then a smaller Solution with value h = 0 64.65 1.33 22.05 0
value h can be assigned, otherwise a larger value h should
Solution with value h = 0.3 64.65 1.33 31.5 0
be given. When h increases, the spread of the fuzzy-output
Solution with value h = 0.5 64.65 1.33 44.11 0
value becomes wider. Tanaka and Watada (Tsaur et al. 2002)
Solution with value h = 0.7 64.65 1.33 73.52 0
suggested that h = 0 when the data set is sufficiently large,
and to use a comparatively higher h as the size of the data set
becomes smaller. Bardossy et al. (1990) selected the value for
h according to the decision maker’s belief in the model, gen- extrapolative fuzzy-output LCD TV demand is estimated and
erally recommending an h value between 0.5 and 0.7. With shown in Table 3, with a lower and upper value for each year.
the property of h, the resultant forecasting interval provides A better forecasting model means that it has a smaller esti-
more flexibility for a decision maker. mated error and can use the training data to obtain a more
accurate extrapolative value (compared to the testing value).
The central value of the fuzzy-output LCD TV demand is
4 Illustrated examples used for defuzzying the fuzzy-output values of column 1
and column 3 in Table 3. Then we find that the fuzzy grey
Since 1998, the growth of the TFT-LCD industry in Taiwan regression model has the smallest forecasting error among the
has been more rapid than the growth in the semiconductor grey model GM(1,1), Watada’s model, and the linear regres-
industry. The initial application of TFT-LCD was in the note- sion model, with a estimated error MAPE (Mean Absolute
book, but it is now also applied to PC and TV monitors. Dis- Percentage Error, MAPE) of 29.44% from year 2002 to 2004.
play Search (Optotech 2005) reported that LCD TV demand In addition, the testing data of the year 2005 for the LCD TV
has surged worldwide, with all regions enjoying at least 41% demand shows that the proposed model extrapolates the LCD
sequential growth. Topology research institute (2005) pre- TV demand in 2005 to be between 20,296,200 units (lower
dicted that LCD TV will have a 50% market share in the value under h = 0) and 20,737,200 units (upper value under
TV industry by 2008 if the LCD TV price can be reduced h = 0), which matches the possible demand forecasting of
through lower panel costs and new capacity optimization. some research institutes. For example, it is believed that the
In order to get a larger market share, TFT-LCD companies LCD TV demand in 2005 was more than 20,000,000 units.
have been forecasting future demand in the LCD TV market The comparison results are plotted in Figs. 3 and 4.
and planning a new generation of production line for large In Fig. 3, the linear regression model forecasting value for
size panels. However, it is difficult to collect the world-wide the 2005 LCD TV demand is 10,260,000, and the grey model
sales of different sizes of LCD TVs so we roughly collected GM(1,1) forecasting value for the 2005 LCD TV demand is
the secondary data from some research institutes. Because 16,698,000 which does not fit the rapid increasing LCD TV
the collected time series data is limited and the secondary demand in 2005. In Fig. 4, based on the empirical results of
data collected from different research institutes is vague, the this example, we found that the predictive capability of the
fuzzy grey regression model as proposed in Sect. 3 is a good fuzzy grey regression model is rather encouraging and that
method to forecast the future demand of LCD TV for the the possible interval of the fuzzy grey regression model is
time period 2002 to 2005. narrower than Watada’s interval. For example, for the fore-
casting demand in 2002, the lower bound and upper bound
4.1 Example 1 for crisp-input LCD TV demand of the proposed model are 1,503,300 and 1,944,300 whereas
the lower bound and upper bound in Watada’s model are
The secondary data of LCD TV demand with crisp-input 1,261,824 and 3,399,960, a range that is too wide for a deci-
value is shown in the second row of Table 1. By inputting sion maker to make a decision on under such an uncertain
the values of LCD TV sales into the model (22) and setting environment. We also found that Watada’s model is a fuzzy
the value h as 0, 0.3, 0.5, and 0.7, respectively, as per the linear regression model that cannot fit the rapid increas-
suggestions of Tanaka et al, and Bardossy et al, the solutions ing LCD TV demand. However, our proposed model is a
of the fuzzy parameters are obtained in Table 2. Next, the piecewise model which can more reliably forecast a rapidly

123
Forecasting analysis by using fuzzy grey regression model for solving limited time series data 1111

Table 3 The extrapolative value F̃ˆt0 with different models (unit: ten thousand)
Forecasting value Fuzzy grey regression Grey model Watada’s Fuzzy regression Linear regression
model with value h=0 GM(1,1) model with value h=0 model

F̃ˆ2002
0 (150.33,194.43) 156.13 (126.1824, 339.996) 243
ˆ
F̃ 0 (349.83,393,93) 343.99 (389.0736, 602.994) 504
2003

F̃ˆ2004
0 (872.52,916.62) 757.89 (651.9648, 865.992) 765
F̃ˆ 0
2005 (2029.62, 2073.72) 1669.80 (914.856, 1128.99) 1026
MAPE 23.12% 29.44% 94.37% 101.16%

3000
Collected data

2500 Lower bound of the

Watada's fuzzy regression
model
Upper bound of the
2000 watada's fuzzy regression
model
Lower bound of the fuzzy
grey regression model
1500
Upper bound of the fuzzy
grey regression model

1000

500

0
1 2 3 4 5

-500

Fig. 3 Comparison results between linear regression model and grey

Fig. 4 Comparison results between fuzzy regression model and fuzzy
model GM(1,1)
grey regression model

Table 4 The sales of LCD TVs per year (unit: ten thousand)
Year 2001 2002 2003 2004 analyze the fuzzy-input data and transfer to useable informa-
tion, we use the fuzzy grey regression model for forecasting
LCD TV Sales (81, 8) (150, 12) (393, 16) (870, 20)
the LCD TV demand, and compare to Watada’s model. By
inputting the fuzzy-input values of the LCD TV sales into
the model (25) and by setting the value h as 0, 0.3, 0.5, and
increasing demand. These evidences show that the perfor- 0.7, respectively, as per the suggestions of Tanaka et al, and
mance of the fuzzy grey regression model is better than that Bardossy et al, then the solutions of the fuzzy parameters
of the grey model GM(1,1), the linear regression model and can be obtained in Table 5. Next, the LCD TV demand can
Watada’s fuzzy regression model under limited information. be estimated, as shown in Table 6, as a fuzzy-output value
with a lower and upper value for each year. The results show
4.2 Example 2 for the fuzzy-input LCD TV demand that the proposed model extrapolates the LCD TV demand in
2005 to be between 20,251,110 units (lower value at h = 0)
In this example, we revised the collected secondary data in and 22,048,610 units (upper value at h = 0), which matches
Table 1 into fuzzy data by adding the spread value, as shown the possible demand forecasting of some research institutes.
in Table 4. For example, the LCD TV demand in 2001 is Figure 5 shows that Watada’s fuzzy regression model has
listed as (81, 8) which means that the central value is 81 and a larger forecasting interval than the fuzzy grey regression
∼
the spread value is 8 to express fuzzy number 81. We know interval, and Watada’s model remains a linear trend captur-
that the grey model GM(1,1) and the statistical regression ing the future demand, while the fuzzy grey regression model
model are not used for analyzing fuzzy-input data. In order to is a piecewise trend that fits the rapidly increasing LCD TV

123
1112 R.-C. Tsaur

Table 5 The solutions for the variables Lower bound of the fuzzy Upper bound of the fuzzy
Variable a0 a1 c0 c1 grey regression model grey regression model
Lower bound of the Upper bound of the
Solution with value h = 0 49.5311 1.3825 12.4553 0 collected data collected data
Solution with value h = 0.3 40.2435 1.4239 0 0.0616 Lower bound of the Upper bound of the
Solution with value h = 0.5 39.7807 1.4262 0 0.0727 Watada's fuzzy regression Watada's fuzzy regression
Solution with value h = 0.7 40.8943 1.4241 0 0.0818 model model
2500

Table 6 The extrapolative values F̃ˆt0 with different models (unit: ten
thousand) 2000

Forecasting value Fuzzy grey regression Tanaka’s Fuzzy regression

model with value h=0 model with value h=0
1500

F̃ˆ2002
0 (137.9983, 185.0289) (110.1544, 356)
ˆ
F̃ 0 (328.7833, 408.9939) (377.0316, 623)
2003 1000
F̃ˆ2004
0 (849.9858, 974.4364) (643.9088, 890)
F̃ˆ 0
2005 (2025.111, 2204.861) (910.786, 1157)
500

0
demand. For example, in the forecasting demand of 2002,
the lower bound and upper bound of the proposed model 1 2 3 4 5

were 1,379,983 and 1,850,289, while the lower bound and -500
upper bound in Watada’s model was 1,101,544 and 3,560,000
which is too wide for a decision maker to make a decision Fig. 5 Comparison results between Watada’s fuzzy regression model
on under an uncertain environment. In addition, as per the and fuzzy grey regression model
forecasting demand in some research institutes, the fore-
casting interval (20,251,110, 22,048,610) in the fuzzy grey (i) Fuzzy grey regression model can provide the decision
regression model is more precise forecasting for the LCD TV makers the best- and worst-possible situations.
demand in 2005 than Watada’s fuzzy regression model. (ii) The required observations are as little as four.

4.3 Discussions and analysis A comparison of four kinds of time-series methods is shown
in Table 7.
It is evident that the reliability and validity of the fuzzy grey
regression model were examined successfully. By using the
collected limited time series data, we obtained the possible 5 Conclusion
forecasting interval for the LCD TV demand. In addition, we
found that when the collected LCD TV demand is fuzzy-input In this paper, based on the basic concepts of the grey model
time series data, then the proposed fuzzy grey regression GM(1,1) and Tanaka’s fuzzy regression model, we proposed
model can still be used for forecasting, and the forecasting a new method (the fuzzy grey regression model) and applied
intervals are narrower allowing the decision maker to easily it to forecasting the LCD TV demand for showing the reliabil-
determine the trend of future LCD TV demand. In short, ity and validity of the proposed method. From the examples
our proposed method can be used to assist managers in the it is evident that the proposed method not only makes good
LCD TV industry to understand the possible interval in the forecasts but also provides the decision makers with the best
macro-economic environment. and worst-possible scenarios. The performance of the fuzzy
Although the basic concept of the grey model GM(1,1) grey regression model is better than the linear regression
and the fuzzy set theory is used to formulate the fuzzy grey model, the grey model GM(1,1) and Watada’s fuzzy regres-
regression model, the output of the fuzzy grey regression sion model. Although Display Search, iSuppli, and PIPA
model requires fewer observations than the linear regression research institutes collected sufficient information from the
model. There are several situations for which the fuzzy grey LCD TV market, in general, they could have made more pre-
regression model appears to be the most appropriate tool, cise forecasts. Using limited data with uncertain information,
such as: and by using the fuzzy grey regression model we were able to

123
Forecasting analysis by using fuzzy grey regression model for solving limited time series data 1113

Table 7 Comparison of four kinds of methods

Grey Model GM(1,1) Linear Regression Model Watada’s Fuzzy Regression Fuzzy Grey Regression Model
Model

Apply a one-order AGO to assem- A functional relationship between The relationship between the The relationship between the
ble the collected time series data independent variables and the independent variables and the input AGO variable and the output
and obtain internal regularity in dependent variable. dependent variable is a fuzzy variable is a fuzzy function
order to manage the disorga- function
nized original data
Point estimation. Provides confidence interval. Provides possibility interval. Provides possibility interval.
Limited time series data Under the assumption of Normal This model is seldom used to solve Limited time series data
distribution, at least 20 or more limited time series data
observations are preferable

forecast the LCD TV demand with crisp-input or fuzzy-input Hsu CC, Chen CY (2003) Applications of improved grey prediction
data, such that the TFT-LCD producers can use it to plan their model for power demand forecasting. Energy Conversion and
Management 44(14):2241–2249
new generation plant expansion and strengthen their compet- Hsu LC (2003) Applying the grey prediction model to the global
itive edge, meet new and ongoing challenges, and maximize integrated circuit industry. Technological Forecasting & Social
their profit in the LCD TV market. Change 70(6):563–574
Kim KJ, Moskowitz H, Koksalan M (1996) Fuzzy versus statistical
Acknowledgments The authors acknowledge the financial support linear regression. European Journal of Operational Research
from the National Science Council of Taiwan, R.O.C., under project 92(2):417–434
number NSC 93-2213-E-364-002. Lin CT, Yang SY (2003) Forecast of the output value if Taiwan’s opto-
electronics industry using the grey forecasting model. Technologi-
cal Forecasting & Social Change 70(2):177–186
Menozzi M, Napflin U, Krueger H (1999) CRT versus LCD: A pilot
References study on visual performance and suitability of two display tech-
nologies for use in office work. Displays 20(1):3–10
Abdalla A, Buckley JJ (2007) Monte Carlo methods in fuzzy linear Moskowitz H, Kim K (1993) On assessing the h value in Fuzzy Linear
regression. Soft Computing 11(10):991–996 Regression. Fuzzy Sets and Systems 58(3):303–327
Bardossy A, Bogardi I, Duckstein L (1990) Fuzzy Regression in Tanaka H, Uejima S, Asai K (1982) Fuzzy linear regression model.
Hydrology. Water Resources Research 26(7):1497–1508 IEEE Transaction on SMC 12(6):903–907
Chang PT (1997) Fuzzy seasonality forecasting. Fuzzy Sets and Tsaur RC, Wang HF, O Yang JC (2002) Fuzzy Regression for Seasonal
Systems 90(1):1–10 Time Series Analysis. International Journal of Information Tech-
Chang PT, Lee ES, Konz SA (1996) Applying fuzzy linear regression nology & Decision Making 1(1):165–175
to VDT legibility. Fuzzy Sets and Systems 80(2):197–204 Tsaur RC (2003) Extrapolating internet users in Taiwan by risk assess-
Chang SC (2005) The TFT–LCD industry in Taiwan: competitive ment. Computers and Mathematics with Applications 46(10–11):
advantages and future developments. Technology in Society 27(2): 1725–1734
199–215 Tseng FM, Tzeng GH, Yu HC, Yuan Benjamin JC (2001) Fuzzy
Cheng CB (2005) Fuzzy process control: construction of control charts ARIMA model for forecasting the foreign exchange market. Fuzzy
with fuzzy numbers. Fuzzy Sets and Systems 154(2):287–303 Sets and Systems 118(1):9–19
Chiao CH, Wang WY (2002) Reliability improvement of fluorescent Watada J (1992) Fuzzy time series analysis and forecasting of sales
lamp using grey forecasting model. Microelectronics Reliability volume. In: Kacprzyk J, Fedrizzi M (eds.), Fuzzy Regression
42(1):127–134 Analysis, Omnitech Press, Warsaw and Physica-Verlag, Heidel-
Deng LJ (1986) Grey forecasting and decision. Huazhong of Society berg 211–227
and Technology Press, Wuhan Zadeh LA (1965) Fuzzy sets, Inform. Control l.8(3):338–353
Deng LJ (1989) Introduction to grey system theory. The Journal of Grey Optotech, Chinese Information Co. LTD, Taipei, n.74 (2005) (In
System 1(1):1–24 Chinese)
Funga RYK, Chen Y, Tang J (2006) Estimating the functional relation- Topology research institute, Open a new world of LCD : Trend Analysis
ships for quality function deployment under uncertainties. Fuzzy in LCD TV market, Topology, Taipei (2005) (In Chinese)
Sets and Systems 157(1):98–120
Hsu CI, Wen YH (2000) Application of grey theory and mutiobjective
programming towards airline network design. European Journal
of Operational Research 127(1):44–68

123