Time series analysis, modeling and applications

W. Pedrycz & S.-M. Chen (Eds.): Time Series Analysis, Model. & Applications, ISRL 47, pp. 119–137.
DOI: 10.1007/978-3-642-33439-9_6 © Springer-Verlag Berlin Heidelberg 2013
Chapter 6
A Novel Choquet Integral Composition
Forecasting Model for Time Series Data Based
on Completed Extensional L-Measure
Hsiang-Chuan Liu*
Abstract. In this study, based on the Choquet integral with respect to complete
extensional L-measure and M-density, a novel composition forecasting model
which composed the time series model , the exponential smoothing model and
GM(1,1) forecasting model was proposed. For evaluating this improved
composition forecasting model, an experiment with the data of the grain
production in Jilin during 1952 to 2007 by using the sequential mean square error
was conducted. Based on the M-density and N- density, the performances of
Choquet integral composition forecasting model with the completed extensional
L-measure, extensional L-measure, L-measure, Lambda-measure and P-measure,
respectively, a ridge regression composition forecasting model and a multiple
linear regression composition forecasting model and the traditional linear
weighted composition forecasting model were compared. The experimental results
showed that the Choquet integral composition forecasting model with respect to
the completed extensional L-measure and M-density outperforms other ones.
Furthermore, for each fuzzy measure, including the completed extensional
L-measure, extensional L-measure, L-measure, Lambda-measure and P-measure,
respectively, the Choquet integral composition forecasting model based on
M-density is better than the one based on N-density.
Keywords: Choquet integral, composition forecasting model, M-density,
completed extensional L-measure.
Hsiang-Chuan Liu
Department of Biomedical Informatics, Asia University,
500, Lioufeng Rd. Wufeng, Taichung 41354, Taiwan, R.O.C.
Graduate Institute of Educational Measurement and Statistics,
National Taichung University of Education
140 Min-Shen Road, Taichung 40306 Taiwan R.O.C
e-mail: lhc@asia.edu.tw

120 H.-C. Liu
1 Introduction
The composition forecasting model is first considered in the work of Bates and
Granger (1969) [1]. They are now in a widespread use in many areas, especially in
economic field. Zhang Wang and Gao (2008) [2] applied the linear composition
forecasting model which composed the time series model, the second-order
exponential smoothing model and GM(1,1) forecasting model in the Agricultural
Economy Research, the GM(1,1) is one of the most frequently used grey
forecasting model, it is a time series forecasting model, encompassing a group of
differential equations adapted for parameter variance, rather than a first order
differential equation [3-4]. In our previous works [5-9], we extended the work of
Zhang, Wang, and Gao by proposing some nonlinear composition forecasting
model which also composed the time series model, the second-order exponential
smoothing model and GM(1,1) forecasting model by using the ridge regression
model [5] and the theory of Choquet integral with respect to some fuzzy measures,
including Sugeno’s λ-measure [13], Zadeh’s P-measure [14] and authors’ fuzzy
measures, L-measure, extensional L-measure and completed extensional L-
measure [6-12]. Since the first two well-known fuzzy measures are univalent
measures, each of them has just one feasible fuzzy measure satisfying the
conditions of its own definition, but the others proposed by our previous works are
multivalent fuzzy measures, all of them have infinitely feasible fuzzy measures
satisfying the conditions of their own definition. The fuzzy measure based
Choquet integral composition forecasting models are supervised methods, by
comparing the mean square errors between the estimated values and the
corresponding true values, each of our multivalent fuzzy measures based
forecasting models has more opportunity to find the better feasible fuzzy measure,
the performances of them are always better than the one based on the univalent
fuzzy measures, λ-measure and P-measure. In addition, the author has proved that
the P-measure is a special case of the L-measure [7], we know that all of the
extended multivalent fuzzy measures of L-measure are at lest as good as their
special case P-measure. However, the λ-measure is not a special case of the L-
measure, so the improved L-measure, called extensional L-measure, was proposed
to contain the λ-measure as a special case [7]. And then, all of the P-measure, λ-
measure and L-measure are special cases of the extensional L-measure. However,
the extensional L-measure does not attend the largest fuzzy measure B-measure, it
is not a completed fuzzy measure, for overcoming this drawback, an improved
extensional L-measure, called completed extensional L-measure was proposed, all
of other above-mentioned fuzzy measures proposed are the special cases of it. The
real data experiment showed that the extensional L-measure Choquet integral
based composition forecasting model is the best one. On the other hand, all of
above mentioned Choquet integral composition forecasting models with some
different fuzzy measures are based on N-density. From the definition of Choquet
integral and fuzzy measures, we know that the Choquet integral can be viewed as
a function of its fuzzy measure, and the fuzzy measure can be viewed as a function
of its fuzzy density function, therefore, the performance of any Choquet integral
is predominate by its fuzzy measure, and the performance of any fuzzy measure is

6 A Novel Choquet Integral Composition Forecasting Model 121
predominate by its fuzzy density function, in other words, the performance of any
Choquet integral is predominate by its fuzzy density function. Since the older
fuzzy density function N-density is based on the linear correlation coefficient, the
new fuzzy density function M-density based on the mean square error is non-
linear, the relations among the composition forecasting model and three given
forecasting models are non-linear as well, hence, in the same Choquet integral
with respect to the same fuzzy measure, the performance of the non-linear fuzzy
density functions is always better than the linear fuzzy density functions.
In this paper, a novel fuzzy measure, called the completed extensional L-
measure, and the new fuzzy density function, M-density, are considered. Based on
the M-density and the proposed completed extensional L-measure, a novel
composition forecasting model is also considered. For comparing the forecasting
efficiency of two fuzzy densities M-density and N-density, is also considered.
2 The Composition Forecasting Model
In this paper, for evaluating the forecasting validation of forecasting model to
sequential data, the sequential mean square error is used, its formal definition is
listed as follows.
Definition 1. Sequential Mean Square Error (SMSE) [9-10]
If t jθ + is the realized value of target variable at time ( )t j+ , |
ˆ
t j tθ + is the forecasted
value of target variable at time ( )t j+ based on training data set from time 1 to
time t,
and ( )
( ) ( )
2
| 1
1
1ˆ ˆ ˆ
h
h
t t j t j t j
j
SMSE
h
θ θ θ+ + − +
=
= − (1)
then ( )
( )ˆ h
tSMSE θ is called the sequential mean square error (SMSE) of the h
forecasted values of target variable from time ( )1t + to time ( )t h+ based on
training data set from time 1 to time t. The composition forecasting model or
combination forecasting model can be defined as follows.
Definition 2. Composition Forecasting Model [9-10]
(i) Let ty be the realized value of target variable at time t.
(ii) Let ,1 ,2 ,, ,...,t t t mx x x be a set of m competing predictors of ty , ˆty be a function f
of ,1 ,2 ,, ,...,t t t mx x x with some parameters, denoted as
( ),1 ,2 ,
ˆ , ,...,t t t t my f x x x= (2)

122 H.-C. Liu
(iii) Let | ,t j t kx + be the forecasted values of ty by competing predictor k at time
( )t j+ based on training data set from time 1 to time t, and for the same function f
as above,
Let ( )| ,1 ,2 ,
ˆ , ,...,t j t t j t j t j my f x x x+ + + += (3)
(iv) Let ( )
( ) ( )
2
| 1
1
1
ˆ ˆ
h
h
t t j t j t j
j
SMSE y y y
h
+ + − +
=
= − (4)
( )
( ) ( )
2
, ,
1
1 h
h
t k t j k t j
j
SMSE x x y
h
+ +
=
= − (5)
For current time t and the future h times, if
( )
( ) ( )
( ),
1
ˆ min
h h
t t k
k m
SMSE y SMSE x
≤ ≤
≤ (6)
then ˆty is called a composition forecasting model for the future h times of
,1 ,2 ,, ,...,t t t mx x x or, in brief, a composition forecasting model of ,1 ,2 ,, ,...,t t t mx x x .
Definition 3. Linear Combination Forecasting Model [9-10]
For given parameters
1
, 1
m
k k
k
Rβ β
=
∈ = , let
,
1
ˆ
m
t k t k
k
y xβ
=
=  (7)
If ˆty is a composite forecasting model of ,1 ,2 ,, ,...,t t t mx x x then ˆty is called a linear
combination forecasting model or linear composition forecasting model,
otherwise, it is called a non-linear combination forecasting model or non-linear
composition forecasting model.
Definition 4. Ridge Regression Composition Forecasting Model [5,9,10]
(i) Let ( )1 2, ,...,
T
tt
y y y y= be realized data vector of target variable from time 1 to
time t, ( )1, 2, ,, , ,...,
T
k k t kt kx x x x= be a forecasted value vector of competing
predictor k of target variable ty from time 1 to time t.
(ii) Let tX be a forecasted value matrix of m competing predictors of target
variable ty from time 1 to time t.
(iii) Let ( )1 2
ˆ ˆ ˆ ˆ, ,...,
T
tt
y y y y= (8)

( ) ( ),1 ,2 ,, ,...,t t t t mf X f x x x= (9)
(iv) Let ( ) ( ) ( ) ( )
( ) ( )
1
,1 ,2 ,, ,...,
T
r r r r T T
t t t m t t m tt t
X X rI X yβ β β β
−
= = + (10)
( ) ( )
ˆ r
t t tt
y f X X β= = (11)
Then ( ) ( )
|
ˆ r
t j t j tt j t
y f X X β+ ++
= = (12)
( )
( ) ( )
| ,1 ,2 ,
,1 ,2 , , ,
1
ˆ , ,...,
, ,...,
t j t t j t j t j m
m
r r
t j t j t j m t k t j kt
k
y f x x x
x x x xβ β
+ + + +
+ + + +
=
=
 = =  
(13)
For current time t and the future h times, if
( )
( ) ( )
( ),
1
ˆ min
h h
t t k
k m
SMSE y SMSE x
≤ ≤
≤ (14)
And ridge coefficient 0r = then ˆty is called a multiple linear regression
combination forecasting model of ,1 ,2 ,, ,...,t t t mx x x . If formula (14) is satisfied
and 0r > , then ˆty is called a ridge regression composition forecasting model
of ,1 ,2 ,, ,...,t t t mx x x . Note that Hoerl, Kenard, and Baldwin (1975) suggested that the
ridge coefficient of ridge regression is
( )
2
22
1
ˆ 1
ˆ ˆ,
t
i tT
i
t
m
r y y
t
σ
σ
β β =
= = − (15)
3 Choquet Integral Composition Forecasting Model
3.1 Fuzzy Measures [6-13]
Definition 5. Fuzzy Measure [6-13]
A fuzzy measure μ on a finite set X is a set function [ ]: 2 0,1X
μ → satisfying the
following axioms:
( ) ( )0, 1Xμ φ μ= = (boundary conditions) (16)
( ) ( )A B A Bμ μ⊆  ≤ (monotonicity) (17)

124 H.-C. Liu
3.2 Fuzzy Density Function [6-10]
Definition 6. Fuzzy Density Function, Density [6-10]
(i) A fuzzy density function of a fuzzy measure μ on a finite set X is a function
[ ]: 0,1d X → satisfying:
( ) { }( ),d x x x Xμ= ∈ (18)
( )d x is called the density of singleton x .
(ii) A fuzzy density function is called a normalized fuzzy density function or a
density if it satisfying
( ) 1
x X
d x
∈
= (19)
Definition 7. Standard Fuzzy Measure [6-10]
A fuzzy measure is called a standard fuzzy measure, if its fuzzy density function is
a normalized fuzzy density function.
Definition 8. N-density [8-10]
Let μ be a fuzzy measure on a finite set { }1 2, ,..., nX x x x= , iy be global response of
subject i and ( )i jf x be the evaluation of subject i for singleton jx , satisfying:
( )0 1, 1,2,..., , 1,2,...,i jf x i N j n< < = = (20)
If ( )
( )( )
( )( )1
, 1,2,...,
j
N j n
j
j
r f x
d x j n
r f x
=
= =

(21)
Where ( )( )jr f x is the linear regression coefficient of iy on ( )jf x satisfying
( )( ) ,
0
j
j
y x
j
y x
S
r f x
S S
= ≥ (22)
2
2
1 1
1 1
N N
y i i
i i
S y y
N N= =
 
= − 
 
 
  (23)
( ) ( )
2
2
1 1
1 1
j
N N
x i j i j
i i
S f x f x
N N= =
 
= − 
  
  (24)

( ) ( ),
1 1 1
1 1 1
j
N N N
y x i i i j i j
i i i
S y y f x f x
N N N= = =
  
= − −  
    
   (25)
then the function [ ]: 0,1Nd X → satisfying { }( ) ( ),Nx d x x Xμ = ∀ ∈ is a fuzzy
density function, called N-density of μ .
Note that
(i) N-density is a normalized fuzzy density function.
(ii) N-density is a linear fuzzy density function based on linear correlation
coefficients
3.3 M-Density [10]
We know that any linear function can be viewed as a special case of some
corresponding non-linear function, In this paper, a non-linear fuzzy density
function based on Mean Square Error, denoted M-density, is proposed, its formal
definition is introduced as follows:
Definition 9. M-density
Let μ be a fuzzy measure on a finite set { }1 2, ,..., nX x x x= , iy be global response of
subject i and ( )i jf x be the evaluation of subject i for singleton jx , satisfying:
( )0 1, 1,2,..., , 1,2,...,i jf x i N j n< < = = (26)
If ( )
( )
( )
1
1
1
, 1,2,...,
j
M j n
j
j
MSE x
d x j n
MSE x
−
−
=
 
 = =
 
 
(27)
Where ( ) ( )( )
2
1
1 N
j i i j
i
MSE x y f x
N =
= − (28)
then the function [ ]: 0,1Md X → satisfying { }( ) ( ),Mx d x x Xμ = ∀ ∈ is a fuzzy
density function, and called M-density of μ .
3.4 Classification of Fuzzy Measures [6-10]
Definition 10. Additive measure, sub-additive measure and supper- additive
measure
(i) A fuzzy measure μ is called an sub-additive measure, if
( ) ( ) ( ), ,A B X A B g A B g A g Bμ μ μφ∀ ⊂ =  < +  (29)

126 H.-C. Liu
(ii) A fuzzy measure μ is called an additive measure, if
( ) ( ) ( ), ,A B X A B g A B g A g Bμ μ μφ∀ ⊂ =  = +  (30)
(iii) A fuzzy measure μ is called a supper-additive measure, if
( ) ( ) ( ), ,A B X A B g A B g A g Bμ μ μφ∀ ⊂ =  > +  (31)
(iv) A fuzzy measure is called a mixed fuzzy measure, if is not a Additive
measure, sub-additive measure and supper- additive measure.
Theorem 1. Let d be a given fuzzy density function of an additive measure, A-
measure, then its measure function [ ]: 2 0,1X
Ag → satisfies
( ) ( )A
x E
E X g E d x
∈
∀ ⊂  =  (32)
3.4 λ-Measure [13]
Definition 10. λ-measure [13]
For a given fuzzy density function d on a finite set X, X n= , a measure is called
λ-measure, if its measure function, [ ]: 2 0,1X
gλ → , satisfying:
(i) ( ) ( )0, 1g g Xλ λφ = = (33)
(ii)
( ) ( ) ( ) ( ) ( )
, 2 , ,X
A B A B A B X
g A B g A g B g A g Bλ λ λ λ λ
φ
λ
∈ = ≠
 = + +
 

(34)
(iii) ( ) ( ) { }( )
1
1 1 0,
n
i i i
i
d x d x g xλλ λ
=
 + = + > = ∏ (35)
Theorem 2. Let d be a given fuzzy density function on a finite set X, X n= ,
Under the condition ofλ-measure, the equation (35) determines the parameter λ
uniquely:
(i) ( ) 1 0
x X
d x λ
∈
>  < ,λ-measure is a sub-additive measure (36)
(ii) ( ) 1 0
x X
d x λ
∈
=  = , λ-measure is an additive measure (37)
(iii) ( ) 1 0
x X
d x λ
∈
<  > , λ-measure is a supper-additive measure (38)

Note that
(i) λ-measure has just one feasible fuzzy measure satisfies the conditions of
its own definition.
(ii) In equation (35), the value of ( )id x is decided first, and then to find the
solution of the measure parameterλ, and ( )
1
1
n
i
i
d xλ
=
 + ∏ can be viewed
as a function of its fuzzy density ( )id x . Therefore, we can say that λ-
measure is predominate by its fuzzy density function.
(iii) λ-measure can not be a mixed fuzzy measure.
3.5 P-Measure [14]
Definition 11. P-measure [14]
P-measure, if its measure function, [ ]: 2 0,1X
Pg → , satisfying:
(i) ( ) ( )0, 1P Pg g Xφ = = (39)
(ii) ( ) ( ) { }( )2 max maxX
P P
x A x A
A g A d x g x∀
∈ ∈
∈  = = (40)
Theorem 3. P-measure is always a sub-additive measure [6-10]
Note that since the maximum of any finite set is unique, hence, P-measure has just
one feasible fuzzy measure satisfies the conditions of its own definition.
3.6 Multivalent Fuzzy Measure [6-10]
Definition 12. Univalent fuzzy measure, multivalent fuzzy measure [4-8]
A fuzzy measure is called a univalent fuzzy measure, if it has just one feasible
fuzzy measure satisfies the conditions of its own definition, otherwise, it is called
a multivalent fuzzy measure.
Note that both λ-measure and P-measure are univalent fuzzy measures.
3.7 L-Measure [6-10]
In my previous work [4], a multivalent fuzzy measure was proposed, which is
called L-measure, since my last name is Liu. Its formal definition is as follows
Definition 13. L-measure [6-10]
L-measure, if its measure function, [ ]: 2 0,1X
Lg → , satisfying:

128 H.-C. Liu
(i) ( ) ( )0, 1L Lg g Xφ = = (41)
(ii) [ ) ( )
( ) ( ) ( )
( ) ( )
1 1 max
0, , max
1
x A
x A
L
x A
x X
A L d x d x
L X A X g d x
n A L A d x
∈
∈
∈
∈
 − −
 
∈ ∞ ≠ ⊂  = +
 − + − 


(42)
Theorem 4. Important Properties of L-measure [6]
(i) For any [ )L 0,∈ ∞ , L-measure is a multivalent fuzzy measure, in other words,
L-measure has infinite fuzzy measure solutions.
(ii) L-measure is an increasing function on L.
(iii) If L 0= then L-measure is just the P-measure.
(iv) L-measure may be a mixed fuzzy measure
Note that
(i) P-measure is a special case of L-measure
(ii) L-measure does not contain additive measure and λ-measure, in other
words, additive measure and λ-measure are not special cases of L-
measure.
3.8 Extensional L-Measure [7]
For overcoming the drawback of L-measure, an improving multivalent fuzzy
measure which containing additive measure and λ-measure., called extensional L-
measure, was proposed by my next previous paper [7], Its formal definition is as
follows;
Definition 14. Extensional L-measure, LE-measure [7]
extensional L-measure, if its measure function, [ ]: 2 0,1E
X
(i) ( ) ( )0, 1E EL Lg g Xφ = = (43)
(ii)
[ )
( )
( ) ( ) ( ) [ ]
( )
( ) ( ) ( )
( ) ( )
( )
1, ,
1 max , 1,0
1 1
, 0,
1
E
x A
x A
L
x A x A
x A
x X
L A X
L d x L d x L
g A A L d x d x
d x L
n A L A d x
∈
∈
∈ ∈
∈
∈
∈ − ∞ ⊂
 + − ∈ −


 
 = − −  
  + ∈ ∞
  − + − 

 


(44)

Theorem 5. Important Properties of LE –measure [7]
(i) For any [ )L 1,∈ − ∞ , LE -measure is a multivalent fuzzy measure, in other
words, LE-measure has infinite fuzzy measure solutions.
(ii) LE-measure is an increasing function on L.
(iii) if L 1= − then LE-measure is just the P-measure.
(iv) if L 0= then LE-measure is just the additive measure.
(v) if L 0= and ( ) 1
x X
d x
∈
= , then LE-measure is just the λ-measure.
(vi) if -1< L< 0 then LE-measure is a supper-additive measure.
(vii) if L> 0 then LE-measure is a sub-additive measure
Note that additive measure, λ-measure and P-measure are two special cases of LE-
measure.
3.9 B-Measure [7]
For considering to extend the extensional L-measure, a special fuzzy measure was
proposed by my previous work as below;
Definition 15. B-measure [7]
For a given fuzzy density function d, a B-measure, Bg , is a measure on a finite set
X, X n= , satisfying:
( )
( ) 1
1 1
x A
B
d x if A
A X g A
if A
∈
 ≤

∀ ⊂  = 
>

(45)
Theorem 6. Any B-measure is a supper-additive measure.
3.10 Comparison of Two Fuzzy Measures [7-10]
Definition 16. Comparison of two fuzzy measures [7-10]
For a given fuzzy density function, ( )d x , on a finite set, X, let 1μ and 2μ be two
fuzzy measures on X,
(i) If ( ) ( )1 2
, ,g A g A A Xμ μ= ∀ ⊂ , then we say that 1μ -measure is equal to
2μ -measure, denoted as
1 2measure masureμ μ− = − (46)
(ii) If ( ) ( )1 2
, ,1g A g A A X A Xμ μ< ∀ ⊂ < < then we say that 1μ -measure is
less than 2μ -measure, or 2μ -measure is larger than 1μ -measure, denoted as

130 H.-C. Liu
1 2measure masureμ μ− < − (47)
(iii) If ( ) ( )1 2
, ,1g A g A A X A Xμ μ≤ ∀ ⊂ < < , then we say that 1μ -measure
is not larger than 2μ -measure, or 2μ -measure is not smaller than 1μ -
measure, denoted as
1 2measure masureμ μ− ≤ − (48)
.
Theorem 7. For any given fuzzy density function, if measureμ − is a fuzzy
measure, then we have
P measure as meaure B measureμ− ≤ − ≤ − (49)
In other words, for any given fuzzy density function, the P-measure is the smallest
fuzzy measure, and the B-measure is the largest fuzzy measure.
3.11 Completed Fuzzy Measure
Definition 17. Completed fuzzy measure [8]
If the measure function of a multivalent fuzzy measure has continuously infinite
fuzzy measure solutions, and both P-measure and B -measure are its limit fuzzy
measure solutions, then this multivalent fuzzy measure is called a completed fuzzy
measure.
Note that both the L –measure and LE –measure are not completed fuzzy
measures, since
( )
( ) ( )
( )
( )
( 1)
lim 1
1
x A x A
L
x Xx X
A L d x d x
d xn A A L d x
∈ ∈
→∞
∈∈
−
= ≠
 − + − 
 

, the B-measure is not a limit fuzzy
measure of the L –measure and LE –measure
3.12 Completed Extensional L-Measure
Definition 18. Completed extensional L-measue, LCE –measure
extensional L-measure, if its measure function, [ ]: 2 0,1CE
X
(i) ( ) ( )0, 1CE CEL Lg g Xφ = = (50)

(ii)
[ )
( )
( ) ( ) ( ) [ ]
( )
( ) ( ) ( )
( ) ( ) ( )
( )
1, ,
1 max , 1,0
1 1
, 0,
1
CE
x A
x A
L
x A x A
x A
x X x A
L A X
L d x L d x L
g A A L d x d x
d x L
n A d x L A d x
∈
∈
∈ ∈
∈
∈ ∈
∈ − ∞ ⊂
 + − ∈ −


  
 = − −  
  + ∈ ∞
  − + − 

 

  (51)
Theorem 7. Important Properties of LCE –measure [7]
(i) For any [ )L 1,∈ − ∞ , LCE -measure is a multivalent fuzzy measure, in other
words, LCE -measure has infinite fuzzy measure solutions.
(ii) LCE -measure is an increasing function on L.
(iii) if L 1= − then LCE -measure is just the P-measure.
(iv) if L 0= then LCE -measure is just the additive measure.
(v) if L 0= and ( ) 1
x X
d x
∈
= , then LCE -measure is just the λ-measure.
(vi) if -1< L< 0 then LCE -measure is a sub-additive measure.
(vii) if L> 0 then LCE -measure is a supper-additive measure
(viii) L → ∞ then LCE -measure is a B- measure
(ix) LCE -measure is a completed fuzzy measure.
Note that additive measure, λ-measure, P-measure and B-measure are special
cases of LCE -measure.
3.13 Choquet Integral
Definition 19. Choquet Integral [9-10]
Let μ be a fuzzy measure on a finite set { }1 2, ,..., mX x x x= . The Choquet integral
of :if X R+→ with respect to μ for individual i is denoted by
( )( ) ( )( ) ( )( )1
1
, 1,2,...,
m
i
C i i ij j j
j
f d f x f x A i Nμ μ−
=
 = − =   (52)
where ( )( )0
0if x = , ( )( )i j
f x indicates that the indices have been permuted so that
( )( ) ( )( ) ( )( )1 2
0 ...i i i m
f x f x f x≤ ≤ ≤ ≤ , ( ) ( ) ( ) ( ){ }1
, ,...,j j j m
A x x x+
= (53)
Note that from Definition 19, for given integrand :if X R+→ , the Choquet
integral can be viewed as a function of the fuzzy measure μ -measure, in other
words, the value of Choquet integral is predominate by its fuzzy measure.

132 H.-C. Liu
Theorem 8. If a λ-measure is a standard fuzzy measure on { }1 2, ,..., mX x x x= ,
and [ ]: 0,1d X → is its fuzzy density function, then the Choquet integral of
:if X R+→ with respect to λ for individual i satisfying
( ) ( )
1
, 1,2,...,
m
C i j i j
j
f d d x f x i Nλ
=
= = (54)
3.14 Choquet Integral Composition Forecasting Model
Definition 20. Choquet Integral Composition Forecasting Model [8]
(i) Let ty be the realized value of target variable at time t,
(ii) Let { }1 2, ,..., mX x x x= be the set of m competing predictors,
(iii) Let :tf X R+→ , ( ) ( ) ( )1 2, ,...,t t t mf x f x f x be m forecasting values of ty by
competing predictors 1 2, ,..., mx x x at time t.
If μ is a fuzzy measure on X , , Rα β ∈ satisfying
( ) ( ),
1
ˆˆ, arg min
N
i C t
t
y f dgμ
α β
α β α β
=
 
= − − 
 
  (55)
1 1
1 1ˆˆ
N N
t t
t t
y f dg
N N
μα β
= =
= −  , ˆ yf
ff
S
S
β = (56)
1 1
1 1ˆˆ
N N
t t
t t
y f dg
N N
μα β
= =
= −  (57)
1 1 1
1 1
1
N N N
i t t t
t t t
yf
y y f dg f dg
N N
S
N
μ μ
= = =
  
− −  
    =
−
    (58)
then ˆˆˆ , 1,2,...,t ty f dg t Nμα β= + = is called the Choquet integral regression
composition forecasting estimator of ty , and this model is also called the Choquet
integral regression composition forecasting model with respect to μ -measure.
Theorem 9. If a λ-measure is a standard fuzzy measure then Choquet integral
regression composition forecasting model with respect to λ-measure is just a linear
combination forecasting model.

4 Experiments and Results
A real data of the grain production with 3 kinds of forecasted values of the time
series model, the exponential smoothing model and GM(1,1) forecasting model,
respectively, in Jilin during 1952 to 2007 from the paper of Zhang, Wang and Gao
[2],was listed in Table 2. For evaluating the proposed new density based
composition forecasting model, an experiment with the above-mentioned data by
using sequential mean square error was conducted.
We arrange the first 50 years grain production and their 3 kinds of forecasted
values as the training set and the rest data as the forecasting set. And the following
N-density and M-density of all fuzzy measures were used
N-density: { }0.3331, 0.3343, 0.3326 (59)
M-density: { }0.2770, 0.3813, 0.3417 (60)
The performances of Choquet integral composition forecasting model with
extensional L-measure, L-measure, λ-measure and P-measure, respectively, a
ridge regression composition forecasting model and a multiple linear regression
composition forecasting model and the traditional linear weighted composition
forecasting model were compared. The result is listed in Table 1.
Table 1 SMSEs of 2 densities for 7 composition forecasting models
Composition forecasting Models
SMSE
N-density M-density
Choquet integral regression
LCE-measure 13149.64 13217.31
LE-measure 13939.84 13398.29
L-measure 14147.83 13751.60
λ-measure 21576.38 19831.86
P-measure 16734.88 16465.98
Ridge regression 18041.92
Multiple linear regression 24438.29
Table 1 shows that the M-density based Choquet integral composition
forecasting model with respect to LCE-measure outperforms other composition
forecasting models. Furthermore, for each fuzzy measure, including the LCE-
measure, LE-measure, L-measure, λ-measure and P-measure, the M-density based
Choquet integral composition forecasting model is better than the N-density based.
5 Conclusion
In this paper, a new density, M-density, was proposed. Based on M-density, a
novel composition forecasting model was also proposed. For comparing the

134 H.-C. Liu
forecasting efficiency of this new density with the well-known density, N-density,
a real data experiment was conducted. The performances of Choquet integral
composition forecasting model with the completed extensional L-measure,
extensional L-measure, λ-measure and P-measure, by using M-density and N-
density, respectively, a ridge regression composition forecasting model and a
multiple linear regression composition forecasting model and the traditional linear
weighted composition forecasting model were compared. Experimental result
showed that for each fuzzy measure, including the LCE-measure, LE-measure, L-
measure, λ-measure and P-measure, the M-density based Choquet integral
composition forecasting model is better than the N-density based, and the M-
density based Choquet integral composition forecasting model outperforms all of
other composition forecasting models.
Acknowledgment. This study is partially supported by the grant of National Science
Council of Taiwan Government (NSC 100-2511-S-468-001).
References
1. Bates, J.M., Granger, C.W.J.: The Combination of Forecasts. Operations Research
Quarterly 4, 451–468 (1969)
2. Zhang, H.-Q., Wang, B., Gao, L.-B.: Application of Composition Forecasting Model in
the Agricultural Economy Research. Journal of Anhui Agri. Sci. 36(22), 9779–9782
(2008)
3. Hsu, C.-C., Chen, C.-Y.: Applications of improved grey prediction model for power
demand forecasting. Energy Conversion and Management 44, 2241–2249 (2003)
4. Kayacan, E., Ulutas, B., Kaynak, O.: Grey system theory-based models in time series
prediction. Expert Systems with Applications 37, 1784–1789, (2010)
5. Hoerl, A.E., Kenard, R.W., Baldwin, K.F.: Ridge regression: Some simulation.
Communications in Statistics 4(2), 105–123 (1975)
6. Liu, H.-C., Tu, Y.-C., Lin, W.-C., Chen, C.C.: Choquet integral regression model
based on L-Measure and γ-Support. In: Proceedings of 2008 International Conference
on Wavelet Analysis and Pattern Recognition (2008)
7. Liu, H.-C.: Extensional L-Measure Based on any Given Fuzzy Measure and its
Application. In: Proceedings of 2009 CACS International Automatic Control
Conference, November 27-29, pp. 224–229. National Taipei University of
Technology, Taipei Taiwan (2009)
8. Liu, H.-C.: A theoretical approach to the completed L-fuzzy measure. In: Proceedings
of 2009 International Institute of Applied Statistics Studies (IIASS), 2nd Conference,
Qindao, China, July 24-29 (2009)
9. Liu, H.-C., Ou, S.-L., Cheng, Y.-T., Ou, Y.-C., Yu, Y.-K.: A Novel Composition
Forecasting Model Based on Choquet Integral with Respect to Extensional L-Measure.
In: Proceedings of the 19th National Conference on Fuzzy Theory and Its Applications
(2011)

10. Liu, H.-C., Ou, S.-L., Tsai, H.-C., Ou, Y.-C., Yu, Y.-K.: A Novel Choquet Integral
Composition Forecasting Model Based on M-Density. In: Pan, J.-S., Chen, S.-M.,
Nguyen, N.T. (eds.) ACIIDS 2012, Part I. LNCS, vol. 7196, pp. 167–176. Springer,
Heidelberg (2012)
11. Choquet, G.: Theory of capacities. Annales de l’Institut Fourier 5, 131–295 (1953)
12. Wang, Z., Klir, G.J.: Fuzzy Measure Theory. Plenum Press, New York (1992)
13. Sugeno, M.: Theory of fuzzy integrals and its applications. Unpublished doctoral
dissertation, Tokyo Institute of Technology, Tokyo, Japan (1974)
14. Zadeh, L.A.: Fuzzy Sets as a Basis for Theory of Possibility. Fuzzy Sets and
Systems 1, 3–28 (1978)
Appendix
Table 2 SMSEs of 2 densities for 6 composition forecasting models
Years Y X1 X2 X3 X4
1952 613.20 490.67 518.60 399.51 472.45
1953 561.45 549.73 570.84 414.09 511.35
1954 530.95 542.83 586.41 429.20 524.94
1955 556.53 549.57 584.31 444.86 530.10
1956 493.64 582.69 591.12 461.09 542.51
1957 429.35 598.64 570.80 477.91 538.81
1958 528.84 610.69 531.14 495.35 524.37
1959 526.60 633.88 540.11 513.43 537.85
1960 394.70 655.07 544.78 532.16 549.04
1961 398.55 672.97 497.45 551.58 531.58
1962 437.16 694.53 465.26 571.71 523.02
1963 501.67 617.26 457.04 592.57 519.94
1964 491.80 738.99 475.53 614.19 547.93
1965 525.10 761.94 484.57 636.61 563.02
1966 597.60 786.18 503.23 659.84 583.82
1967 647.74 810.67 543.38 683.91 616.78
1968 622.15 835.87 589.95 708.87 653.65
1969 498.70 862.13 612.17 734.74 677.54
1970 738.80 889.12 580.22 761.55 672.02
1971 713.05 916.86 647.93 789.33 721.78

136 H.-C. Liu
Table 2 (continued)
1972 556.99 945.56 684.75 818.14 754.99
1973 783.00 975.15 650.45 847.99 749.50
1974 858.15 1005.63 711.01 878.93 796.71
1975 906.50 1037.08 780.79 911.01 849.50
1976 755.50 1069.53 846.59 944.25 900.60
1977 728.35 1102.98 833.96 978.70 909.05
1978 914.70 1137.47 813.29 1014.40 913.61
1979 903.34 1173.05 867.92 1051.40 960.20
1980 859.60 1209.74 900.44 1089.80 995.23
1981 921.91 1247.58 905.13 1129.60 1015.54
1982 1000.04 1286.60 930.45 1170.80 1047.82
1983 1477.98 1326.84 976.21 1213.50 1092.01
1984 1634.46 1368.34 1187.28 1257.80 1227.91
1985 1225.26 1411.14 1391.77 1303.70 1360.87
1986 1397.71 1455.27 1376.80 1351.30 1373.74
1987 1675.81 1500.79 1428.01 1400.60 1423.79
1988 1693.25 1547.73 1565.57 1451.70 1522.16
1989 1351.29 1596.14 1664.67 1504.60 1600.13
1990 2046.52 1646.06 1600.61 1559.50 1589.13
1991 1898.87 1697.54 1814.80 1616.50 1732.23
1992 1840.31 1750.64 1904.70 1675.40 1807.73
1993 1900.90 1805.39 1940.83 1736.60 1854.61
1994 2015.70 1861.86 1984.40 1799.90 1906.49
1995 1992.40 1920.09 2053.78 1865.60 1973.62
1996 2326.60 1980.15 2089.00 1933.70 2022.96
1997 1808.30 2042.08 2235.53 2004.30 2134.67
1998 2506.00 2105.95 2137.39 2077.40 2112.74
1999 2305.60 2171.82 2328.13 2153.20 2251.02
2000 1638.00 2239.75 2381.33 2231.80 2314.77
2001 1953.40 2309.80 2161.43 2313.20 2229.36
2002 2214.80 2382.04 2123.07 2397.60 2245.19

Table 2 (continued)
2003 2259.60 2456.54 2192.19 2485.10 2321.52
2004 2510.00 2533.38 2254.69 2575.80 2395.57
2005 2581.21 2612.61 2390.11 2669.80 2511.18
2006 2720.00 2694.33 2508.40 2767.20 2618.82
2007 2454.00 2778.60 2560.38 2831.50 2677.95
Y: realized value of target variable
X1: Fitting value of time series model
X2: Fitting value of exponential smoothing model
X3: Fitting value of GM(1,1) model
X4: Fitting value of composition forecasting model

Time series analysis, modeling and applications

More Related Content

What's hot (20)

Viewers also liked (8)

Similar to Time series analysis, modeling and applications (20)

More from Springer (20)

Recently uploaded (20)

Time series analysis, modeling and applications