You are on page 1of 16

Fuzzy Optimization and Decision Making

https://doi.org/10.1007/s10700-018-9298-z

Uncertain time series analysis with imprecise observations

Xiangfeng Yang1 · Baoding Liu2

© Springer Science+Business Media, LLC, part of Springer Nature 2018

Abstract
Time series analysis is a method to predict future values based on previously observed
values. Assuming the observed values are imprecise and described by uncertain vari-
ables, this paper proposes an approach of uncertain time series. By employing the
principle of least squares, a minimization problem is derived to calculate the unknown
parameters in the uncertain time series model. In addition, residual and confidence
interval are also proposed. Finally, some numerical examples are given.

Keywords Time series analysis · Uncertainty theory · Principle of least square ·


Residual analysis · Confidence interval

1 Introduction

Time series is a sequence of observed values arranged in chronological order, and the
main goal of time series analysis is to predict future value from the knowledge of the
time series at its current time and also previous time. That is, via time series analysis,
the inner relationships reflected from observed values will be obtained, and the future
tendency can be forecasted. The early time series model can be traced back to the
work of Yule (1927), in which he proposed an autoregressive model to represent time
series. Later on, Walker (1931) developed a moving-average model. But time series
analysis still does not attract the attention of researchers until Box and Jenkins (1970)
published a book about time series analysis, in which they developed a systematic
approach that enables practitioners to use time series analysis in forecasting.

B Baoding Liu
liu@tsinghua.edu.cn
Xiangfeng Yang
yangxf@uibe.edu.cn

1 School of Information Technology and Management, University of International Business and


Economics, Beijing 100029, China
2 Department of Mathematical Sciences, Tsinghua University, Beijing 100084, China

123
X. Yang, B. Liu

Traditional time series analysis is a method to predict future values based on


previously observed values which are precisely observed. But in many cases, the
observations are imprecise, such as carbon emissions, social benefit and oil reserves.
When those imprecise observed values are described as fuzzy sets, Song and Chissom
(1993a, b, 1994) first proposed fuzzy time series model based on fuzzy set theory
and applied it to the forecast of university enrollments. Since then, fuzzy time series
analysis has drawn much attention to the researchers, and was further investigated
by many scholars, such as Sullivan and Woodall (1994), Hwang (2001), Tseng et al.
(2001), Chen (2002), Lee et al. (2006), Domańska and Wojtylak (2012), Egrioglu et al.
(2013), Lee and Chou (2014) and Cai et al. (2015).
However, if imprecisely observed value, like carbon emission, is assumed to be a
fuzzy set, say a trapezoidal one (390, 400, 420, 430) in ppm, then from fuzzy set theory
we can conclude the following three propositions, (i) the carbon emission is “exactly
410 ppm” with possibility measure 1, (ii) the carbon emission is “not 410 ppm” with
possibility measure 1, (iii) “exactly 410 ppm” is as possible as “not 410 ppm”. The
first proposition says we are 100% sure that the carbon emission is “exactly 410 ppm”,
neither less nor more. What a coincidence it should be! It is doubtless that the belief
degree of “exactly 410 ppm” is almost zero, and nobody is so naive to expect that
“exactly 410 ppm” is the true carbon emission. The second proposition sounds good.
The third proposition says “exactly 410 ppm” and “not 410 ppm” have same possibility
measure. Thus we have to regard them “equally likely”. Consider a bet: you get $100
if the carbon emission is “exactly 410 ppm”, and pay $100 if the carbon emission is
“not 410 ppm”. Do you think the bet is fair? It seems that no one thinks so. Hence
the conclusion (iii) is unacceptable because “exactly 410 ppm” is almost impossible
compared with “not 410 ppm”. This paradox shows that those imprecise observed
values like carbon emission cannot be quantified by a fuzzy set. In fact, a lot surveys
showed that none of data collectors think their observations follow the laws of fuzzy
set theory.
For this reason, Liu (2007) declared that uncertain variable is a better suitable tool
to describe imprecise observed value. From then on, many scholars used uncertain
variables to model imprecise observations in different fields, such as Wen et al. (2017),
Lio and Liu (2018a, b), Yao (2018), Yao and Liu (2018), and Nejad and Ghaffari-
Hadigheh (2018). In this paper, we propose an uncertain time series model to predict
future values based on previously imprecise observed values which are described by
uncertain variables.
The rest of this paper is organized as follows. Section 2 reviews some basic knowl-
edge about uncertainty theory. Section 3 introduces an uncertain time series model,
and gives a minimization problem to calculate the unknown parameters via the prin-
ciple of least squares. Section 4 considers the residual analysis of disturbance term.
Section 5 develops a confidence interval to forecast the next value. Section 6 gives
some examples, and Sect. 7 makes some conclusions.

123
Uncertain time series analysis with imprecise observations

2 Preliminaries

This section introduces some fundamental concepts and properties in uncertainty the-
ory including uncertain measure, uncertain variable and expected value.

Definition 1 (Liu 2007) Let L be a σ -algebra on a nonempty set Γ . A set function


M : L → [0, 1] is called an uncertain measure if it satisfies the following axioms,
Axiom 1 M{Γ } = 1 for the universal set Γ ;
Axiom 2 M{Λ} + M{Λc } = 1 for any event Λ;
Axiom 3 For every countable sequence of events Λ1 , Λ2 , . . ., we have
∞  ∞
 
M Λi ≤ M{Λi }.
i=1 i=1

In order to provide the operational law, Liu (2009) defined the product uncertain
measure on the product σ -algebre L, called product axiom.
Axiom 4 Let (Γk , Lk , Mk ) be uncertainty spaces for k = 1, 2, . . . . The product
uncertain measure M is an uncertain measure satisfying
∞  ∞
 
M Λk = Mk {Λk }
k=1 k=1

where Λk are arbitrarily chosen events from Lk for k = 1, 2, . . ., respectively.

Definition 2 (Liu 2007) An uncertain variable ξ is a function from an uncertainty space


(Γ , L, M) to the set of real numbers such that for any Borel set B of real numbers,
the set

{ξ ∈ B} = {γ ∈ Γ |ξ(γ ) ∈ B}

is an event.

In order to describe uncertain variable in practice, uncertainty distribution Φ :  →


[0, 1] of an uncertain variable ξ is defined as Φ(x) = M {ξ ≤ x}. If ξ is an uncertain
variable with regular uncertainty distribution Φ, then the inverse function Φ −1 (α) is
called the inverse uncertainty distribution of ξ .

Definition 3 (Liu 2009) The uncertain variables ξ1 , ξ2 , . . . , ξm are said to be indepen-


dent if
m 
 
m
M {ξi ∈ Bi } = M {ξi ∈ Bi }
i=1 i=1

for any Borel sets B1 , B2 , . . . , Bm of real numbers.

123
X. Yang, B. Liu

The operational law of uncertain variables was proved by Liu (2010) to calculate
the inverse uncertainty distribution of strictly monotonous function as the following
theorem.

Theorem 1 (Liu 2010) Let ξ1 , ξ2 , . . . , ξn be independent uncertain variables with


regular uncertainty distributions Φ1 , Φ2 , . . . , Φn , respectively. If the function f (x1 ,
x2 , . . . , xn ) is strictly increasing with respect to x1 , x2 , . . . , xm and strictly decreasing
with xm+1 , xm+2 , . . . , xn , then the uncertain variable

ξ = f (ξ1 , . . . , ξm , ξm+1 , . . . , ξn )

is an uncertain variable with inverse uncertainty distribution

Ψ −1 (α) = f (Φ1−1 (α), . . . , Φm−1 (α), Φm+1


−1
(1 − α), . . . , Φn−1 (1 − α)).

The expected value operator of uncertain variable, proposed by Liu (2007), is the
average value of uncertain variable in the sense of uncertain measure, and represents
the size of uncertain variable.

Definition 4 (Liu 2007) Let ξ be an uncertain variable. Then the expected value of ξ
is defined as
+∞ 0
E[ξ ] = M{ξ ≥ x}dx − M{ξ ≤ x}dx
0 −∞

provided that at least one of the two integrals is finite.

For an uncertain variable ξ with regular uncertainty distribution Φ(x), its expected
value can be obtained by
1
E[ξ ] = Φ −1 (α)dα. (1)
0

Theorem 2 (Liu 2010) Let ξ and η be independent uncertain variables with finite
expected values. Then for any real numbers a and b, we have

E[aξ + bη] = aE[ξ ] + bE[η].

The variance operator of uncertain variable, proposed by Liu (2007), provides a


degree of the spread of the distribution around its expected value.

Definition 5 (Liu 2007) Let ξ be an uncertain variable with finite expected value e.
Then the variance of ξ is defined by

V [ξ ] = E[(ξ − e)2 ].

123
Uncertain time series analysis with imprecise observations

For an uncertain variable ξ with regular uncertainty distribution Φ(x) and finite
expected value e, its variance can be obtained by
1
2
V [ξ ] = Φ −1 (α) − e dα. (2)
0

Theorem 3 (Sheng and Kar 2015) Let ξ be an uncertain variable with regular uncer-
tainty distribution Φ(x). Then
+∞
E[ξ 2 ] = (Φ −1 (α))2 dα.
−∞

3 Uncertain time series

Traditional time series analysis is a method to predict future values based on previously
observed values which are precisely observed. However, in an environment where the
previously observed values are evaluated imprecisely through domain experts’ belief
degrees, the traditional time series model cannot be applied directly since imprecise
observed values cannot be treated as precise observed values. By interpreting the
observed values as uncertain variables, traditional time series model can be naturally
extended to become uncertain time series model.
An uncertain times series is a sequence of imprecisely observed values each of
which is described by an uncertain variable. Mathematically, an uncertain time series
is represented by

X = {X 1 , X 2 , . . . , X n }

where X t are imprecisely observed values (uncertain variables) at times t, t =


1, 2, . . . , n, respectively. A basic problem of uncertain time series is to predict the
value of X n+1 based on previously observed values X 1 , X 2 , . . . , X n .
The simplest approach for modeling uncertain time series is an uncertain autore-
gressive model,


k
X t = a0 + ai X t−i + t (3)
i=1

where a0 , a1 , . . . , ak are unknown parameters, k is called the order of the autoregres-


sive model, and t is a disturbance term.
The least squares estimate of a0 , a1 , . . . , ak in the autoregressive model (3) is the
solution of the minimization problem
⎡ 2 ⎤

n 
k
min E ⎣ X t − a0 − ai X t−i ⎦. (4)
a0 ,a1 ,...,ak
t=k+1 i=1

123
X. Yang, B. Liu

Denoting the optimal solution by a0∗ , a1∗ , . . . , ak∗ , the fitted autoregressive model is
determined by


k
X t = a0∗ + ai∗ X t−i . (5)
i=1

Theorem 4 Let X 1 , X 2 , . . . , X n be imprecisely observed values characterized in term


of independent uncertain variables with regular uncertainty distributions Φ1 (x),
Φ2 (x), . . . , Φn (x), respectively. Then the least squares estimate of a0 , a1 , . . . , ak in
the autoregressive model


k
X t = a0 + ai X t−i + t
i=1

solves the minimization problem

n
 2
 1 
k
−1
min Φt−1 (α) − a0 − ai Υt−i (α, ai ) dα (6)
a0 ,a1 ,...,ak
t=k+1 0 i=1

where
 −1
−1 Φt−i (1 − α), if ai ≥ 0
Υt−i (α, ai ) = −1
Φt−i (α), if ai < 0

for i = 1, 2, . . . , k.
Proof For each index t, we obtain that the inverse uncertainty distribution of


k
X t − a0 − ai X t−i
i=1

is


k
−1
Ψt−1 (α) = Φt−1 (α) − a0 − ai Υt−i (α, ai ).
i=1

from Theorem 1. It follows from Theorem 3 that


⎡ 2 ⎤
k 1
2
E ⎣ X t − a0 − ai X t−i ⎦ = Ψt−1 (α) dα.
i=1 0

Therefore, the minimization problem (4) is equivalent to (6). This theorem is thus
proved.

123
Uncertain time series analysis with imprecise observations

4 Residual analysis

In k-order uncertain autoregressive model (3), we further assume that disturbance


terms k+1 , k+2 , . . . , n are iid uncertain variables (hereafter called iid hypothesis).
In this section, we consider the residual analysis of uncertain time series. And, in
residual analysis, a basic problem is to find the estimators ê and σ̂ 2 of the expected
value and variance of disturbance term, respectively.

Definition 6 Let X 1 , X 2 , . . . , X n be imprecisely observed values, and let the fitted


autoregressive model be


k
X t = a0∗ + ai∗ X t−i .
i=1

Then for each index t (t = k + 1, k + 2, . . . , n), the difference between the actual
observed value and the value predicted by the fitted model,


k
ˆt := X t − a0∗ − ai∗ X t−i
i=1

is called the tth residual.

Then a suitable approach to estimate the expected value of disturbance term is the
average of expected values of residuals, i.e.,

1 
n
 
ê := E ˆt ,
n−k
t=k+1

and the variance of disturbance term can be estimated by

1 
n  
σ̂ 2 := E (ˆ t − ê)2
n−k
t=k+1

where ˆt are the tth residuals, t = k + 1, k + 2, . . . , n, respectievly.

Theorem 5 Let X 1 , X 2 , . . . , X n be imprecisely observed values characterized in term


of independent uncertain variables with regular uncertainty distributions Φ1 (x),
Φ2 (x), . . . , Φn (x), and let the fitted autoregressive model be


k
X t = a0∗ + ai∗ X t−i .
i=1

123
X. Yang, B. Liu

Then the estimated expected value of disturbance terms under the iid hypothesis is

n 1
 
1  
k
−1 ∗ ∗ −1 ∗
ê = Φt (α) − a0 − ai Υt−i (α, ai ) dα (7)
n−k 0 t=k+1 i=1

and the estimated variance is

n 1
 2
1  
k
−1 ∗ ∗ −1 ∗
σ̂ =
2
Φt (α) − a0 − ai Υt−i (α, ai ) − ê dα (8)
n−k 0 t=k+1 i=1

where
 −1
−1 Φt−i (1 − α), if ai∗ ≥ 0
Υt−i (α, ai∗ ) = −1
Φt−i (α), if ai∗ < 0

for i = 1, 2, . . . , k.

Proof Since


k
ˆt = X t − a0∗ − ai∗ X t−i
i=1

and X t , X t−1 , . . . , X t−k are independent, we obtain that the inverse uncertainty dis-
tribution of ˆt is


k
−1
Ψ̂t−1 (α) = Φt−1 (α) − a0∗ − ai∗ Υt−i (α, ai∗ )
i=1

from Theorem 1. Based on equations (1) and (2), we can get


1
E[ˆ t ] = Ψ̂t−1 (α)dα.
0

and

2  1
2
E ˆt − ê = Ψ̂t−1 (α) − ê dα,
0

respectively. The theorem is thus proved.

123
Uncertain time series analysis with imprecise observations

5 Forecast value and confidence interval

This section proposes a confidence interval to predict the next value X n+1 for k-order
uncertain autoregressive model (3). From Sects. 3 and 4, we already have obtained the
fitted autoregressive model


k
X t = a0∗ + ai∗ X t−i
i=1

and the estimators ê and σ̂ 2 of disturbance term t . We further assume that disturbance
term n+1 follows normal uncertainty distribution N(ê, σ̂ ), and n+1 is independent
of X 1 , X 2 , . . . , X n .
For predicting the next value X n+1 , a forecast uncertain variable X̂ n+1 is defined
as follows,


k
X̂ n+1 := a0∗ + ai∗ X n+1−i + n+1 .
i=1

And the expected value of X̂ n+1 is called forecast value of X n+1 . From independence
of X n , X n−1 , . . . , X n+1−k and n+1 (Theorem 2), we can obtain that the forecast value
is
 
μ = E X̂ n+1
 
k
∗ ∗
= E a0 + ai X n+1−i + n+1
i=1

k
    (9)
= a0∗ + ai∗ E X n+1−i + E n+1
i=1

k
 
= a0∗ + ai∗ E X n+1−i + ê.
i=1

If uncertain observed values X 1 , X 2 , . . . , X n have regular uncertainty distributions


Φ1 (x), Φ2 (x), . . . , Φn (x), respectively, then the inverse uncertainty distribution of
X̂ n+1 is


k
−1 −1
Φ̂n+1 (α) = a0∗ + ai∗ Υn+1−i (α, ai∗ ) + Φ −1 (α)
i=1

123
X. Yang, B. Liu

from Theorem 1, where


 −1
−1
Φn+1−i (α), if ai∗ ≥ 0
Υn+1−i (α, ai∗ ) =
−1
Φn+1−i (1 − α), if ai∗ < 0

for i = 1, 2, . . . , k, and Φ −1 (α) is the inverse uncertainty distribution of N(ê, σ̂ ), i.e.,



−1 σ̂ 3 α
Φ (α) = ê + ln .
π 1−α
−1
From inverse uncertainty distribution Φ̂n+1 (α), we can derive the uncertainty distri-
bution Φ̂n+1 (x) of X̂ n+1 .
Take α (e.g., 95%) as the confidence level, and find the minimum value b such that

Φ̂n+1 (μ + b) − Φ̂n+1 (μ − b) ≥ α. (10)

Since

M{μ − b ≤ X̂ n+1 ≤ μ + b} ≥ Φ̂n+1 (μ + b) − Φ̂n+1 (μ − b) ≥ α,

we suggest that the α confidence interval of X n+1 is [μ − b, μ + b], which is often


abbreviated as

μ ± b.

6 Numerical examples

Example 1 Assume X 1 , X 2 , . . . , X 18 are imprecisely observed values that are char-


acterized in term of independent linear uncertain variables. See Table 1. Let us show
how the uncertain time series is used to forecast the next value X 19 .

In order to forecast it, we employ a 3-order uncertain autoregressive model,

X t = a0 + a1 X t−1 + a2 X t−2 + a3 X t−3 + t .

Table 1 Linear uncertainty distribution in Example 1

X1 X2 X3 X4 X5 X6

L(335, 347) L(338, 350) L(340, 354) L(343, 359) L(344, 364) L(346, 366)
X7 X8 X9 X 10 X 11 X 12
L(350, 366) L(355, 369) L(360, 372) L(362, 376) L(365, 381) L(370, 384)
X 13 X 14 X 15 X 16 X 17 X 18
L(373, 390) L(379, 391) L(380, 398) L(384, 402) L(388, 410) L(390, 415)

123
Uncertain time series analysis with imprecise observations

By solving the minimization problem,

18
 2
 1 
3
−1
min Φt−1 (α) − a0 − ai Υt−i (α, ai ) dα
a0 ,a1 ,a2 ,a3 0
t=4 i=1

where
 −1
−1
Φt−i (1 − α), if ai ≥ 0
Υt−i (α, ai ) =
−1
Φt−i (α), if ai < 0

for i = 1, 2, 3, we have the least squares estimate

(a0∗ , a1∗ , a2∗ , a3∗ ) = (43.3189, 0.1922, 0.0415, 0.6728).

Therefore, the fitted autoregressive model is

X t = 43.3189 + 0.1922X t−1 + 0.0415X t−2 + 0.6728X t−3 .

Based on the following two equations,

 
1  
18 1 3
−1
ê = Φt−1 (α) − a0∗ − ai∗ Φt−i (1 − α) dα,
15 0
t=4 i=1
 2
1  
18 1 3
−1
σ̂ 2 = Φt−1 (α) − a0∗ − ai∗ Φt−i (1 − α) − ê dα,
15 0
t=4 i=1

we get the estimators ê and σ̂ 2 of disturbance term t are

ê = 0.0000, σ̂ 2 = 88.7226,

respectively. And the forecast value of X 19 is


3
 
μ = a0∗ + ai∗ E X 19−i + ê = 401.6407.
i=1

Furthermore, take the confidence level α = 95%, and 19 is assumed to follow normal
uncertainty distribution N(ê, σ̂ ). We find that

b = 27.4935

is the minimum value such that

Φ̂19 (μ + b) − Φ̂19 (μ − b) ≥ 0.95

123
X. Yang, B. Liu

Table 2 Zigzag uncertainty distribution in Example 2

X1 X2 X3 X4 X5

Z(600, 612, 620) Z(615, 630, 650) Z(640, 666, 685) Z(700, 725, 740) Z(735, 758, 770)
X6 X7 X8 X9 X 10
Z(740, 755, 760) Z(730, 745, 755) Z(705, 713, 720) Z(680, 685, 692) Z(630, 636, 655)
X 11 X 12 X 13 X 14 X 15
Z(605, 611, 620) Z(560, 568, 590) Z(505, 520, 550) Z(500, 510, 512) Z(485, 494, 508)

where uncertainty distribution Φ̂19 can be derived from its inverse uncertainty distri-
bution


3
σ̂ 3 α
−1 −1
Φ̂19 (α) = a0∗ + ai∗ Φ19−i (α) + ln .
π 1−α
i=1

Thus, the 95% confidence interval of X 19 is

401.6407 ± 27.4935.

Example 2 Assume X 1 , X 2 , . . . , X 15 are imprecisely observed values that are charac-


terized in term of independent zigzag uncertain variables. See Table 2. Let us show
how the uncertain time series is used to forecast the next value X 16 .
In order to forecast it, we employ a 3-order uncertain autoregressive model,

X t = a0 + a1 X t−1 + a2 X t−2 + a2 X t−3 + t .

By solving the minimization problem,

15
 2
 1 
3
−1
min Φt−1 (α) − a0 − ai Υt−i (α, ai ) dα
a0 ,a1 ,a2 ,a3 0
t=4 i=1

where
 −1
−1
Φt−i (1 − α), if ai ≥ 0
Υt−i (α, ai ) =
−1
Φt−i (α), if ai < 0

for i = 1, 2, 3, we have the least squares estimate

(a0∗ , a1∗ , a2∗ , a3∗ ) = (116.2659, 1.1905, 0.0000, −0.3790).

Therefore, the fitted autoregressive model is

X t = 116.2659 + 1.1905X t−1 − 0.379X t−3 .

123
Uncertain time series analysis with imprecise observations

Based on the following two equations,

1

1 
15
−1 −1
ê = Φt−1 (α) − a0∗ − a1∗ Φt−1 (1 − α) − a3∗ Φt−3 (α) dα,
12 0
t=4
1
2
1 
15
−1 −1
σ̂ 2 = Φt−1 (α) − a0∗ − a1∗ Φt−1 (1 − α) − a3∗ Φt−3 (α) − ê dα,
12 0
t=4

we get the estimators ê and σ̂ 2 of disturbance term t are

ê = 0.2480, σ̂ 2 = 403.3562,

respectively. And the forecast value of X 16 is

μ = a0∗ + a1∗ E [X 15 ] + a3∗ E [X 13 ] + ê = 507.3598.

Furthermore, take the confidence level α = 95%, and 16 is assumed to follow normal
uncertainty distribution N(ê, σ̂ ). We find that

b = 61.6727

is the minimum value such that

Φ̂16 (μ + b) − Φ̂16 (μ − b) ≥ 0.95

where uncertainty distribution Φ̂16 can be derived from its inverse uncertainty distri-
bution

−1 −1 −1 σ̂ 3 α
Φ̂16 (α) = a0∗ + a1∗ Φ15 (α) + a3∗ Φ13 (1 − α) + ê + ln .
π 1−α

Thus, the 95% confidence interval of X 16 is

507.6092 ± 61.6727.

Example 3 Assume X 1 , X 2 , . . . , X 16 are imprecisely observed values that are charac-


terized in term of independent normal uncertain variables. See Table 3. Let us show
how the uncertain time series is used to forecast the next value X 17 .

In order to forecast it, we employ a 2-order uncertain autoregressive model,

X t = a0 + a1 X t−1 + a2 X t−2 + t .

123
X. Yang, B. Liu

Table 3 Normal uncertainty distribution in Example 3

X1 X2 X3 X4 X5 X6 X7 X8

N(42, 2) N(44, 1) N(46, 3) N(49, 2) N(52, 4) N(55, 2) N(57, 1) N(60, 3)


X9 X 10 X 11 X 12 X 13 X 14 X 15 X 16
N(62, 2) N(65, 3) N(67, 4) N(70, 5) N(72, 2) N(82, 3) N(85, 2) N(87, 2)

By solving the minimization problem,

16
 2
 1 
2
−1
min Φt−1 (α) − a0 − ai Υt−i (α, ai ) dα
a0 ,a1 ,a2 0
t=4 i=1

where
 −1
−1 Φt−i (1 − α), if ai ≥ 0
Υt−i (α, ai ) = −1
Φt−i (α), if ai < 0

for i = 1, 2, we have the least squares estimate

(a0∗ , a1∗ , a2∗ ) = (6.4084, 0.8147, 0.1382).

Therefore, the fitted autoregressive model is

X t = 6.4084 + 0.8147X t−1 + 0.1382X t−2 .

Based on the following equations,

 
1  
16 1 2
−1
ê = Φt−1 (α) − a0∗ − ai∗ Φt−i (1 − α) dα,
14 0
t=3 i=1
 2
1  
16 1 2
−1
σ̂ =
2
Φt−1 (α) − a0∗ − ai∗ Φt−i (1 − α) − ê dα,
14 0
t=3 i=1

we get the estimators ê and σ̂ 2 of disturbance term t are

ê = 0.0000, σ̂ 2 = 29.2464,

respectively. And forecast value of X 17 is


3
 
μ = a0∗ + ai∗ E X 16−i + ê = 89.0357.
i=1

123
Uncertain time series analysis with imprecise observations

Furthermore, take the confidence level α = 95%, and 17 is assumed to follow normal
uncertainty distribution N(ê, σ̂ ). We find that

b = 14.7727

is the minimum value such that

Φ̂17 (μ + b) − Φ̂17 (μ − b) ≥ 0.95

where uncertainty distribution Φ̂17 can be derived from its inverse uncertainty distri-
bution


2
σ̂ 3 α
−1 −1
Φ̂17 (α) = a0∗ + ai∗ Φ17−i (α) + ln .
π 1−α
i=1

Thus, the 95% confidence interval of X 17 is

89.0357 ± 14.7727.

7 Conclusion

This paper firstly proposed a methodology of uncertain time series analysis to predict
future values based on previously observed values which are described by uncertain
variables. According to the principle of least squares, a minimization problem was
derived to calculate the unknown parameters in the uncertain time series model. Fur-
thermore, residual of disturbance term was analyzed, and a confidence interval was
also defined to predict the future value.

Acknowledgements The authors gratefully acknowledge the financial support provided by National Natural
Science Foundation of China (Grants Nos. 61573210 and 61873329).

References
Box, G. E. P., & Jenkins, G. M. (1970). Time series analysis: Forecasting and control. San Francisco:
Holden-Day.
Cai, Q., Zhang, D., Zheng, W., & Leung, S. C. (2015). A new fuzzy time series forecasting model combined
with ant colony optimization and auto-regression. Knowledge-Based Systems, 74, 61–68.
Chen, S. M. (2002). Forecasting enrollments based on high order fuzzy time series. Cybernetics and Systems,
33(1), 1–16.
Domańska, D., & Wojtylak, M. (2012). Application of fuzzy time series models for forecasting pollution
concentrations. Expert Systems with Applications, 39(9), 7673–7679.
Egrioglu, E., Yolcu, U., Aladag, C. H., & Kocak, C. (2013). An ARMA type fuzzy time series forecasting
method based on particle swarm optimization. Mathematical Problems in Engineering, 2013, Article
ID 935815.
Huarng, K. (2001). Effective lengths of intervals to improve forecasting in fuzzy time series. Fuzzy Sets
and Systems, 123(3), 387–394.

123
X. Yang, B. Liu

Lee, L. W., Wang, L. H., Chen, S. M., & Leu, Y. H. (2006). Handling forecasting problems based on
two-factors high-order fuzzy time series. IEEE Transactions on Fuzzy Systems, 14(3), 468–477.
Lee, H. S., & Chou, M. T. (2014). Fuzzy forecasting based on fuzzy time series. International Journal of
Computer Mathematics, 81, 781–789.
Lio, W., & Liu, B. (2018a). Uncertain data envelopment analysis with imprecisely observed inputs and
outputs. Fuzzy Optimization and Decision Making, 17(3), 357–373.
Lio, W., & Liu, B. (2018b). Residual and confidence interval for uncertain regression model with imprecise
observations. Journal of Intelligent and Fuzzy Systems, 35(1), 2573–2583.
Liu, B. (2007). Uncertainty theory (2nd ed.). Berlin: Springer.
Liu, B. (2009). Some research problems in uncertainty theory. Journal of Uncertain Systems, 3(1), 3–10.
Liu, B. (2010). Uncertainty theory: A branch of mathematics for modeling human uncertainty. Berlin:
Springer.
Nejad, Z. M., & Ghaffari-Hadigheh, A. (2018). A novel DEA model based on uncertainty theory. Annals
of Operations Research, 264, 367–389.
Sheng, Y. H., & Kar, S. (2015). Some results of moments of uncertain variable through inverse uncertainty
distribution. Fuzzy Optimization and Decision Making, 14(1), 57–76.
Song, Q., & Chissom, B. S. (1993a). Forecasting enrollments with fuzzy time series: Part I. Fuzzy Sets and
Systems, 54, 1–9.
Song, Q., & Chissom, B. S. (1993b). Fuzzy time series and its models. Fuzzy Sets and Systems, 54, 269–277.
Song, Q., & Chissom, B. S. (1994). Forecasting enrollments with fuzzy time series: Part II. Fuzzy Sets and
Systems, 62, 1–8.
Sullivan, J., & Woodall, W. H. (1994). A comparison of fuzzy forecasting and Markov modeling. Fuzzy
Sets and Systems, 64(3), 279–293.
Tseng, F. M., Tzeng, G. H., Yu, H. C., & Yuan, B. J. C. (2001). Fuzzy ARIMA model for forecasting the
foreign exchange market. Fuzzy Sets and Systems, 118(1), 9–19.
Wen, M. L., Zhang, Q. Y., Kang, R., & Yang, Y. (2017). Some new ranking criteria in data envelopment
analysis under uncertain environment. Computers and Industrial Engineering, 110, 498–504.
Yao, K. (2018). Uncertain statistical inference models with imprecise observations. IEEE Transactions on
Fuzzy Systems, 26(2), 409–415.
Yao, K., & Liu, B. (2018). Uncertain regression analysis: An approach for imprecise observations. Soft
Computing, 22(17), 5579–5582.
Yule, G. U. (1927). On a method of investigating periodicities in disturbed series with special reference to
Wolfer’s sunspot numbers. Philosophical Transactions of the Royal Society of London, 226, 267–298.
Walker, G. T. (1931). On periodicity in series of related terms. Proceedings of the Royal Society of London,
Series A, 131, 518–532.

123

You might also like