This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

Digital Object Identifier

Time Series Forecasting Based on

Complex Network Analysis
SHENGZHONG MAO1 , FUYUAN XIAO2 ,∗
1
School of Hanhong, Southwest University, Chongqing 400715, China
2
School of Computer and Information Science, Southwest University, Chongqing 400715, China
∗
Corresponding author: Fuyuan Xiao (e-mail: [email protected], [email protected]).
The authors greatly appreciate the reviews’ suggestions and the editor’s encouragement. This research was funded by the Chongqing
Overseas Scholars Innovation Program (No. cx2018077).

ABSTRACT Time series forecasting, especially from the perspective of network, has been a hot research
topic. In this paper, based on the analysis of complex network, a novel method is proposed for more accurate
time series predictions. Firstly, time series data are mapped into a network by visibility graph. Then, the link
prediction method is adopted to calculate similarity index. Considering that node distance is an important
factor in the network, we take it into account to determine the weight coefficients and improve the predictive
results. To fully verify the validity of the proposed method, it is applied to some representative time series
data sets with different characteristics. The data values are recorded daily, monthly, and yearly, respectively.
The error measurement and correlation analysis show that our method has a good prediction performance.
It is believed that our work will not only contribute to time series forecasting in theory, but also take effect
in practice.

INDEX TERMS Complex network, Link prediction, Node distance, Time series, Visibility graph

I. INTRODUCTION to achieve better predictability [26].

Time series is a data sequence in chronological order [1]. These theories and models are considerably valuable.
There are many typical examples of time series, like solar However, they still have their limitations. Although SMA
irradiance [2], probabilistic prediction [3] and so forth [4], model can predict future values directly and it is easy to
[5]. Time series research includes many aspects, such as implement, it is less accurate. ARIMA model can improve
time series analysis [6], data aggregation and storage [7], the accuracy but it needs many observed values to train
[8], time series prediction [9], [10]. Time series forecasting parameters when modelling. Besides, it is only useful for
is to predict future data or variation tendency by analyzing some specific forecast targets so it can not be applicable
historical data and it is a hot research topic because of its to predict different kinds of time series. Therefore, it is
wide applications in finance [11], construction costs predic- urgent to find new methods that can improve prediction
tion [12], and some other fields. accuracy and are easy to implement. Since these existing
To make accurate predictions, many forecasting methods direct prediction methods have different shortcomings, we
have been proposed [13]. Simple moving average (SMA) can consider the problem of time series prediction from other
[14] and exponential smoothing (ES) [15] are two simple perspectives. Complex network has been a popular academic
forecasting methods. Besides, autoregressive integrated and research direction for its widespread applications [27]–[29].
moving average (ARIMA) [16] and seasonal ARIMA models One possible solution is to use the network approach as
are developed that can study linear and stationary time series. complex network has a close relationship with the time series
Commonly, uncertainties and fluctuations may bring some [30], [31], and the network can be used to identify time series
forecasting errors [17]–[19]. Researchers have conducted information [32], [33]. In some ways, complex network can
many studies on uncertainties [20], [21]. Due to the efficiency also make predictions through reasonable analysis [34]. The
to handle uncertainty [22], [23], fuzzy sets theory has been real systems are very complicated with many factors interact-
investigated a lot and it also can be used in time series ing with each other [35]–[37]. As a result, complex network
prediction [24], [25]. Apart from that, some new methods analysis is widely used due to its efficiency to model the
have been proposed by many other researchers and scientists relationship between each factor or node [38]. It is promising

VOLUME 4, 2016 1

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

to analyze time series in network models. A. VISIBILITY GRAPH

Recently, a novel perspective that complex network can Lacasa et al. first proposed visibility graph to characterise
be used to study time series has attracted much attention. time series from the perspective of complex network [39].
The visibility graph, proposed by Lacasa et al. is considered The visibility algorithm, which can transform a time series
to be a bridge between time series and complex network into a network, is defined as follows.
[39]. A time series can be mapped into a graph by visibility Two time series data (t1 , y1 ) and (t2 , y2 ) are considered
algorithm. The properties of time series are also inherited in as two visible and connected nodes in a network, if any data
the graph. Based on visibility graph, many researchers have s- value (t3 , y3 ) between them fulfills (1).
tudied time series through network. Jiang et al. used visibility t2 − t3
graph to aggregate time series data and successfully applied y3 < (y1 − y2 ) + y2 . (1)
t2 − t1
it to water management [40]. Lacasa et al. studied visibility
graph to research Fractional Brownian Motion (FBM) which As presented in Fig. 1, a vertical bar represents a node
could be used to study a real time series appearing in diverse and the edge connecting two vertical bars represents a link
scientific fields [41]. Taking advantages of visibility graph in the network. Two nodes can be connected if they meet the
theories, Donner and Donge found out the potentials and requirement of visibility criteria.
possible pitfalls by geophysical time series [42]. Inspired
B. LINK PREDICTION
by these researches, we can use the visibility algorithm to
convert the time series into a network. Then time series can Link prediction is able to explore potential links in the
be studied in the network using visibility graph. network. It is believed that two nodes tend to be linked if
the have higher similarity index. Based on LRW, Liu and
Besides, the missing link forecasting is supposed to be
Lü proposed an efficient method to predict missing links by
accounted urgently [43], [44]. It should be pointed out that
defining the node similarity [50].
modeling uncertainty of link in complex network has been
In a network, the probability that a random walker de-
paid great attention based on belief function [45]–[47] and
parting from node x is denoted by ~πx . In t steps (t is a
entropy function [48], [49]. In order to solve this problem,
large enough value to ensure the walker can fully walk in
Liu and Lü introduced a link prediction method based on
the network), ~πx (t) meets the following requirement:
local random walk (LRW ) [50]. Considering the relation-
ship between nodes and links in the network [51], [52], link
~πx (t) = P T ~πx (t − 1). (2)
prediction is an efficient way to explore wether there exists a
link between two nodes. It should pointed out that P T is the transpose of transition
In this paper, we propose a novel method for time series probability matrix P. The probability of moving from node x
a
forecasting based on complex network analysis. Specifically, to node y in one step is denoted by ~πxy . Pxy = kxy x
, where
a time series is first transformed into a network by visibility axy = 1 if point x is linked to point y, otherwise axy = 0.
graph. Then link prediction method is adopted to calculate Besides, kx means the degree of point x. ~πx (0) is an N × 1
the similarity between two nodes. Considering that node vector indicating the initial state of node x, where the x-th
distance is a vital element in the network, we take it into element is equal to 1 and the others are 0.
account to determine the weight coefficients and obtain the Equation (3) presents the definition of the similarity be-
prediction results. To demonstrate the validity of proposed tween node x and node y [50].
method, different kinds of time series datasets, including kx ky
LRW
Taiwan Capitalization Weighted Stock Index (TAIEX), State Sxy (t) = πxy (t) + πyx (t), (3)
2 |E| 2 |E|
Bank of India Share (SBI) share price at BSE, Construction
cost index (CCI), and university enrollment are adopted in where |E| is the total number of edges in the network. Finally,
the experimental section. The comparison results and error according to [50], the higher similarity between node x and
LRW
analysis show that our method has a good prediction perfor- node y can be obtained by adding up the results of Sxy in
mance. each step. The result is calculated by (4) after a superposed
random walk (SRW ) .
The structure of this paper as follows: Section 2 introduces
some basic concepts. In Section 3, the proposed method is t
X
SRW LRW
illustrated in detail. Section 4 demonstrates the predictability Sxy (t) = Sxy (l). (4)
of the method by some different experiments. Section 5 l=1

summarizes the paper and gives a brief conclusion.

C. NODE DISTANCE
According to [34], the node distance in a visibility graph is
II. PRELIMINARIES defined by (5).
In this section, we will give a brief introduction to some of di→j = |ti − tj | , (5)
the basic theories involved in the proposed method.
where ti and tj are the corresponding time values in point
(ti , yi ) and node (tj , yj ), respectively.
2 VOLUME 4, 2016

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

𝑡
1 .2 3 .4 5 .......6 7

FIGURE 1: The description of visibility graph: A time series including 7 data values is converted to a visibility graph. The
vertical bars represent nodes and the edges connecting two bars represent links in the network.

III. THE PROPOSED METHOD Step 5 Make predictions

In this section, the time series prediction method including Based on the consideration of node similarity and node
five main steps is proposed. distance, the prediction method is determined.
Step 1 Transform a time series to a visibility graph Firstly, it is in consideration of the two following sig-
A given time series T = {(t1 , y1 ), (t2 , y2 ), ..., (tN , yN )} nificant factors that we combine node (tM , yM ) and node
can be transformed into a visibility graph by (1) and the (tN , yN ) to determine the future node (tN +1 , yN +1 ).
corresponding figure is shown in the lower part of Fig. 1. 1) The most similar node (tM , yM )
Step 2 Calculate node similarity Time series forecasting predicts future values based on
The similarity of any two nodes is firstly calculated by previously observed values. The node (tM , yM ) is the
LRW
(3). Then, to sum the results of Sxy , a higher similarity representative of the previous data. It is considered to
is obtained based on (4). carry all the historical information so it can be used to
predict the future value.
Step 3 Find out the most similar node
2) The last observed node (tN , yN )
The similarities between the last node N and the
The node (tN , yN ) is the closest point to the future n-
preceding (N − 1) nodes are represented by SSRW =
  ode (tN +1 , yN +1 ). Besides, it will be directly connect-
S1N , S2N , ..., SM N , ..., S(N −1)N . The maximum value of
ed to node (tN +1 , yN +1 ). Therefore, the last observed
S SRW is denoted as SM N . Accordingly, node (tM , yM )
value yN is supposed to have an direct impact on the
corresponding to SM N is identified as the most similar point
predicted value yN +1 .
to node (tN , yN ).
As node (tM , yM ) and node (tN , yN ) has the highest
Step 4 Determine node distance
similarity, they are linearly connected to calculate the future
Based on (5), the distance between two most similar points
value of yN +1 .
(tM , yM ) and (tN , yN ), is calculated by (6).
yN − yM
dM →N = |tN − tM | , (6) yN +1 = yN + (tN +1 − tN ). (9)
tN − tM
If the node to be predicted is denoted as (tN +1 , yN +1 ), the
As shown in Fig. 2, node (t3 , y3 ) and node (t7 , y7 ) are
distance between the last node (tN , yN ) and its next point
directly linked to determine the node (tN +1 , yN +1 ), i.e.
(tN +1 , yN +1 ) is calculated by (7). Similarly, the distance
(t8 , y8 ).
between node (tM , yM ) and future node (tN +1 , yN +1 ) is
determined by (8). Furthermore, to improve the forecasting results, the node
distance obtained from Step 4 is taken into account to deter-
dN →N +1 = |tN +1 − tN | . (7) mine the weight coefficients of yN +1 and yN , respectively.
Commonly, the further the node (tM , yM ) is from the
dM →N +1 = |tN +1 − tM | . (8)
point (tN , yN ), the more important the node (tM , yM ) is
For example, in Fig. 2, the distance between node (t4 , y4 ) as it carries more historical information. Conversely, if node
and node (t7 , y7 ), i.e. d47 , is equal to 3. (tM , yM ) is pretty close to node (tN , yN ), the importance of
VOLUME 4, 2016 3

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

denotes the weight coefficient of the last observed value yN .

y Fig. 4 describes it in a more vivid way.
(𝑡𝑁 , 𝑦𝑁 ) As can be seen from (10) and (11), if dM →N increases,
(𝑡𝑁+1 , 𝑦𝑁+1 ) the weight coefficient of yN +1 , i.e. wN +1 , becomes larger
(𝑡𝑀 , 𝑦𝑀 ) while wN gets smaller. Conversely, wN +1 will be smaller and
wN gets larger when dM →N decreases. Therefore, the final
prediction results are calculated by (12).
ŷ N +1 = wN +1 × yN +1 + wN × yN . (12)

The proposed method is summarized by Algorithm 1 and

Fig. 5 shows the process in a more vivid way.

Algorithm 1 The process of the proposed method

1 2 3 4 5 6 7 .8 t Input: Time series dataset T : N data values
Output: The prediction value of ŷ N +1
FIGURE 2: The two most similar points, node (t3 , y3 ) and 1: function T IME SERIES PREDICTION
node (t7 , y7 ) are directly connected to determine the value of 2: Transform time series T into a graph
y8 . 3: Calculate node similarity Sxy SRW

4: Find two most similar points (tM , yM ) and (tN , yN )

5: Determine node distance dM →N
node (tM , yM ) is weakened because it contains less informa- 6: yN +1 ← yN + ytN −yM
(tN +1 − tN )
N −tM
tion about the past and then takes almost the same effect as 7: ŷ N +1 ← wN +1 × yN +1 + wN × yN
node (tN , yN ) for future predictions. return (tN +1 , ŷ N +1 )
As shown in Fig. 3, node (tM , yM ) and node (tN , yN ) 8: end function
becomes closer from tM 1 to tM 3 and the distance dM →N
is getting smaller at the same time.

Step 1 From time series to visibility graph

(𝑡𝑁 ，𝑦𝑁 )
Step 2 Calculate node similarity

Steps Step 3 Find out the most similar node

(𝑡𝑀1 , 𝑦𝑀 )
(𝑡𝑀3 , 𝑦𝑀 )
(𝑡𝑀2 , 𝑦𝑀 )
Step 4 Determine node distance

𝑡
𝑑𝑀𝑁 𝑑𝑁(N+1)
𝑑𝑀(N+1) Step 5 Make predictions

FIGURE 3: As node (tM , yM ) moves from tM 1 to tM 3 , the

distance dM →N is getting smaller. The historical information
FIGURE 5: The flowchart of the proposed method.
that node (tM , yM ) carries is less and its importance is
weakened meanwhile.

On the basis of such consideration, the weight coefficients IV. EXPERIMENTS AND ANALYSIS
are defined by (10) and (11). In this section, the proposed method is applied to some
representative datasets in different fields. These data values dM N  tN  t
dM →N
wN +1 = , (10) with different characteristics are published daily, monthly,
dM →N +1
and yearly, respectively.
dN →N +1 Determination of weight dM N
wN = , (11) In 2orderStep
Part to evaluate
5 the experimental results, we usefour
N  N 1
dM →N +1 coefficients
error indicators, i.e. mean absolute difference (MAD), mean d M  N 1

where wN +1 denotes the weight coefficient of yN +1 and wN absolute percentage error (MAPE), root mean square error
4 ŷ𝑁+1
VOLUME 4, = 𝜔𝑁 𝑦𝑁 + 𝜔𝑁
2016

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

y
𝑦𝑁+1 𝜔𝑁+1
(𝑡𝑁 ，𝑦𝑁 )
𝑦𝑁 𝜔𝑁

(𝑡𝑀 ，𝑦𝑀 )

𝑡𝑀 𝑡𝑁 𝑡𝑁+1 t

FIGURE 4: wN +1 (blue curve) is the weight coefficient of yN +1 and wN (red curve) is the weight coefficient of yN .

(RMSE), and normalized root mean squared error (NRMSE). Table 1 and plotted in Fig. 6. As can be seen, the proposed
They are defined by (13)-(16). method has the minimum error compared to other models.
N
1 X 2) SBI share price forecasting
M AD = |ŷ(t) − y(t)| (13)
N t=1 The SBI share prices fluctuate greatly, so they are ideal
N
for testing the predictability of the method. The time series
1 X |ŷ(t) − y(t)| data values from April 2008 to March 2010 are chosen as
M AP E = (14)
N t=1 y(t) a universe dataset. The data from April 2008 to February
v 2010 are adopted to predict the data values from August
u
u1 X N 2008 to March 2010. Table 2 lists the error measurement and
2
RM SE = t [ŷ(t) − y(t)] (15) correlation analysis of some classic and recent methods.
N t=1
As can be seen, the proposed method has minimum the
s
N
RM SE compared with other methods. In addition, correla-
1
P 2 tion coefficient (R > 0.9) indicates that the predicted and
N [ŷ(t) − y(t)]
t=1 true values are highly correlated.
N RM SE = (16)
ymax − ymin
TABLE 2: The comparison of error measurement and corre-
where ŷ(t) is the predicted value, y(t) is the true value and N lation analysis in SBI prediction
is the total number of ŷ(t).
Prediction methods RMSE MAPE(%) R R2
A. FINANCIAL TIME SERIES FORECASTING
In order to evaluate the predictive performance of the pro- Pathak and Singh [55] 205.96 8.95 0.8686 0.7544
posed method in financial field, we use it to predict TAIEX
and SBI share price at BSE. Joshi and Kumar [56] 200.17 9.52 0.8821 0.7781
Bisht et al. [57] 327.22 1.49 0.9821 0.9645
1) TAIEX forecasting
Taiwan Capitalization Weighted Stock Index (TAIEX) is a Proposed method 199.03 7.97 0.9056 0.8201
stock market index that releases every day. As one of the main
indicators of Taiwan’s economic trend, it fluctuates greatly
and attracts many investors and economists [53], [54]. B. CCI FORECASTING
In this case, the index values from 1991 to 1999 are Construction Cost Index (CCI) published by Engineering
selected as the whole data sets. Considering the data go News Record (ENR) monthly. Many civil engineers and cost
through a long span time, they are divided into 9 groups by analysts have studied on CCI since it contains vital price
year. In each group, the index values of the first ten months information about building industry [58], [59].
are applied to predict the last two months. The errors results A CCI data set from January 1990 to July 2014 (295 data
of RMSE for three different forecasting methods are listed in values), is adopted as the universe set in this study. As shown
VOLUME 4, 2016 5

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

TABLE 1: The errors results of RMSE for TAIEX forecasting

Prediction methods 1991 1992 1993 1994 1995 1996 1997 1998 1999 Average

Chen et al. [53] 80 60 110 112 79 54 148 167 149 107

Yu et al. [54] 61 67 105 135 70 54 133 151 142 102
Propsed method 46 42 72 94 58 52 133 119 108 80

180

Chen et al. Yu et al. Proposed method

150

120
RMSE

90

60

30

0
1991 1992 1993 1994 1995 1996 1997 1998 1999 Average
Year

FIGURE 6: Comparison of RMSE for TAIEX forecasting.

in Fig. 7, it increases during a long period while fluctuates a dicted values (red curve) of CCI. As can be seen, the red
lot in a short time. To show the predictability of the proposed curve is very close to the black one, which roughly indicates
method, all the 295 data values are adopted. Specifically, that the proposed method has a good predictive performance.
the data values from January 1990 to June 2014 are used to However, there still exists continuous fluctuations that may
predict data from March 1991 to July 2014. cause forecasting errors (As shown in rectangular boxes in
Fig. 8).
To better illustrate the prediction performance of the
10000
method, it is compared with classical time series prediction
Actual value
method and network approach, including Simple Moving
9000
Average (SMA) model [60] and Zhang et al.’s method [34].
Table 3 lists the forecasting errors of the three methods in
8000
detail and Fig. 9 compares the errors in a more vivid way.
The results show that the proposed method can improve
CCI

7000
the accuracy compared with both classic and recent network
8000 methods. Although the improvement is not so big, it is still
6000
considerably important for cost analysts to make budgets.
CCI

7900
In the engineering field, even 0.1% improvement can create
5000 7800
200 205 210
huge benefits, especially when it comes to millions or billions
Time
of construction projects.
4000
0 50 100 150 200 250 300
Time C. UNIVERSITY ENROLLMENT FORECASTING
FIGURE 7: A CCI data set from January 1990 to July 2014. In this experiment, the proposed method is used to predict the
It increases during a long period but fluctuates in a short time. enrollment of the University of Alabama.
The enrollment data from 1971 to 1992 is chosen as a
Fig. 8 presents the actual values (black curve) and pre- universal data set. Likewise, the first two data values are
6 VOLUME 4, 2016

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

10000
Actual value
Predicted value

9000

8000 5850

5800
CCI

7000
5750
84 86 88
8050
6000 8000

7950

7900
5000
7850
200 202 204 206 208 210

4000
0 50 100 150 200 250 300
Time

FIGURE 8: The actual values (black curve) and predicted values (red curve) of CCI.

TABLE 3: The comparison for CCI forecasting of the three methods

Prediction methods MAD RMSE MAPE(%) NRMSE(%) R R2

SMA(k=1) 21.59 32.73 0.3110 0.6350 0.9998 0.9997

Zhang et al. [34] 20.05 29.33 0.2889 0.5690 0.9998 0.9997
Proposed method 19.30 28.16 0.2797 0.5462 0.9998 0.9997
RMSE
MAD

0 5 10 15 20 25 30 35
NRMSE(%)
MAPE(%)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

Proposed method Zhang et al. SMA(k=1)

FIGURE 9: The comparison of CCI forecasting errors by the measurement of four indicators.

VOLUME 4, 2016 7

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

adopted to predict the enrollment of 1973. Then the predicted series based on complex network theories. Besides, the link
enrollments from 1973 to 1992 can be obtained after this step prediction method is applied to explore potential links be-
is iterated for 20 times. tween two nodes. The node with highest similarity index and
The actual enrollment of the University of Alabama and the last observed node are directly linked to predict yN +1 .
predicted results from five different models are presented in The node distance is also a crucial factor in the network,
Table 4. The experimental errors and statistic performance so it is adopted for weight determination. The further node
of the methods are listed in Table 5. It is obvious that our (tM , yM ) is away from node (tN , yN ), the more previous
method has improved the forecasting accuracy compared information it carries. Consequently, the weight coefficient
with the first three methods, but less accurate than Kumar and of initial prediction results, i.e. wN +1 , becomes larger while
Gangwar’s method. Besides, the correlation coefficients of wN gets smaller accordingly.
the proposed method indicate that there is a good correlation The proposed method takes full advantages of the prop-
between observed and predicted values. From Fig. 10, we erties of the nodes themselves in the network, inheriting
can see that the actual and the predicted enrollments by the characteristics of time series. It is on the basis of such
the proposed method undergo the almost same variation analyses that the improved forecasting model is constructed.
tendency. In order to fully demonstrate effectiveness of the proposed
method, we adopt some classic time series datasets. Com-
D. ANALYSIS pared with other models and methods, the proposed method
According to the forecasting results and their errors analysis, achieves more accurate prediction. It is convince that our
we discuss the above experiments as follows. method is applicable to finance, construction industry, and
1) Firstly, the proposed method is applied to the field of some other fields.
finance. The TAIEX is recorded every day and fluctu- Although the proposed method has a good prediction
ates wildly. It is divided into groups by year because performance, it also can be improved in many aspects. For
there are too many data values and they experience example, the data mining method can be used to deal with the
such a long time. As can be seen from Table 1, the pro- time series data sets before mapped into the network. It may
posed method has the best prediction accuracy by the help to construct a more better network and further improve
measurement of RMSE, both in each group and over- the prediction accuracy. We will focus on this work in the
all average. Besides, in SBI share price forecasting, future.
our method also has the minimum value of RM SE.
Therefore, it is believed that our method is applicable REFERENCES
[1] Y. Guo, X. Song, N. Li, and D. Fang, “An efficient missing data prediction
to financial time series prediction.
method based on kronecker compressive sensing in multivariable time
2) CCI is an significant indicator in construction industry. series,” IEEE Access, vol. 6, pp. 57 239–57 248, 2018.
As shown in Fig. 7, the CCI data show an increasing [2] Y. Hirata and K. Aihara, “Improving time series prediction of solar
irradiance after sunrise: Comparison among three methods for time series
trend on the whole, but vary a lot from month to month.
prediction,” Solar Energy, vol. 149, pp. 294–301, 2017.
The proposed method requires at least two values to [3] J. M. Amigó, Y. Hirata, and K. Aihara, “On the limits of probabilistic
make a prediction for the first step, so the first two forecasting in nonlinear times series analysis,” Chaos: An Interdisciplinary
Journal of Nonlinear Science, vol. 26, no. 12, p. 123114, 2016.
data values are adopted. Table 1 lists the errors of each
[4] M. Hu, Z. Ji, K. Yan, Y. Guo, X. Feng, J. Gong, X. Zhao, and L. Dong, “De-
method. Although it seems that the accuracy has only tecting anomalies in time series data via a meta-feature based approach,”
improved a little by the proposed method, it may save a IEEE Access, 2018.
[5] F. Karim, S. Majumdar, H. Darabi, and S. Chen, “Lstm fully convolutional
lot of costs when applied to huge construction projects.
networks for time series classification,” IEEE Access, vol. 6, pp. 1662–
The R value (0.9998) is so close to 1, indicating that 1669, 2018.
the proposed method has achieved high correlation. [6] M. Fukino, Y. Hirata, and K. Aihara, “Coarse-graining time series data:
Recurrence plot of recurrence plots and its application for music,” Chaos:
3) Like CCI prediction, the first two values of enrollment
An Interdisciplinary Journal of Nonlinear Science, vol. 26, no. 2, p.
are first input to obtain the third data value. All the 023116, 2016.
forecasting results will be calculated by repeating the [7] M. Tahmassebpour, “A new method for time-series big data effective
storage,” Ieee Access, vol. 5, pp. 10 694–10 699, 2017.
first step. As shown in Fig. 10, the predicted values
[8] W. Jiang, B. Wei, J. Zhan, C. Xie, and D. Zhou, “A visibility graph
(red curve) are close to the actual ones (black curve). power averaging aggregation operator: A methodology based on network
Meanwhile, the proposed method also shows a better analysis,” Computers &amp; Industrial Engineering, vol. 101, pp. 260–
268, 2016.
predictive accuracy than other four methods. The ac-
[9] P. Li, J. Zhang, C. Li, B. Zhou, Y. Zhang, M. Zhu, and N. Li, “Dynamic
curate prediction of enrollment for the next year helps similar sub-series selection method for time series forecasting,” IEEE
school to go in normal operation. Access, 2018.
[10] H. S. Milad, U. Farooq, M. E. El-Hawary, and M. U. Asad, “Neo-fuzzy
integrated adaptive decayed brain emotional learning network for online
V. CONCLUSION time series prediction.” IEEE Access, vol. 5, pp. 1037–1049, 2017.
In this paper, based on the analysis of complex network, a [11] M. M. Rounaghi and F. N. Zadeh, “Investigation of market efficiency
novel method for time series forecasting is proposed. Firstly, and financial stability between s&amp;p 500 and london stock exchange:
Monthly and yearly forecasting of time series stock returns using arma
the time series is transformed into a network by visibility model,” Physica A: Statistical Mechanics and its Applications, vol. 456,
graph. This provides a new point of view to study on time pp. 10–21, 2016.

8 VOLUME 4, 2016

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

TABLE 4: The forecasting results of student enrollment.

Pathak and Gangwar and Kumar and

Year Actual Chou [61] Proposed
Singh [55] Kumar [62] Gangwar [63]

1971 13055
1972 13563 14025 13250 14586 13693
1973 13867 14568 13750 14586 13693 13817.0
1974 14696 14568 13750 15363 14867 14019.0
1975 15460 15654 14500 15363 15287 15073.7
1976 15311 15654 15375 15442 15376 15842.0
1977 15603 15654 15375 15442 15376 15236.5
1978 15861 15654 15625 15442 15376 15650.7
1979 16807 16197 15875 15442 16523 15961.3
1980 16919 17283 16833 17064 16606 17076.4
1981 16388 17283 16833 17064 17519 17162.2
1982 15433 16197 16500 15438 16606 16520.6
1983 15497 15654 15500 15442 15376 15429.6
1984 15145 15654 15500 15442 15376 15501.1
1985 15163 15654 15125 15363 15287 15113.5
1986 15984 15654 15125 15363 15287 15136.0
1987 16859 15654 16833 15438 16523 16105.8
1988 18150 16197 16667 17064 17519 17296.5
1989 18970 17283 18125 19356 19500 18795.5
1990 19328 18369 18750 19356 19000 19380.0
1991 19337 19454 19500 19356 19500 19547.0
1992 18876 19454 19500 19356 19500 19538.5

TABLE 5: The forecasting error measurement and statistical performance analysis of the five methods

Prediction methods MAD RMSE MAPE(%) NRMSE(%) R R2

Chou [61] 605.00 781.47 3.6075 12.4399 0.8866 0.7860

Pathak and Singh [55] 493.71 646.67 2.9883 10.2940 0.9373 0.8785
Gangwar and Kumar [62] 476.24 642.68 2.9672 10.2305 0.9243 0.8544
Kumar and Gangwar [63] 386.24 493.56 2.3366 7.8568 0.9594 0.9205
Proposed method 455.64 559.36 2.7858 8.8984 0.9426 0.8886

VOLUME 4, 2016 9

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

22,000
Actual value
Predicted value
20,000
Student enrollment

18,000

16,000

14,000

12,000

10,000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Time

FIGURE 10: The actual values (black curve) and predicted values (red curve) of student enrollment.

[12] T. Moon et al., “Forecasting model of construction cost index based on fluctuation trends and particle swarm optimization machine learning,”
vecm with search query,” KSCE Journal of Civil Engineering, vol. 22, Symmetry, vol. 9, no. 7, p. 124, 2017.
no. 8, pp. 2726–2734, 2018. [27] B. Wei and Y. Deng, “A cluster-growing dimension of complex networks:
[13] H. Guan, S. Guan, and A. Zhao, “Forecasting model based on neutrosophic From the view of node closeness centrality,” Physica A: Statistical Me-
logical relationship and jaccard similarity,” Symmetry, vol. 9, no. 9, p. 191, chanics & Its Applications, p. 10.1016/j.physa.2019.01.125, 2019.
2017. [28] D. R. Amancio, O. N. O. Jr, and L. D. F. Costa, “Identification of literary
[14] S. Guan and A. Zhao, “A two-factor autoregressive moving average model movements using complex networks to represent texts,” New Journal of
based on fuzzy fluctuation logical relationships,” Symmetry, vol. 9, no. 10, Physics, vol. 14, no. 21, pp. 1759–1761, 2012.
p. 207, 2017. [29] L. Yin and Y. Deng, “Toward uncertainty of weighted networks: An
[15] C. C. Holt, “Forecasting seasonals and trends by exponentially weighted entropy-based model,” Physica A: Statistical Mechanics and its Applica-
moving averages,” International journal of forecasting, vol. 20, no. 1, pp. tions, vol. 508, pp. 176–186, 2018.
5–10, 2004. [30] R. V. Donner, M. Small, J. F. Donges, N. Marwan, Y. Zou, R. Xiang, and
[16] F.-M. Tseng, H.-C. Yu, and G.-H. Tzeng, “Combining neural network J. Kurths, “Recurrence-based time series analysis by means of complex
model with seasonal time series arima model,” Technological Forecasting network methods,” International Journal of Bifurcation and Chaos, vol. 21,
and Social Change, vol. 69, no. 1, pp. 71–87, 2002. no. 04, pp. 1019–1046, 2011.
[17] Y. Han and Y. Deng, “A hybrid intelligent model for assessment of [31] S. Liu, X. Gao, W. Fang, Q. Sun, S. Feng, X. Liu, and S. Guo, “Modeling
critical success factors in high risk emergency system,” Journal of Ambient the complex network of multidimensional information time series to
Intelligence and Humanized Computing, vol. 9, no. 6, pp. 1933–1953, characterize the volatility pattern evolution,” IEEE Access, vol. 6, pp.
2018. 29 088–29 097, 2018.
[18] F. Xiao, “A novel multi-criteria decision making method for assessing [32] L. Fei, Q. Zhang, and Y. Deng, “Identifying influential nodes in complex
health-care waste treatment technologies based on d numbers,” Engineer- networks based on the inverse-square law,” Physica A: Statistical Mechan-
ing Applications of Artificial Intelligence, vol. 71, pp. 216–225, 2018. ics and its Applications, vol. 512, pp. 1044–1059, 2018.
[19] H. Mo and Y. Deng, “A new mada methodology based on d numbers,” [33] Z.-K. Gao, Y.-X. Yang, P.-C. Fang, Y. Zou, C.-Y. Xia, and M. Du,
International Journal of Fuzzy Systems, vol. 20, no. 8, pp. 2458–2469, “Multiscale complex network for analyzing experimental multivariate time
2018. series,” EPL (Europhysics Letters), vol. 109, no. 3, p. 30005, 2015.
[20] Y. Li and Y. Deng, “Generalized ordered propositions fusion based on [34] R. Zhang, B. Ashuri, and Y. Deng, “A novel method for forecasting
belief entropy,” International Journal of Computers Communications & time series based on fuzzy logic and visibility graph,” Advances in Data
Control, vol. 13, no. 5, pp. 792–807, 2018. Analysis and Classification, vol. 11, no. 4, pp. 759–783, 2017.
[21] R. Sun and Y. Deng, “A new method to identify incomplete frame of [35] Y. Wang, S. Wang, and Y. Deng, “A modified efficiency centrality to
discernment in evidence theory,” IEEE Access, vol. 7, no. 1, pp. 15 547– identify influential nodes in weighted networks,” Pramana, vol. 92, no. 4,
15 555, 2019. p. 68, 2019.
[22] F. Xiao, “A hybrid fuzzy soft sets decision making method in medical [36] P. Zhu, X. Song, L. Liu, Z. Wang, and J. Han, “Stochastic Analysis of
diagnosis,” IEEE Access, vol. 6, pp. 25 300–25 312, 2018. Multiplex Boolean Networks for Understanding Epidemic Propagation,”
[23] X. Gao and Y. Deng, “The negation of basic probability assignment ,” IEEE Access, 2018.
IEEE Access, vol. 7, no. 1, p. 10.1109/ACCESS.2019.2901932, 2019. [37] W. Deng and Y. Deng, “Entropic methodology for entanglement mea-
[24] A. Zhao, S. Guan, and H. Guan, “A computational fuzzy time series sures,” Physica A: Statistical Mechanics and its Applications, vol. 512,
forecasting model based on gem-based discretization and hierarchical pp. 693–697, 2018.
fuzzy logical rules,” Journal of Intelligent &amp; Fuzzy Systems, vol. 31, [38] H. Yang, Y. Deng, and J. Jones, “Network division method based on
no. 6, pp. 2795–2806, 2016. cellular growth and physarum-inspired network adaptation,” International
[25] Y. Han and Y. Deng, “An enhanced fuzzy evidential DEMATEL method Journal of Unconventional Computing, vol. 13, no. 6, pp. 477–491, 2018.
with its application to identify critical success factors,” Soft computing, [39] L. Lacasa, B. Luque, F. Ballesteros, J. Luque, and J. C. Nuño, “From
vol. 22, no. 15, pp. 5073–5090, 2018. time series to complex networks: The visibility graph,” Proceedings of the
[26] J. Jia, A. Zhao, and S. Guan, “Forecasting based on high-order fuzzy- National Academy of Sciences, vol. 105, no. 13, pp. 4972–4975, 2008.

10 VOLUME 4, 2016

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2906268, IEEE Access

[40] W. Jiang, B. Wei, Y. Tang, and D. Zhou, “Ordered visibility graph average
aggregation operator: An application in produced water management,”
Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 27, no. 2,
p. 023117, 2017.
[41] L. Lacasa, B. Luque, J. Luque, and J. C. Nuno, “The visibility graph:
A new method for estimating the hurst exponent of fractional brownian
motion,” EPL (Europhysics Letters), vol. 86, no. 3, p. 30001, 2009.
[42] R. Donner and J. Donges, “Visibility graph analysis of geophysical time
series: Potentials and possible pitfalls,” Acta Geophysica, vol. 60, no. 3,
pp. 589–623, 2012.
[43] L. Lü and T. Zhou, “Link prediction in complex networks: A survey,”
Physica A: Statistical Mechanics and its Applications, vol. 390, no. 6, pp.
1150–1170, 2011.
[44] L. Lü, L. Pan, T. Zhou, Y.-C. Zhang, and H. E. Stanley, “Toward link pre-
dictability of complex networks,” Proceedings of the National Academy of
Sciences, vol. 112, no. 8, pp. 2325–2330, 2015.
[45] L. Yin, X. Deng, and Y. Deng, “The negation of a basic probability
assignment,” IEEE Transactions on Fuzzy Systems, vol. 27, no. 1, pp. 135–
143, 2019.
[46] F. Xiao, “Multi-sensor data fusion based on the belief divergence measure
of evidences and the belief entropy,” Information Fusion, vol. 46, no. 2019,
pp. 23–32, 2019.
[47] H. Zhang and Y. Deng, “Engine fault diagnosis based on sensor data
fusion considering information quality and evidence theory,” Advances in
Mechanical Engineering, vol. 10, no. 11, p. 1687814018809184, 2018.
[48] F. Xiao, “An improved method for combining conflicting evidences based
on the similarity measure and belief function entropy,” International Jour-
nal of Fuzzy Systems, vol. 20, no. 4, pp. 1256–1266, 2018.
[49] B. Kang, Y. Deng, K. Hewage, and R. Sadiq, “Generating Z- number
based on OWA weights using maximum entropy,” International Journal
of Intelligent Systems, vol. 33, no. 8, pp. 1745–1755, 2018.
[50] W. Liu and L. Lü, “Link prediction based on local random walk,” EPL
(Europhysics Letters), vol. 89, no. 5, p. 58007, 2010.
[51] P. Xu, R. Zhang, and Y. Deng, “A novel visibility graph transformation of
time series into weighted networks,” Chaos, Solitons & Fractals, vol. 117,
pp. 201–208, 2018.
[52] M. Li, Q. Zhang, and Y. Deng, “Evidential identification of influential
nodes in network of networks,” Chaos, Solitons & Fractals, vol. 117, pp.
283–296, 2018.
[53] S.-M. Chen, “Forecasting enrollments based on fuzzy time series,” Fuzzy
sets and systems, vol. 81, no. 3, pp. 311–319, 1996.
[54] H.-K. Yu, “Weighted fuzzy time series models for taiex forecasting,”
Physica A: Statistical Mechanics and its Applications, vol. 349, no. 3-4,
pp. 609–624, 2005.
[55] H. K. Pathak and P. Singh, “A new bandwidth interval based forecasting
method for enrollments using fuzzy time series,” Applied Mathematics,
vol. 2, no. 04, p. 504, 2011.
[56] B. P. Joshi and S. Kumar, “A computational method of forecasting based
on intuitionistic fuzzy sets and fuzzy time series,” in Proceedings of
the International Conference on Soft Computing for Problem Solving
(SocProS 2011) December 20-22, 2011. Springer, 2012, pp. 993–1000.
[57] K. Bisht, D. K. Joshi, and S. Kumar, “Dual hesitant fuzzy set-based
intuitionistic fuzzy time series forecasting,” in Ambient Communications
and Computer Systems. Springer, 2018, pp. 317–329.
[58] S. Shahandashti and B. Ashuri, “Forecasting engineering news-record
construction cost index using multivariate time series models,” Journal of
Construction Engineering and Management, vol. 139, no. 9, pp. 1237–
1243, 2013.
[59] S. Hwang, “Time series models for forecasting construction costs using
time series indexes,” Journal of Construction Engineering and Manage-
ment, vol. 137, no. 9, pp. 656–662, 2011.
[60] B. Ashuri and J. Lu, “Time series analysis of enr construction cost index,”
Journal of Construction Engineering and Management, vol. 136, no. 11,
pp. 1227–1237, 2010.
[61] M.-T. Chou, “Long-term predictive value interval with the fuzzy time
series,” Journal of Marine Science and Technology, vol. 19, no. 5, pp. 509–
513, 2011.
[62] S. S. Gangwar and S. Kumar, “Probabilistic and intuitionistic fuzzy sets–
based method for fuzzy time series forecasting,” Cybernetics and Systems,
vol. 45, no. 4, pp. 349–361, 2014.
[63] S. Kumar and S. S. Gangwar, “Intuitionistic fuzzy time series: an approach
for handling nondeterminism in time series forecasting,” IEEE Transac-
tions on Fuzzy Systems, vol. 24, no. 6, pp. 1270–1281, 2016.

VOLUME 4, 2016 11

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.