Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90

Contents lists available at ScienceDirect

Commun Nonlinear Sci Numer Simulat

journal homepage: www.elsevier.com/locate/cnsns

Research paper

Multivariate multiscale entropy of financial markets

Yunfan Lu∗, Jun Wang
School of Science, Beijing Jiaotong University, Beijing 100044, PR China

a r t i c l e i n f o a b s t r a c t

Article history: In current process of quantifying the dynamical properties of the complex phenomena in
Received 31 December 2016 financial market system, the multivariate financial time series are widely concerned. In this
Revised 10 April 2017
work, considering the shortcomings and limitations of univariate multiscale entropy in an-
Accepted 24 April 2017
alyzing the multivariate time series, the multivariate multiscale sample entropy (MMSE),
Available online 25 April 2017
which can evaluate the complexity in multiple data channels over different timescales, is
Keywords: applied to quantify the complexity of financial markets. Its effectiveness and advantages
Multivariate multiscale entropy analysis have been detected with numerical simulations with two well-known synthetic noise sig-
Nonlinear complexity behavior nals. For the first time, the complexity of four generated trivariate return series for each
Globe stock market stock trading hour in China stock markets is quantified thanks to the interdisciplinary ap-
Multivariate volatility series plication of this method. We find that the complexity of trivariate return series in each
Shuffled multivariate return hour show a significant decreasing trend with the stock trading time progressing. Fur-
ther, the shuffled multivariate return series and the absolute multivariate return series are
also analyzed. As another new attempt, quantifying the complexity of global stock markets
(Asia, Europe and America) is carried out by analyzing the multivariate returns from them.
Finally we utilize the multivariate multiscale entropy to assess the relative complexity of
normalized multivariate return volatility series with different degrees.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction

Nonlinear science, an important and challenging subject, has got many contributions about the modeling and the behav-
ior analysis in many branch field, such as the financial market analysis [1–3], the physiologic system [4], biological signal
[5], laster doppler flowmetry system [6], oil-water flow [7], the multi-hydro-turbine governing systems [8–10] and human
gait dynamics [11]. It is also noteworthy that the empirical studies which focus on the discovery of nonlinear features in
financial dynamics have been attracting extensive attentions in recent years. Considered as a complex evolved nonlinear
system, the financial market has the empirically observed stylized facts, such as fat tails phenomenon and power law
of logarithmic returns [1,12–14], long-term memory and volatility clustering [15,16], autocorrelation and cross-correlation
[17,18], multifractality of volatility and multifractal behaviors [2,19,20], complexity of financial systems [3,18]. These studies,
aiming at extending and consolidating the known stylized facts, are also essential for risk management and has aroused
wide attention from the community of researchers in nonlinear financial dynamical systems. Nowadays, many research
literatures have arisen to explore the nonlinear natures of financial markets [2,14,19,21–25]. One of the most significant
aspects is to quantify the complexity of financial time series. The recently introduced multiscale entropy (MSE) approach
proposed by Costa et al. [4] has become a useful and popular method to quantify the complexity of signals in different

∗
Corresponding author.
E-mail address: yunfanlu@bjtu.edu.cn (Y. Lu).

http://dx.doi.org/10.1016/j.cnsns.2017.04.028
1007-5704/© 2017 Elsevier B.V. All rights reserved.
78 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90

research fields, such as the biomedical and physiologic time series [5,6], the vibration of rotary machine [26], the human
gait pattern at different ages [11,27], and the financial time series [3]. The MSE method (that evaluates sample entropy from
univariate time series, coarse grained on multiple scales, and thus indicates long-range correlations over a range of temporal
scales within a complex system) has been proven to be able to distinguish physiological time series with different degrees
of complexity. However, it is also shown that this MSE method has some shortcomings and limitations, especially it is not
suited for multivariate time series that are routinely recorded in experimental, physical and biological systems thanks to the
recent developments in sensor technology. In the MSE method, multivariate time series are considered as a set of individual
time series, and they are separately calculated in each variable to assess their complexity. However, this is only adequate if
each variable in a multivariate signal is statistically independent or at the very least uncorrelated (which is usually not the
case). To generalize the univariate MSE to the multivariate case, Ahmed and Mandic introduced multivariate sample entropy
(MSampEn), and evaluate its evolution over multiple time scales to perform the multivariate multiscale entropy (MMSE)
analysis [28]. The introduced multivariate multiscale entropy (MMSE) has been shown substantial advantages in operating
on any number of data channels simultaneously, and providing a dynamical complexity measure for multivariate data
observed from the same system, especially if there is a large degree of uncertainty or coupling underlying the measured
system dynamics. The approach has been supported by simulations on both synthetic and real world multivariate processes,
such as gait, wind, and physiological data [7,28,29]. However, the multivariate multiscale entropy (MMSE) has not been
applied in analyzing the multivariate time series in financial market.
In the present work, we detect the effectiveness and advantages of the MMSE method in complexity analysis of multi-
variate time series and try to determine appropriate parameters. For example, the effective length of the multivariate time
series with small experimental error on two well-known synthetic noise signals, white noise (with uncorrelated fluctuation)
and 1/f noise (with correlated fluctuations), is studied. Then the MMSE method is applied in the complexity analysis
of China stock markets with four generated trivariate return series for each stock trading hour, also with their shuffled
multivariate return series and absolute multivariate return series. Finally, a new approach of analyzing the complexity of
stock markets for three major regions (Asia, Europe and America) of the world is carried out by studying the multivariate
stock returns (which are derived from the logarithmic returns of the stock indices) from them. The multivariate return
series with different degrees of volatility of three regions are also quantified by using the MMSE analysis. The logarithmic
returns of a stock index is defined as:

R(t ) = ln P (t ) − ln P (t − t ) (1)

where P (t ) denote the price of a given stock market index, and the corresponding price return R(t) is defined as the
change of the logarithm of the price in a given time interval t [30–32].

2. Multivariate multiscale entropy analysis

2.1. Multivariate sample entropy

Recently, Ahmed & Mandic [28] proposed the multivariate sample entropy (MSampEn), which enables entropy calcula-
tion for multichannel data by taking into account both within- and cross-channel dependencies, and then introduced it into
the multivariate multiscale entropy (MMSE) analysis. We show the multivariate sample entropy (MSampEn) in Algorithm 1,
and by extending the standard univariate sample entropy in Ref. [4], we introduce the multivariate multiscale entropy
(MMSE) analysis in Algorithm 2. In Algorithm 1, the multivariate sample entropy method is based on the estimation of
the conditional probability that two similar sequences will remain similar when the next data point is included. This is
achieved by calculating the average number of neighboring delay vectors for a given tolerance level r and repeating the
process after increasing the embedding dimension from m to m + 1. Fig. 1 illustrates the principle behind multivariate
sample entropy calculation with two high-frequency (five-minute interval) returns of Shanghai Stock Exchange (SSE) and
Shenzhen Stock Exchange (SZSE) from February 16, 2015 to February 18, 2015, separately denoted as x(t) and y(t). In order
to show clearly, each of the return series only shows 100 points in Fig. 1(a). For the illustration of principle, when the
embedding dimension m = 2, we assume that the time lag vector τ = [1, 1] and the embedding vector M = [1, 1], then
the composite bivariate delay vectors are [x(t), y(t)] as shown in Fig. 1(b), where t denotes the time as sample index. In
the process of MSampEn calculation, for any such vector (e.g., [x(42), y(42)]), we need to count the number of neighbors
which are within a distance r (tolerance level), illustrated by a circle centered at [x(42), y(42)] with radius r in Fig. 1(b).
And then, upon increasing the embedding dimension from m = 2 to m = 3, the embedding vector M evolves to two new
embedding vectors, M = [2, 1] and M = [1, 2], according to Step 4 in Algorithm 1. That is the reason that we have two
possible subspaces: (i) the subspace of all the vectors [x(t ), x(t + 1 ), y(t )] is shown in Fig. 1(c), and (ii) the subspace of all
the vectors [x(t ), y(t ), y(t + 1 )] is shown in Fig. 1(d). Similarly, for any such vector (e.g., [x(42), x(43), y(42)]), we also need
to count the number of neighbors which are within a distance r (tolerance level), illustrated by a sphere centered at [x(42),
x(43), y(42)] with radius r in Fig. 1(c). For any such vector (e.g., [x(42), y(42), y(43)]) in Fig. 1(d), we should do the similar
counting. We employ this rigorous approach to compare composite delay vectors (to find the neighbors) not only within
each subspace but also across all the subspaces, thus fully catering for both within- and cross-channel correlations.
 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90 79

Fig. 1. (a) Two high-frequency (five-minute interval) returns from SSE and SZSE separately denoted as x(t) and y(t) with the length of 100. (b) 2D-plot of
composite delay vectors [x(t), y(t)]. (c) 3D-plot of vectors [x (t ), x(t + 1 ), y(t )]. (d) 3D-plot of vectors [x(t ), y(t ), y(t + 1 )].

From Algorithm 2, we can obtain the multivariate MSE (MMSE) plot as a demonstration of the experimental results,
in which the MSampEn (as a function of scale factor λ) is utilized to evaluate the relative complexity of normalized
multichannel time series. The MMSE analysis has its own interpretations, which is outlined as follows [28]:

(i) For the majority of time scales, if the MSampEn values with multiple scales for the multivariate time series X are
larger than those for the multivariate time series Y, the signal X is more dynamically complex than the signal Y.
(ii) The multivariate time series X containing useful information at the smaller scales is indicated by a monotonic decrease
shown in the multivariate MSE plot, which is typical for completely random time series.
(iii) If the MSampEn values with multiple scales keep constant or the multivariate MSE plot show a monotonic increase,
the multivariate time series X exhibits long-range correlations, and this multivariate system shows complex generating
dynamics.

Algorithm 1. Multivariate sample entropy (MSampEn).

Step 1: For a p-variate time series {xk,i }N i=1
(k = 1, 2, . . . , p), N is the number of samples in every channel, the embedding
vector is M = [m1 , m2 , . . . , m p ] ∈ R p , the time lag vector is τ = [τ1 , τ2 , . . . , τ p ]. By forming N − νm composite delay vectors,
the multivariate embedded reconstruction [33] is defined as
Xm (i ) = [x1,i , x1,i+τ1 , . . . , x1,i+(m1 −1)τ1 , x2,i , x2,i+τ2 , . . . ,
x2,i+(m2 −1)τ2 , . . . , x p,i , x p,i+τ p , . . . , x p,i+(m p −1)τ p ], ( i = 1 , 2 , . . . , N − νm ) (2)
p
where m = k=1 mk , ν m is defined as νm = max{M} × max{τ }.
Step 2: Considering the maximum norm [34], the distance d[Xm (i), Xm (j)], between any two composite delay vectors
Xm (i) and Xm (j), is defined as
d[Xm (i ), Xm ( j )] = max {|x(i + l − 1 ) − x( j + l − 1 )|}. (3)
l=1,...,m

Step 3: For the obtained composite delay vector Xm (i) and a given threshold r, the number of instances, denoted as
Pim , is calculated for which d[Xm (i ), Xm ( j )] ≤ r (1 ≤ j ≤ N − νm , j = i ), the frequency of its occurrence is calculated as
80 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90

Bm
i
(r ) = 1 m
N−νm −1 Pi . By taking the average over all i, Bm (r) is defined as


N− νm
1
Bm ( r ) = i (r ).
Bm (4)
N − νm
i=1

Step 4: The dimension in Step 1 increases from m to m + 1. Considering the previous space with the embed-
ding vector, M = [m1 , m2 , . . . , mk , . . . , m p ] (k = 1, 2, · · · , p), the system can evolve to p different forms of space
with the new embedding vector, [m1 , m2 , . . . , mk + 1, . . . , m p ], by changing the specific variable k from 1 to p
gradually and simultaneously keeping the dimension of other variables unchanged. Supposing mk = max{M},
νmk +1 = (mk + 1 ) × max{τ } = (max{M} + 1 ) × max{τ } = νm + max{τ }; If mk = max {M}, there is νmk +1 = νm . Thus a to-
p 
tal of k=1
( N − νm k
+1
) vectors Xm+1 (i ) ∈ Rm+1 (i = 1, 2, . . . , kp=1 (N − νmk
+1
)) are obtained, and the overall embedding
dimension of the system undergoes the change from m to m + 1.
Step 5: For a given Xm+1 (i ), the number of instances, Pim+1 , is calculated for which d[Xm+1 (i ), Xm+1 ( j )] ≤ r (1 ≤ j ≤
p
k=1
( N − νmk
+1
), j = i), the frequency of its occurrence is calculated, Bm i
+1
(r ) is defined as
1
Bm +1
(r ) =  p Pim+1 . (5)
i
k=1
(N − νmk +1 ) − 1
Finally, by taking the average over all i in (m + 1 ) dimensional space, Bm+1 (r ) is defined as
p
(N−νmk +1 )
1 k=1 
Bm+1 (r ) =  p Bm +1
(r ). (6)
k=1
(N − ν k
m+1
) i=1
i

Step 6: For the tolerance level r, MSampEn is estimated as

B (r ) 
m+1
MSampEn(M, τ , r, N ) = − ln . (7)
Bm ( r )
Algorithm 2. Multivariate multiscale entropy (MMSE) Analysis.
Step 1: For a p-variate time series {xk,i }N
i=1
(k = 1, 2, . . . , p), N is the number of samples in every channel, λ is a scale
factor, the corresponding coarse-grained multivariate time series is defined as
λl

1 N
yλk,l = xk,i , (1 ≤ l ≤ , k = 1, 2, · · · , p ). (8)
λ i=λ(l−1)+1 λ

Step 2: The multivariate sample entropy MSampEn is evaluated for each coarse-grained multivariate time series yλk,l by using
steps in the Algorithm 1, and then MSampEn can be plotted as the function of scale factor λ.

2.2. Effectiveness test of MMSE analysis.

Before applying the MMSE analysis to analyze the return series from the real stock markets, to illustrate the performance
and effectiveness of this method, we generate multivariate sequences consisting of white noise (with uncorrelated fluctu-
ations) or 1/f noise (with correlated fluctuations) [3,28], these two types of data are commonly used in statistical finance
and signal analysis. When studying the statistical properties of time series, the effective length of the sequence is a very
important parameter. In order to assess the sensitivity of the MMSE method to the length of the measured multivariate time
series at multiple scales, we evaluate the multivariate sample entropy of a 3-channel white noise and a 3-channel 1/f noise,
both with different sample size (N = 50 0, 10 0 0, 20 0 0, 50 0 0, 10 0 0 0, 20 0 0 0) in each channel. The values of the parameters
used to calculate MSampEn in this section are mk = 2, τk = 1, and r = 0.20 × δ (δ is the standard deviation of data) [35] for
each data channel. And in order to improve the accuracy of the test, each case of experiments is carried out 100 times in-
dependently, and we calculate the corresponding mean and standard deviation (SD) of MSampEn values at each scale. Some
important corresponding results are shown as represented in Tables 1 and 2. The MSampEn values for the 3-channel white
noise with different data length in each channel are shown in Fig. 2(a), the error bar at each scale represents the SD of the
MSampEn values. All MSampEn curves for the 3-channel white noise of different lengths show a significant monotonically
decreasing on multiple scales. The resulting MSampEn curves are getting smoother as the length of the measured time
series increases, especially when the data length is less than 20 0 0. It is also interesting to note that when the scale factor λ
< 6, the MSampEn value increases with the length of the measured data. For a clearer observation, we show the standard
deviation as a function of the scale factor λ in Fig. 2(b). For a fixed scale factor, the standard deviation of the MSampEn
value for the 3-channel white noise decreases as the sequence length increases. For the fixed sequence length, the standard
deviation curve of MSampEn values is roughly declining, especially when the sequence length is relatively small (e.g. N <
20 0 0). For each standard deviation curve, as the sequence length increases, the fluctuation of the curve decreases gradually,
and when the sequence length is large enough (e.g. N > 20 0 0), the curve is basically stable. For the 3-channel 1/f noise, we
take the same experimental process, the corresponding experimental results are shown in Fig. 2(c) and (d). The MSampEn
results of the 3-channel 1/f noise are significantly more affected by the sequence length than the 3-channel white noise.
 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90 81

Table 1
Means and SDs of MSampEn values for 3-channel white noise with different data length (N = 50 0, 10 0 0, 20 0 0, 50 0 0, 10 0 0 0, 20 0 0 0).

Mean Scale factor(τ )

1 2 3 5 7 10 12 15 20

N = 500 1.0927 0.8747 0.7633 0.6143 0.4392 0.3420 0.3083 0.2539 0.2296
N = 10 0 0 1.2087 0.9421 0.7883 0.6168 0.4826 0.3669 0.3053 0.2560 0.1885
N = 20 0 0 1.2968 0.9969 0.8358 0.6254 0.5093 0.3718 0.3139 0.2500 0.1849
N = 50 0 0 1.3476 1.0422 0.8546 0.6331 0.5055 0.3727 0.3151 0.2438 0.1726
N = 10 0 0 0 1.3722 1.0481 0.8627 0.6437 0.5083 0.3767 0.3135 0.2495 0.1686
N = 20 0 0 0 1.3787 1.0519 0.8663 0.6438 0.5082 0.3755 0.3130 0.2449 0.1682

SD Scale factor(τ )

1 2 3 5 7 10 12 15 20

N = 500 0.1364 0.1036 0.0983 0.0512 0.0701 0.0515 0.0691 0.0758 0.0705
N = 10 0 0 0.1045 0.0479 0.0416 0.0495 0.0509 0.0336 0.0390 0.0285 0.0406
N = 20 0 0 0.0434 0.0344 0.0259 0.0317 0.0299 0.0262 0.0330 0.0277 0.0243
N = 50 0 0 0.0187 0.0143 0.0166 0.0182 0.0133 0.0160 0.0156 0.0155 0.0109
N = 10 0 0 0 0.0086 0.0087 0.0106 0.0124 0.0130 0.0141 0.0105 0.0100 0.0083
N = 20 0 0 0 0.0056 0.0061 0.0051 0.0075 0.0098 0.0100 0.0093 0.0078 0.0064

Table 2
Means and SDs of MSampEn values for 3-channel 1/f noise with different data length (N = 50 0, 10 0 0, 20 0 0, 50 0 0, 10 0 0 0, 20 0 0 0 ).

Mean Scale factor(τ )

1 2 3 5 7 10 12 15 20

N = 500 0.9483 0.7749 0.6873 0.4415 0.0140 – – – –

N = 10 0 0 1.0714 0.9238 0.8475 0.7225 0.5964 0.3948 0.4312 0.1393 0.0902
N = 20 0 0 1.1248 1.0133 0.9523 0.8466 0.7578 0.7276 0.6248 0.6239 0.5330
N = 50 0 0 1.1372 1.0702 1.0308 0.9808 0.9315 0.8735 0.8755 0.8400 0.7730
N = 10 0 0 0 1.1392 1.0744 1.0444 1.0137 0.9851 0.9422 0.9314 0.8992 0.8604
N = 20 0 0 0 1.1405 1.0810 1.0596 1.0280 1.0109 0.9842 0.9727 0.9512 0.9313

SD Scale factor(τ )

1 2 3 5 7 10 12 15 20

N = 500 0.1143 0.1610 0.3353 0.2935 0.4689 – – – –

N = 10 0 0 0.0424 0.0685 0.1003 0.1535 0.2065 0.3075 0.2434 0.4460 0.4225
N = 20 0 0 0.0312 0.0512 0.0693 0.0823 0.0982 0.0919 0.1746 0.1603 0.1760
N = 50 0 0 0.0271 0.0290 0.0331 0.0576 0.0530 0.0605 0.0694 0.0627 0.0818
N = 10 0 0 0 0.0168 0.0213 0.0243 0.0282 0.0361 0.0289 0.0271 0.0458 0.0511
N = 20 0 0 0 0.0164 0.0202 0.0246 0.0213 0.0279 0.0308 0.0314 0.0387 0.0475

Unlike the experimental results of 3-channel white noise, when the length of the 3-channel 1/f noise is not large enough,
the MSampEn value will disappear as the scale increases, for instance, the MSampEn curves with the sequence length
N = 50 0, 10 0 0. When the sequence length N > 20 0 0, the default value disappears. For a fixed scale factor λ, the MSampEn
value of the 3-channel 1/f noise increases significantly as the sequence length increases. At the same time, as the length
of the sequence increases, the overall downward trend of the MSampEn curve becomes weaker and tends to be stable
in Fig. 2(c). Different from the results of the 3-channel white noise in Fig. 2(b), there is a rising trend in the standard
deviation curve of MSampEn values for the 3-channel 1/f noise, and the rising trend is weakened and eventually stabilized
in Fig. 2(d), as the length of the sequence increases. Therefore, the length of the sequence has a great effect on the error of
the experimental results, especially when the data length is relatively small. It is clear that the increase in sequence length
can reduce the error of single experiment results to a certain extent. And when the sequence length is not large enough,
the scale factor λ also has a relative greater impact on the experimental results. Considering that the error range of the
experimental results is reduced below 0.1, the effective length of the sequence to be measured in the MMSE method should
be larger than 20 0 0. Considering the interpretations of the MMSE analysis in the last subsection, these results also show a
physical justification, that the correlated 1/f noise is structurally more complex than the uncorrelated white noise.
In order to study the effect of single channel data on the experiment results of the MMSE analysis for the multivariate
time series (in which different channels contain different noise realizations), we generate four trivariate time series. For the
first one all the data channels are realizations of mutually independent 1/f noise, then the number of variates that represent
1/f noise decreases from 3 to 0 and simultaneously the number of data channels that represent independent white noise in-
creases from 0 to 3, so that the total number of variates is always three. These trivariate time series are denoted as (3f, 0w),
(2f, 1w), (1f, 2w), (0f, 3w) in turn. Then we calculate the MSampEn values of these four trivariate time series, and each case
of experiments is still carried out 100 times independently. Some corresponding important results are shown as represents
in Table 3. Fig. 2(e) shows the MMSE curves for the cases considered; notice that as the number of variates representing 1/f
82 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90

Fig. 2. (a) MSampEn values for 3-channel white noise with different data length (N = 50 0, 10 0 0, 20 0 0, 50 0 0, 10 0 0 0, 20 0 0 0 ). (b) SDs of MSampEn values
(error bars) of white noise with multiple scales for different data length. (c) MSampEn values for 3-channel 1/f noise with different data length (N =
50 0, 10 0 0, 20 0 0, 50 0 0, 10 0 0 0, 20 0 0 0 ). (d) SDs of MSampEn values (error bars) of 1/f noise with multiple scales for different data length. (e) MMSE curves
for four multichannel time series generated by white noise and 1/f noise. (f) SDs of MSampEn values (error bars) for four generated multichannel time
series.

noises increases, the MSampEn value at higher scale factors λ increases. And when all the three data channels contain 1/f
noise, the complexity at larger scales is the highest. At scale factor λ = 1, the opposite trend occurs, so each MMSE curve is
coincident at scale factor λ = 2. Fig. 2(f) shows the standard deviations of the MSampEn values as a function of scale factor
λ for the cases considered, it is discovered that the increase of the number of variates representing 1/f noises can cause
an increase in the resulting errors, and these errors increase with the scale factor λ. The analysis in Fig. 2(e) confirms that,
as desired, the more variables or channels within a multivariate time series exhibit long-range correlations (e.g. 1/f noise),
the higher the overall complexity of the underlying multivariate system is. However, Fig. 2(f) shows that it also leads to an
increase in the error of the experimental results. Fortunately, the error range is essentially below 0.1.
 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90 83

Table 3
Means and SDs of MSampEn values for four multichannel time series generated by white noise and 1/f noise.

Mean Scale factor(τ )

1 2 3 5 7 10 12 15 20

(3f, 0w) 1.1206 1.0341 0.9953 0.9160 0.9014 0.8195 0.7884 0.7197 0.7196
(2f, 1w) 1.1952 1.0535 0.9857 0.8964 0.8402 0.7902 0.7438 0.7216 0.6615
(1f, 2w) 1.2477 1.0299 0.9023 0.7551 0.6647 0.5728 0.5245 0.4798 0.4228
(0f, 3w) 1.3112 1.0240 0.8409 0.6366 0.4990 0.3785 0.3169 0.2417 0.1716

SD Scale factor(τ )

1 2 3 5 7 10 12 15 20

(3f, 0w) 0.0293 0.0497 0.0489 0.0668 0.0640 0.0897 0.1077 0.1046 0.1652
(2f, 1w) 0.0257 0.0375 0.0393 0.0394 0.0595 0.0602 0.0704 0.0611 0.0806
(1f, 2w) 0.0256 0.0204 0.0263 0.0320 0.0310 0.0305 0.0249 0.0385 0.0450
(0f, 3w) 0.0325 0.0186 0.0257 0.0231 0.0141 0.0212 0.0189 0.0128 0.0171

Fig. 3. (a) Schematic diagrams for illustrating the grouping process of 12 groups of five-minute returns. (b) The X1 of SSE and SZSE and the boxplots of all
12 groups of five-minute returns for SSE and SZSE.

3. Nonlinear complexity for financial time series

3.1. Complexity analysis of china stock market

In this section, we apply the MMSE analysis and validate its utility in studying the complexity of financial time series.
Inspired by Ref. [3], we select the high-frequency (five-minute interval) returns of Shanghai Stock Exchange (SSE) and
Shenzhen Stock Exchange (SZSE) from November 1, 2013 to February 18, 2016 with 26784 data points (five-minute returns).
These two classical stock markets in China trade 4 hours a day, which is from 9:35 a.m. to 11:30 a.m., and then from 1:00
p.m. to 3:00 p.m., which means that we can get 48 high-frequency (five-minute interval) data every trading day. We take
a 20-min interval in each day, and divide the 48 five-minute returns into 12 groups, each group has 4 five-minute returns.
Then we connect each group of data with the corresponding set of data on the same time of the next day, finally generate
12 groups of five-minute return series, denoted as X1 , X2 , , X12 , the length of each group is 2232. In order to visualize the
grouping process, we use the schematic diagram in Fig. 3(a) for a better illustration. Dk1 in pink is the trading day, xk2 ,k3 in
black is a five-minute return value in the trading day, Xk4 in blue is a group of five-minute returns. The colored boxs clearly
show the formation of the grouping process. In order to observe the numerical values of each group of five-minute returns,
we take the X1 of SSE and SZSE as represents and show these high-frequency returns in Fig. 3(b). And the boxplots of each
group of five-minute returns from the SSE and the SZSE are displayed in Fig. 3(b) correspondingly. These boxplots show
simple statistical properties, and clearly show the similarity (e.g. the mean of the data) and differences (e.g. the range of the
data) between the groups of the five-minute returns. Then we generate 4 trivariate time series denoted as Y1 = [X1 , X2 , X3 ]
for the 1st hour, Y2 = [X4 , X5 , X6 ] for the 2nd hour, Y3 = [X7 , X8 , X9 ] for the 3rd hour, Y4 = [X10 , X11 , X12 ] for the 4th hour.
In order to study the complexity of stock markets in different time periods, we apply the MMSE analysis to study
these four trivariate time series (Y1 , Y2 , Y3 , Y4 ) from the SSE and the SZSE, this series represents the stock time series
for each hour respectively. The MMSE values are displayed in Fig. 4, in which the isometric subgraphs are adopted for a
better obversion. Combining the interpretations of the MMSE analysis with the subgraphs in Fig. 4 and the MMSE results
84 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90

Fig. 4. MMSE results of trivariate time series (Y1 , Y2 , Y3 , Y4 ) from SSE and SZSE.

Table 4
MSampEn values of trivariate time series (Y1 , Y2 , Y3 , Y4 ) from SSE and SZSE.

Scale factor(τ )

1 2 3 5 7 10 12 15 20

Data: SSE
Y1 0.5766 0.4007 0.2806 0.1760 0.1567 0.1207 0.0938 0.0706 0.0638
Y2 0.5297 0.4104 0.2928 0.2069 0.1646 0.1306 0.1210 0.0902 0.0699
Y3 0.4428 0.3442 0.2548 0.1632 0.1373 0.1243 0.1136 0.0935 0.0959
Y4 0.3935 0.2686 0.2078 0.1466 0.1029 0.1067 0.0992 0.0872 0.0772
Mean 0.4856 0.3560 0.2590 0.1732 0.1404 0.1206 0.1069 0.0854 0.0767
SD 0.0717 0.0564 0.0326 0.0221 0.0238 0.0087 0.0109 0.0088 0.0120
Data: SZSE
Y1 0.7528 0.5456 0.3935 0.2463 0.2149 0.1394 0.1038 0.0891 0.0754
Y2 0.6917 0.5508 0.4003 0.2643 0.2025 0.1613 0.1446 0.0919 0.0701
Y3 0.5658 0.4206 0.3092 0.2090 0.1465 0.1365 0.1185 0.1067 0.1047
Y4 0.4366 0.3351 0.2668 0.1740 0.1306 0.1171 0.0956 0.0925 0.0762
Mean 0.6118 0.4630 0.3424 0.2234 0.1736 0.1386 0.1156 0.0950 0.0816
SD 0.1215 0.0904 0.0565 0.0348 0.0358 0.0156 0.0186 0.0069 0.0135

in Table 4, we find that all of MMSE curves show monotonic decreases, which means that the multivariate return series
contain useful information at the smaller scales. For all of the time scales (especially for smaller scales), the MSampEn
values of four trivariate time series from the SZSE are correspondingly larger than those from the SSE, which shows that
the stock returns in each hour from the SZSE are more dynamically complex than that from the SSE. In Fig. 4, for both
the SZSE and the SSE, the MMSE values of Y1 , Y2 , Y3 and Y4 progressively decrease in the smaller scales (when the scale
factor τ ∈ [1, 5]), and all gradually tend to be less than 0.1 in the larger scales (when the scale factor τ ∈ [15, 20]). And the
error bars (SD) of the mean lines in the last subgraph show that the differences are gradually decreasing with the increases
of the scale factor. It is also clearly shown in Table 4, the MSampEn values of Y1 , Y2 , Y3 and Y4 from the SSE at τ = 1
are 0.5766, 0.5297, 0.4428, 0.3935 respectively, and the MSampEn values of Y1 , Y2 , Y3 and Y4 from the SZSE at τ = 1 are
0.7528, 0.6917, 0.5658, 0.4366 respectively. These results indicate that the complexity of stock market returns in each hour
are different and show a significant decreasing trend with the stock trading time progressing. The system complexity of the
stock market in the forenoon is significantly higher than that in the afternoon.
To study the influence of the temporal order of stock return series on the complexity of financial systems, we perform
the MMSE analysis on two sets of shuffled multivariate data, which are separately chosen from the stock returns of SSE
and SZSE and randomly shuffled with respect to their temporal order [3,36]. The relevant results are displayed in Fig. 5 and
Table 5, compared with the original stock returns. The error bars in the figures represent the standard deviations (SDs) of
the MMSE values by analyze 100 independent samples of shuffled multivariate data in each case. In Fig. 5 and Table 5, it is
clearly shown that, as the time scale τ increases, the MMSE values derived from these two sets of shuffled multivariate time
 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90 85

Fig. 5. (a) MMSE results of shuffled multivariate returns compared with the original multivariate returns from SSE. (b) MMSE results of shuffled multivari-
ate returns compared with the original multivariate returns from SZSE.

Table 5
Means and SDs of MSampEn values of shuffled multivariate returns from SSE and SZSE.

Scale factor(τ )

1 2 3 5 7 10 12 15 20

Data: Shuffled SSE

Mean(S-Y1 ) 0.9914 0.8112 0.6903 0.5325 0.4468 0.3398 0.2965 0.2473 0.1828
Mean(S-Y2 ) 1.0235 0.8450 0.7275 0.5652 0.4504 0.3473 0.2950 0.2327 0.1679
Mean(S-Y3 ) 0.9846 0.8205 0.6970 0.5472 0.4346 0.3292 0.2821 0.2287 0.1680
Mean(S-Y4 ) 0.9482 0.7916 0.6898 0.5457 0.4493 0.3460 0.2946 0.2396 0.1728
SD(S-Y1 ) 0.0167 0.0205 0.0211 0.0262 0.0233 0.0251 0.0176 0.0307 0.0307
SD(S-Y2 ) 0.0208 0.0277 0.0197 0.0325 0.0247 0.0204 0.0262 0.0217 0.0163
SD(S-Y3 ) 0.0199 0.0175 0.0235 0.0222 0.0168 0.0192 0.0282 0.0230 0.0171
SD(S-Y4 ) 0.0174 0.0228 0.0237 0.0254 0.0253 0.0172 0.0252 0.0191 0.0250
Data: Shuffled SZSE
Mean(S-Y1 ) 1.0497 0.8289 0.6940 0.5381 0.4347 0.3351 0.2800 0.2293 0.1687
Mean(S-Y2 ) 1.0854 0.8646 0.7253 0.5533 0.4449 0.3450 0.2986 0.2384 0.1693
Mean(S-Y3 ) 1.0187 0.8163 0.6997 0.5378 0.4340 0.3291 0.2851 0.2244 0.1637
Mean(S-Y4 ) 0.9405 0.7881 0.6849 0.5447 0.4559 0.3431 0.2904 0.2422 0.1774
SD(S-Y1 ) 0.0191 0.0244 0.0270 0.0190 0.0244 0.0282 0.0221 0.0229 0.0209
SD(S-Y2 ) 0.0215 0.0219 0.0197 0.0228 0.0222 0.0238 0.0226 0.0176 0.0161
SD(S-Y3 ) 0.0189 0.0232 0.0173 0.0224 0.0189 0.0197 0.0202 0.0171 0.0168
SD(S-Y4 ) 0.0241 0.0193 0.0278 0.0209 0.0199 0.0219 0.0216 0.0207 0.0202

series decrease more severely than the corresponding original data. At each scale τ , the MMSE values of the shuffled data
are all larger than those of the original data, which means that the random shuffle of the multivariate stock return increases
its complexity. Different from the clear distinctions of the MMSE curves of the original data, the MMSE curves of the
shuffled data are almost coincide and show the monotonic decreases more severely than those of the original data, which
are qualitatively similar to the curves of the multivariate white noise time series already shown in Fig. 2(a). At each scale
τ , the MMSE values of the shuffled data are all larger than those of the original data, for each scale the SDs of the MMSE
values are all smaller than 0.05. These results show that the random shuffle of the stock returns increases the complexity,
and the temporal order of the stock return series may have great influence on the complexity of financial systems.
Considering the widely acknowledged fact that the absolute return [37–39] can be used as a proxy of the volatility of
stock time series, we also make the MMSE analysis on two sets of absolute multivariate returns separately chosen from the
aforementioned stock returns. The related experimental results are clearly shown in Fig. 6 and Table 6. Compared with the
original data, the results of the absolute multivariate data show obviously differences, on the smaller scales their MMSE
values become lower, conversely, on the larger scales the MMSE values become higher, and naturally the slopes of their
MMSE curves get smoother. The two sets of absolute multivariate returns from the SSE and the SZSE show the experimental
results in similar pattern. The MMSE curves exhibit a tendency to close to the level, considering the interpretations of the
MMSE analysis, the absolute multivariate time series tend to exhibit long-range correlations.

3.2. Complexity analysis of global stock markets

It has been proved that the multivariate MSE method gets more comprehensive and systematic analysis on the complex
system by performing the analysis of multivariate data at the same time. In this section, we use the MMSE method to make a
86 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90

Fig. 6. (a) MMSE results of absolute multivariate returns compared with the original multivariate returns from SSE. (b) MMSE results of absolute multi-
variate returns compared with the original multivariate returns from SZSE.

Table 6
MSampEn values of absolute multivariate returns from SSE and SZSE.

Scale factor(τ )

1 2 3 5 7 10 12 15 20

Data: Absolute SSE

|Y1 | 0.4413 0.2829 0.2096 0.1637 0.1431 0.1198 0.1165 0.0949 0.1124
|Y2 | 0.3863 0.2861 0.2375 0.1745 0.1478 0.1638 0.1600 0.1461 0.1666
|Y3 | 0.3347 0.2233 0.1690 0.1335 0.1360 0.1349 0.1349 0.1402 0.1392
|Y4 | 0.2853 0.1903 0.1492 0.1226 0.1014 0.1252 0.0942 0.1142 0.1107
Mean(a) 0.3619 0.2456 0.1913 0.1486 0.1321 0.1359 0.1264 0.1238 0.1322
SD(a) 0.0581 0.0406 0.0344 0.0212 0.0182 0.0170 0.0242 0.0206 0.0229
Data: Absolute SZSE
|Y 1 | 0.6029 0.3870 0.3021 0.2058 0.1815 0.1352 0.1367 0.1096 0.1440
|Y2 | 0.5412 0.3803 0.3065 0.2338 0.2093 0.1790 0.1766 0.1546 0.1556
|Y3 | 0.4235 0.2900 0.2164 0.1584 0.1470 0.1284 0.1266 0.1279 0.1154
|Y4 | 0.3405 0.2302 0.1941 0.1450 0.1230 0.1507 0.1340 0.1236 0.1174
Mean(a) 0.4770 0.3219 0.2548 0.1857 0.1652 0.1483 0.1435 0.1289 0.1331
SD(a) 0.1018 0.0653 0.0502 0.0357 0.0329 0.0195 0.0195 0.0163 0.0172

new approach to analyze the complexity of stock markets for three different regions - Asia, Europe and America, and to clar-
ify the difference of them by studying the daily logarithmic returns of stock indices. For the Asia, we choose the HSI (Hang
Seng Index in China), the KOSPI (Korea Composite Stock Price Index in Korea) and the Nikkei (Nikkei Index in Japan). For the
Europe, we choose the DAX (DAX Index in Germany), the CAC (CAC Index in France) and the FTSE (Financial Times and Stock
Exchange in England). For the America, we choose the S & P (Standard & Poor’s Index), the NASDAQ (NASDAQ Index) and
the DJI (Dow Jones Index). The daily closing prices of them are all from the same time bucket, which is from January 2, 1991
to July 1, 2016, with about 60 0 0 data points (Even though the appropriate adjustments and deletion on the data are carried
out for different non-trading time among these markets, there are still some slight differences). Then, we obtain 9 daily
logarithmic returns and recombine them to build three multivariate time series separately for Asia, Europe and America.
Considering the algorithm of MMSE analysis, this method still work and turn to be the MSE method if a single variate is
input into this method [28,29]. So we first do the MMSE analysis on each univariate return (HSI, KOSPI, Nikkei), then on the
corresponding multivariate return (Asia), the results are shown in Fig. 7(a). The data from Europe and America are analyzed
with the same process, and their results are shown in Fig. 7(b) and (c) respectively. From all the figures in Fig. 7, it can
be easily discovered that all of MMSE plots of the returns show the same trend that they decrease with the scale factor
λ, which means that all of the returns contain the useful information at the smaller scales. Only the differences between
them are the value ranges (degrees of complexity displayed with the histograms of different heights distinctly) to make it
more convenient to show their differences. Observed the MMSE results in smaller scales, there is an intriguing finding that,
different from Asia in Fig. 7(a) and Europe in Fig. 7(b) (in which the MMSE values of multivariate returns are both lower
than all of the corresponding univariate returns), the MMSE values of multivariate returns of America in Fig. 7(c) are higher
than the corresponding univariate returns. To make a more clear comparison among these three regions, the MMSE results
 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90 87

Fig. 7. (a) MMSE results of each univariate return (HSI, KOSPI, Nikkei) and the corresponding multivariate return of Asia. (b) MMSE results of each univariate
return (DAX, CAC, FTSE) and the corresponding multivariate return of Europe. (c) MMSE results of each univariate return (S & P, NASDAQ, DJI) and the
corresponding multivariate return of America. (d) MMSE results of multivariate returns of Asia, Europe and America for comparison.

of their multivariate returns are exhibited in Fig. 7(d), which presents that the multivariate returns from America show the
highest degree of complexity, followed by those from Asia and then Europe with smaller scales.
The absolute returns with different power exponents α (i.e., |R(t)|α ) have been shown to carry varying levels of volatility
behaviors [37]. In Section 3.1, the MMSE curves of absolute multivariate time series show the smoother slopes and exhibit
a tendency to close to the level. In the following, the absolute multivariate returns from three regions are analyzed with
different power exponents α , in order to find out the relationship between the levels of volatility behaviors and the slope of
MMSE curves. The three absolute multivariate returns from Asia, Europe and America with power exponent α are separately
denoted as Asia − |MR|α , Europe − |MR|α and America − |MR|α . It is a remarkable fact that the multivariate time series
do not necessarily have the same amplitude range among its data channels. In order to avoid the bias of the distances
calculated on embedded vectors toward the variable with largest amplitude ranges, we propose to normalize all the data
channels to the same amplitude range [0, 1] as a prior choice for the operation of the MMSE algorithm. In this section,
we take a sequence of α as {0.25, 0.5, 0.75, 1, 1.25, 1.5, 1.75, 2}. To make a clear show of the normalized multivariate
returns with different α , we choose the normalized multivariate return series from the Nikkei for example, labeled as
|RNi |α , considering that its original multivariate returns have the highest MSampEn values in all of the nine returns shown
in Fig. 7(a)–(c). Fig. 8 presents the normalized multivariate returns |RNi |α (α = 0.25, 0.5, 0.75, 1, 1.25, 1.5, 1.75, 2) compared
with the normalized original multivariate returns of the Nikkei. It is observed that, with the same variation 1, these
normalized |RNi |α have different fluctuation ranges, and the ranges get bigger as the power exponent α increases.
Further, we normalize each channel of Asia − |MR|α , Europe − |MR|α and America − |MR|α (α =
0.25, 0.5, 0.75, 1, 1.25, 1.5, 1.75, 2), and apply the MMSE analysis on them to calculate their complexity. The MMSE results
for Asia − |MR|α , Europe − |MR|α and America − |MR|α are separately presented in Fig. 9(a)–(c). From these figures it can be
observed that for Asia − |MR|α , Europe − |MR|α and America − |MR|α , the MSampEn curves with a fixed α totally decrease
88 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90

Fig. 8. Plots of normalized multivariate returns |RNi |α (α = 0.25, 0.5, 0.75, 1, 1.25, 1.5, 1.75, 2) compared with the normalized original multivariate returns
of Nikkei.

Fig. 9. (a) MMSE results for Asia − |MR|α . (b) MMSE results for Europe − |MR|α . (c) MMSE results for America − |MR|α . Where α=
0.25, 0.5, 0.75, 1, 1.25, 1.5, 1.75, 2.

gradually and finally almost reach the level as the scale factor λ increases. Morever, with the gradual increase of the power
exponent α , the MMSE curves reach the constant more quickly. Finally when α = 2, the MMSE curves almost keep level,
which can be conveniently observed through the color gradient of the three-dimensional graph and the directed arrow in
Fig. 9. These findings indicate that each of Asia − |MR|α , Europe − |MR|α and America − |MR|α with different levels of volatil-
ity behaviors, unlike the original multivariate returns, exhibits long-range correlations and dynamically complex structures
across multiple time scales in different degrees. The MSampEn behaviors for the multivariate volatility time series show a
 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90 89

great similar trend to that of multivariate 1/f noise series, which has the strongly correlated fluctuations and its multivariate
system shows complex generating dynamics with different degrees. In addition, it is noteworthy that the MSampEn values
of the multivariate volatility time series (Asia − |MR|α , Europe − |MR|α and America − |MR|α ), with the time scale factor
getting smaller step by step and as the power exponent α gets bigger gradually, show a decreasing dynamical complexity.

4. Conclusion

In this present paper, for the first time, the multivariate multiscale entropy (MMSE) analysis is introduced into the
research of financial stock markets. First, we confirm that, for the MMSE method, each channel length of multivariate time
series should be larger than 20 0 0 to quantify their complexity, and the more channels within a multivariate time series
exhibit long-range correlations, the higher the overall complexity of the underlying multivariate system is. After that, for
a new trial, we apply the MMSE method to quantify the complexity of four generated trivariate time series for each stock
trading hour in China stock markets, the SSE and the SZSE. The results illustrate three new findings: the multivariate return
series from the China stock markets contain useful information at the smaller scales, the stock returns in each hour from
the SZSE are more dynamically complex than that from the SSE, and the complexity of stock returns in each hour shows a
significant decreasing trend with the stock trading time progressing. Further, the MMSE analysis on the shuffled multivariate
return series indicates that the random shuffle of the stock return series increases its complexity, and the temporal order of
the stock return series may have great influence on the complexity of the financial system. And the analysis of the absolute
multivariate time series shows a tend to exhibit long-range correlations. Moreover, there is another new attempt, we utilize
the MMSE method to analyze the complexity of stock markets for three major regions - Asia, Europe and America. The
empirical results show that the multivariate stock returns from America show the highest degree of complexity, followed by
those from Asia and then Europe with smaller scales. Finally, we wield the MMSE analysis to assess the relative complexity
of normalized multivariate return volatility series with the different levels of volatility behavior. It has been identified
that, unlike the original multivariate returns, all of the normalized multivariate return volatility series exhibit long-range
correlations and dynamically complex structures across multiple time scales. It is also noteworthy that the MSampEn values
of the multivariate volatility time series (Asia − |MR|α , Europe − |MR|α and America − |MR|α ), with the time scale factor
getting smaller and the power exponent α getting bigger gradually, show a decreasing dynamical complexity.

Acknowledgment

The authors were supported in part by National Natural Science Foundation of China Grant No. 71271026.

References

[1] Gabaix X, Gopikrishnan P, Plerou V, Stanley HE. A theory of power-law distributions in financial market fluctuations. Nature 2003;423:267–70.
[2] Hong WJ, Wang J. Multiscale behavior of financial time series model from potts dynamic system. Nonlinear Dyn 2014;78:1065–77.
[3] Niu HL, Wang J. Quantifying complexity of financial short-term time series by composite multiscale entropy measure. Commun Nonlinear Sci Numer
Simul 2015;22:375–82.
[4] Costa M, Goldberger AL, Peng CK. Multiscale entropy analysis of complex physiologic time series. Phys Rev Lett 2002;89(6):068102.
[5] Costa M, Goldberger AL, Peng CK. Multiscale entropy analysis of biological signals. Phys Rev E 2005;71:021906.
[6] Humeau A, Mahe G, Chapeau-Blondeau F, Rousseau D, Abraham P. Multiscale analysis of microvascular blood flow: a multiscale entropy study of laser
doppler flowmetry time series. IEEE Trans Biomed Eng 2011;58:2970–3.
[7] Gao ZK, Ding MS, Geng H, Jin ND. Multivariate multiscale entropy analysis of horizontal oil-water two-phase flow. Physica A 2015;417:7–17.
[8] Xu BB, Wang FF, Chen DY, Zhang H. Hamiltonian modeling of multi-hydro-turbine governing systems with sharing common penstock and dynamic
analyses under shock load. Energy Convers Manage 2016;108:478–87.
[9] Xu BB, Chen DY, Zhang H, Zhou R. Dynamic analysis and modeling of a novel fractional-order hydro-turbine-generator unit. Nonlinear Dyn
2015;81:1263–74.
[10] Zhang H, Chen DY, Xu BB, Wang FF. Nonlinear modeling and dynamic analysis of hydro-turbine governing system in the process of load rejection
transient. Energy Convers Manage 2015;90:128–37.
[11] Costa M, Peng CK, Goldberger AL, Hausdorff JM. Multiscale entropy analysis of human gait dynamics. Physica A 2003;330:53–60.
[12] Drozdz S, Forczek M, Kwapien J, Oswiecimka P, Rak R. Stock market return distributions: from past to present. Physica A 2007;383:59–64.
[13] Drozdz S, Kwapien J, Gruemmer F, Ruf F, Speth J. Are the contemporary financial fluctuations sooner converging to normal? Acta Phys Pol B
2003;34:4293–306.
[14] Lux T. Financial power laws: empirical evidence. Models and Mechanisms. Cambridge: Cambridge University Press; 2008.
[15] Oh G, Kim S, Eom C. Long-term memory and volatility clustering in high-frequency price changes. Physica A 2008;387:1247–54.
[16] Lo AW. Long-term memory in stock market prices. Econometrica 1991;59:1279–313.
[17] Plerou V, Gopikrishnan P, Rosenow B, Amaral L, Stanley HE. Econophysics: financial time series from a statistical physics point of view. Physica A
20 0 0;279:443–56.
[18] Mantegna RN, Stanley HE. An Introduction To Econophysics: Correlations And Complexity In Finance. Cambridge: Cambridge University Press; 20 0 0.
[19] Drozdz S, Kwapien J, Oswiecimka P, Rak R. Quantitative features of multifractal subtleties in time series. Europhys Lett 20 09;88:60 0 03.
[20] Fang W, Wang J. Statistical properties and multifractal behaviors of market returns by ising dynamic systems. Int J Mod Phys C 2012;23:1250023.
[21] Lopes AM, Machado JAT. Analysis of temperature time-series: embedding dynamics into the MDS method. Commun Nonlinear Sci Numer Simulat
2014;19:851–71.
[22] Machado JAT, Duarte FB, Duarte GM. Analysis of stock market indices through multidimensional scaling. Commun Nonlinear Sci Numer Simulat
2011;16:4610–18.
[23] Kantz H, Schreiber T. Nonlinear Time Series Analysis. Cambridge: Cambridge University Press; 1997.
[24] Lu YF, Wang J, Niu HL. Nonlinear multi-analysis of agent-based financial market dynamics by epidemic system. Chaos 2015;25:103103.
[25] Machado JAT. Complex dynamics of financial indices. Nonlinear Dyn 2013;74:287–96.
[26] Wu SD, Wu PH, Wu CW, Ding JJ, Wang CC. Bearing fault diagnosis based on multiscale permutation entropy and support vector machine. Entropy
2012;14(8):1343–56.
90 Y. Lu, J. Wang / Commun Nonlinear Sci Numer Simulat 52 (2017) 77–90

[27] Bisi MC, Stagni R. Complexity of human gait pattern at different ages assessed using multiscale entropy: from development to decline. Gait Posture
2016;47:37–42.
[28] Ahmed MU, Mandic DP. Multivariate multiscale entropy: a tool for complexity analysis of multichannel data. Phys Rev E 2011;84:3067–76.
[29] Ahmed MU, Mandic DP. Multivariate multiscale entropy analysis. IEEE Signal Process Lett 2012;19:91–4.
[30] Lamberton D, Labeyre B. Introduction to stochastic calculus applied to finance. Chapman and Hall/CRC; 20 0 0.
[31] Ross SM. An Introduction To Mathematical Finance. Cambridge: Cambridge University Press; 1999.
[32] Yu Y, Wang J. Lattice oriented percolation system applied to volatility behavior of stock market. J Appl Stat 2012;39(4):785–97.
[33] Cao L, Mees A, Judd K. Dynamic from multivariate time series. physica d. Nonlinear Phenomena 1998;121(1-2):75–88.
[34] Grassberger P, Schreiber T, Schaffrath C. Nonlinear times sequence analysis. Int J Bifur Chaos 1991;1:521.
[35] Lake DE, Richmann JS, Griffin MP, Moorman JR. Sample entropy analysis of neonatal heart rate variability. Am J Physiol Heart Circ Physiol
2002;283:789–97.
[36] Leontitsis A. A note on shuffled financial surrogates. Math Comput Model 2003;38:33–40.
[37] Niu HL, Wang J. Volatility clustering and long memory of financial time series and financial price model. Digit Signal Process 2013;23:489–98.
[38] Wang F, Yamasaki K, Havlin S, Stanley HE. Scaling and memory of intraday volatility return intervals in stock markets. Phys Rev E 2006;73:026117.
[39] Yang G, Wang J, Deng W. Nonlinear analysis of volatility duration financial series model by stochastic interacting dynamic system. Nonlinear Dyn
2015;80:701–13.