Complex and Composite Entropy Fluctuation Behaviors of Statistical Physics Interacting Financial Model

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Physica A 517 (2019) 97–113

Contents lists available at ScienceDirect

Physica A
journal homepage: www.elsevier.com/locate/physa

Complex and composite entropy fluctuation behaviors of


statistical physics interacting financial model

Guochao Wang , Shenzhou Zheng, Jun Wang
School of Science, Beijing Jiaotong University, Beijing 100044, PR China

highlights

• A novel complex stochastic interacting financial price model is introduced.


• Stylized facts, complexity behavior and multifractal property of the model are investigated.
• New entropy-based approach called composite distance fuzzy entropy is established.
• C-FuzzyEn has higher estimation accuracy and stronger robustness than FuzzyEn.
• Empirical results show the feasibility of the proposed model.

article info a b s t r a c t

Article history: In an attempt to investigate and reproduce the fluctuation dynamics of price changes in
Received 12 July 2018 financial markets, a novel complex random interacting financial price model is developed
Received in revised form 25 September 2018 by combining voter interacting system with compound Poisson process. In this model, the
Available online 9 November 2018
dissemination of investors’ investment attitudes can simulate normal and frequent small
Keywords: price fluctuations of financial markets, while the random drastic jumps can capture undue
Complex interacting financial model and rare large price fluctuations as results of sudden economic events and political events
Voter interacting system in the markets. In order to verify the rationality of the model, some statistical characteristics
Compound Poisson process of the model returns including fat-tail, power-law scaling, complexity and multifractal are
Complexity analysis compared with real returns by probability density functions, composite distance fuzzy
Composite distance fuzzy entropy
entropy (C-FuzzyEn), matching energy and multifractal detrended fluctuation analysis.
Matching energy
The C-FuzzyEn is a new entropy-based approach, based on fuzzy entropy and complexity-
invariant, to measure the complexity behaviors of return series. The effectiveness analysis
of C-FuzzyEn indicates that it has a higher estimation accuracy and stronger robustness
than fuzzy entropy. The empirical results show that the price model is able to reproduce
some important statistical properties of financial markets to a certain extent, and the
complexity of returns for the proposed model increases when infection intensity or jump
intensity increases.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction

Financial market can be regarded as a complex and dynamical evolving system filled with many interacting influence
agents and factors, of which the fluctuation behavior is one of the most important character to depict the strong dynamics
and measure the risks of financial markets. Recently, the fluctuation behavior of financial markets has become increasingly

∗ Corresponding author.
E-mail address: wangguochao@bjtu.edu.cn (G. Wang).

https://doi.org/10.1016/j.physa.2018.11.014
0378-4371/© 2018 Elsevier B.V. All rights reserved.
98 G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113

popular. Among them, constructing financial price model has played an important role to better understand the mechanism
of price fluctuation behavior or reproduce some quite nontrivial statistical properties of financial markets called stylized
empirical facts, such as fat tails property, power-law scaling behavior, absence of autocorrelation [1–18]. For example,
Ghysels et al. [2] introduced a new stochastic volatility duration (SVD) model based on the standard exponential model
with gamma heterogeneity and its factor representation to study the liquidity on financial markets and risk on car insurance
contracts. Kostanjčar et al. [14] developed a network model based on complex network science and cooperative game theory
to generate trading prices and test the existence of tipping points in financial markets, which was proved the applicability
by S&P500 index. Oświecimka et al. [19] developed an agent-based computational economical toy model for the emergence
of money, which can reveal emergence and collapse of money and the related competition effects. In this process, some
scholars find that there exist some similarities between financial markets and statistical physics systems, they all are made
up of numerous interacting ‘‘agents’’ or particles, and the complicated interaction among these agents or particles leads to
a complex system with nonlinear properties. Therefore, a large number of complex stochastic interacting financial price
models have been developed based on various theories of statistical physics systems, including contact system, exclusion
process, Potts dynamic system and Ising dynamic system [20–35]. For instance, a financial interacting model is constructed
on the basis of contact system, where interaction among investors determined by the mechanism of the contact system leads
to the stock price fluctuation [20]. A stochastic interacting epidemic system is applied to develop financial stock price model,
where the dispersal of the information and the transformation of investors’ attitudes in financial markets are mimicked
by contagion of virus and influence of epidemic individuals in the epidemic process respectively [36]. For the financial
interacting Ising price model, a group of investors holding the same attitude can be simulated by a cluster of parallel spins
in Ising dynamic systems [37]. In the above financial price models, interactions among investors for different models are
mimicked from different aspects and perspectives, but they can reproduce and explain the majority of stylized facts existed
in financial markets.
The research on fluctuation properties of financial markets is also an important aspect in understanding the price
fluctuation behaviors, such as long-term correlation, multifractal property, complexity behavior and chaotic property
[38–53], where the complexity behavior analysis is one of most significant methods to measure the inherent fluctuation
characteristic of financial time series. Therefore, plenty of approaches have been developed to quantify the complexity
behaviors, of which entropy-based approach is the most one common kind, such as permutation entropy [54], sample
entropy [55], fuzzy entropy [56,57]. These entropy-based approaches have been employed in various fields to measure
the complexity behaviors of corresponding time series, such as physical, physiology, mechanical engineering, biology and
financial markets [58]. Moreover, there exist a variety of other complexity measures approaches except the entropy-based
approach. For example, the Lempel–Ziv complexity (LZC) method is introduced to indicate the rate of the gradual buildup of
new patterns in time series [59]. A Shiner–Davison–Landsberg (SDL) method has been developed to measure the complexity
based on notions of order and disorder [60]. It is worth noting that each approach has both advantages applicable to different
complex systems or fields, which requires people to make the appropriate choice according to requirements of complexity
measure. In fact, there is no a single definition to be accepted as a global unique definition for the complexity, and it is
difficult to find a unique and mathematically rigorous method to exclusively measure the complexity behaviors of complex
systems [61]. In this paper, the ‘‘complexity’’ can be viewed as irregularity and randomness of a system’s intrinsic structure.
We generally think that the larger the complexity is, the stronger the irregularity and randomness the time series has, and it
is more unpredictable. Furthermore, multifractal property analysis is also a significant aspect of financial time series, because
the multifractal property can grasp the most essential characteristics of financial markets. Multifractal detrended fluctuation
analysis (MF-DFA), as a generalization of standard detrended fluctuation analysis (DFA), is an important method to study
the multifractal property of time series by using a variable moment [19,62].
In this paper, a novel complex stochastic interacting price model, combining voter system and compound Poisson process,
is developed to describe the complex dynamics mechanism of price fluctuation in financial markets. We assume that the
prices fluctuation in financial markets is mainly determined by two parts. The first part is normal and frequent small price
fluctuation, which is mainly determined by the interaction effect of investors. In this model, the voter interacting system
is applied to simulate the change of investors’ attitudes towards the financial markets. The investors can influence each
other and change their investment attitudes with the spread of markets information, which is assumed as the main factor
of normal and frequent small price. The second part is the rare large and discontinuous price fluctuation. It is worth noting
that financial markets sometimes will have sudden and discontinuous fluctuation because of sudden economic events and
political events [33,63]. Therefore, considering the random jumps in financial modeling has become increasingly popular
in reproducing the price fluctuation behaviors. In this model, the compound Poisson process is used to generate random
drastic jumps, which can capture undue and rare large price fluctuation. For simplicity, let above two price fluctuations
be independent. The return series data from the proposed model are selected to compare with real market indexes data,
including Shanghai Stock Exchange (SSE) Composite Index, Shenzhen Stock Exchange (SZSE) Component Index and Hushen
300 (HS300) index, Dow Jones Industrial Average (DJIA), and Gaussian noise series with same sample length, which can
show the reasonable of stochastic interacting price model. Firstly, empirical probability density function analysis is used
to investigate some stylized facts of all returns series. Moreover, new entropy-based approach called composite distance
fuzzy entropy (C-FuzzyEn) is proposed based on fuzzy entropy (FuzzyEn) method and complexity-invariant distance (CID) to
investigate the complexity behaviors of returns series and absolute returns series. The accuracy test and robustness to noise
are investigated in order to verify the effectiveness of C-FuzzyEn method compared with FuzzyEn method. Furthermore, a
G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113 99

new approach of complexity measure named matching energy (ME) is applied to investigate the complexity behaviors of
returns and corresponding absolute returns. In addition, the multifractal properties of above returns are also investigated
by multifractal detrended fluctuation analysis (MF-DFA) method. The empirical results confirm the proposed model can
simulate the financial markets in terms of some important fluctuation properties to a certain extent.
In general, there are two main innovations in this paper. One is the modeling of a novel complex stochastic interacting
financial price model based on voter interacting system and compound Poisson process. The feasibility of the proposed model
is verified by comparing some statistical characteristics of returns series with the real markets, such as fat-tail property,
power-law scaling behavior, complexity behavior and multifractal property. The other is the presentation of C-FuzzyEn
method by FuzzyEn and CID method, which can better investigate the complexity behaviors of price fluctuations for the
proposed model and the real markets as a novel complexity analysis method. The organization of this paper is as follows.
Section 2 briefly introduces the proposed complex stochastic interacting financial price model. Section 3 studies the fat-tail
and power-law scaling behaviors of returns. Section 4 introduces the new efficiency analysis of C-FuzzyEn method, and
studies the complexity behaviors of returns and absolute returns. Section 5 investigates the complexity behaviors of returns
and absolute returns by matching energy method. Section 6 studies the multifractal properties of returns by MF-DFA method.
Finally, we summarize the empirical results of this paper in Section 7.

2. Stochastic interacting financial price model

In this section, a novel complex financial stochastic interacting price model is developed by voter interacting system
and compound Poisson process, where the dissemination of investors’ investment attitudes according to the dynamical
mechanism of voter model can simulate frequent small price fluctuations, and random drastic jumps derived from compound
Poisson process can capture rare large price fluctuations. Next, we will provide a brief introduction to the evolution
mechanism of the above model.

2.1. Voter interacting system

As one of statistical physics systems, voter interacting system is used to describe the change mechanism of the voters’
behaviors towards a variety of information [20,21,23,28]. Considering the voters located at sites of d-dimensional space Zd ,
each voter holds the supportive attitude (‘‘1") or opposing attitude (‘‘0") towards some important information at independent
exponential times, and the attitudes of every voter will be influenced by their neighbors (or adjacent voters). Let ηs be a set
of voters who hold supportive attitude, where ηs (x) ∈ {0, 1} for x ∈ Zd . The voter stochastic dynamics is a Markov process,
the generator of which can be denoted as [20,21,28]

Ω g(η) = c(x, η)[g(ηx ) − g(η)] (1)
x∈Zd
d
where the function g on {0, 1}Z depends on finitely many coordinates, and ηx (z) can be given as
1 − η(x),
{
if y = x
ηx (y) = (2)
η(y), if y ̸ = x
c(x, η) is the transition rate, which is described as [28]
⎧ ∑

⎪ λ p(x, y)η(y), if η(x) = 0

y:|y−x|≤L
c(x, η) = ∑ (3)

⎪ p(x, y)[1 − η(y)], if η(x) = 1

y:|y−x|≤L

where p(x, y) is the transition probability, L represents the distance range of neighbors for every voters, and

y:|y−x|≤L
p(x, y) = 1. For each voter in site x ∈ Zd , the voter holding supportive attitude (‘‘1") or opposing attitude (‘‘0") choose
another voter in neighbor site y (x and y satisfying the condition 0 < |x − y| ≤ L) with probability p(x, y) to adopt this
{0}
voter’s attitude on the basis of a Poisson process at rate 1 or λ respectively. Let ηs (x) denote the state of x ∈ Zd at time s
{θ }
with initial point {0}, and ηs represent the voter model with initial distribution vθ .
In this system, λ is the infection intensity parameter, and there exists a ‘‘critical value’’ λc for the process of voter model,
which can be defined as [21,28]

λc = inf{λ : P(|ηs{0} | > 0, for all s ≥ 0) > 0} (4)


where the |·| is the cardinality function and P is the corresponding probability. If λ < λc , there is convex set C so that on
{0} {0}
Ω∞ = {ηs ̸= ∅, for all s}, we have (1 − ϵ )sC ∩ Zd ⊂ ηs ⊂ (1 + ϵ )sC ∩ Zd for any ϵ > 0 and for all s sufficiently
large [21,28]. Therefore, the process becomes vacant exponentially fast for λ < λc . On the contrary, If λ > λc , there
exists P(ηs ̸ = ∅) ≤ e−γ (λ)s for some positive γ (λ), which indicates that the process survives with the positive probability.
{0}

Moreover, the critical value λc depends on the d-dimensional space Zd .


100 G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113

2.2. Modeling financial price dynamics

Then, a novel complex financial stochastic interacting price model is introduced, where the normal price fluctuations in
stock markets can be simulated by the change of investors’ investment attitudes on the basis of voter interacting system
theory, and rare large and discontinuous price fluctuations coming from sudden economic events and political events are
described by random drastic jumps derived from compound Poisson process.
We assume that there exist M + 1 investors participating in stock trading in stock markets, who are correspondingly
located at the sites {−M /2, . . . , −1, 0, 1, . . . , M /2} ⊂ Z of one-dimensional space. In the constant trading time length l
of every trading day t (t ∈ {1, 2, . . . , N }), assume each investor can trade at most one unit number of the stock at each
time. At the starting of each trading day, assume only the investor at the origin site 0 can receive some important market
information and will hold one of three investment attitudes including buying attitude, selling attitude and neutral attitude.
{0}
Let a random variable ξt with symbolism ‘‘+1’’, ‘‘−1’’, ‘‘0’’ denote the buying attitude, selling attitude and neutral attitude
of this investor with probability p+1 , p−1 or p0 respectively (p+1 + p−1 + p0 = 1). Then the investor will spread information
and send bullish, bearish or neutral signal to his neighbor investors according to the interaction mechanism of voters in voter
interacting system. After receiving market information and signal, these investors will influence each other and change their
investment attitudes, which is supposed to be the main factor of normal price fluctuations for stock markets.
According to the above descriptions, for time s (s ∈ [0, l]) in the tth trading day, we can give the aggregate excess demand
At (s) as follows
{0}
{0} |ηs |
At (s) = ξt × , s ∈ [0, l], t ∈ {1, 2, . . . , N } (5)
m+1
{0} ∑M /2 {0}
where |ηs | = y=−M /2 ηs (y). Then we denote the stock price P (t) derived from the normal and frequent small price
fluctuations in stock financial market as follows [44]
t
{∑ }
P (t) = P (t − 1)exp{α At (s)}, P (t) = P (0)exp α At (s) (6)
k=1

where α (>0) represents the depth parameter of the market, it can adjust the sensitivity of normal price fluctuations.
In the following, the rare large and drastic price fluctuations coming from sudden economic events and political events
are considered, which can be described by random drastic jumps derived from compound Poisson process [33,63]. Let a
random variable ςt denote a Poisson process with intensity γ and a random variable B (ϖk ) be the kth jump size, which
is an independent identical standard normally distributed (i.i.d.) sequence. It is noting that ζt and B (ϖk ) are mutually
independent. Then, the stock price P jump (t) derived from the compound Poisson jump in stock financial market is given
by
ςt
{ ∑ }
P jump (t) = P (0)exp β B (ϖk ) (7)
k=1

where β (β > 0) can adjust the sensitivity of drastic price fluctuations. Then, the overall stock price fluctuations Pt at time
t will mainly be composed of above two parts P (t) and P jump (t) in the probability space (Ω × Ω̃ ), described as
t
{ ∑ ςt }

Pt = P (0)exp α A (ωk ) + β B (ϖk ) (8)
k=1 k=1

where P (0) is the initial stock price at trading day 0. Finally, we translate stock price Pt to the stock logarithmic returns
r(t) [44], described as

r(t) = ln Pt − ln Pt −1 , t ∈ {1, 2, . . . , N }. (9)

In the above construction process of the price model, the model seems to be used to simulate microstructure of price
fluctuations for a single stock. However, a single stock is the basis of corresponding stock index and stock market. Studying
the microstructure of price fluctuations for a single stock contributes to better understand the price fluctuations for the
corresponding stock market index, and a single stock generally has the similar statistical characteristics with an entire
stock market index. Therefore, through simulating some important stocks, the proposed model can describe the complex
dynamics mechanism of price fluctuations for the stock market. In this paper, we perform the statistical analysis on the stock
logarithmic return series derived from above price model, where the number of investors M + 1 = 1000, the number of
trading days N = 3000, the initial density θ = 0.02, the range L = 1, the dimension d = 1, and the initial stock price
P (0) = 100. For two most important parameters, the infection intensity λ in voter model is selected as 3 and 5, and the
jump intensity γ in compound Poisson process is set as 5, 10 and 20. Three real price series of market indexes SSE, SZSE and
HS300 are from 1st August 2005 to 29st November 2017 with 3001 data points (from webset of hk.finance.yahoo.com). The
price series and the corresponding returns for the SZSE and the proposed model with λ = 3 and γ = 5 are displayed in
Fig. 1.
G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113 101

Fig. 1. (a) Plots of stock prices and corresponding returns for SZSE. (b) Plots of stock prices and corresponding returns for the proposed model of λ = 3 and
γ = 5.

Fig. 2. Box plots of real returns (SSE, SZSE, HS300), Gaussian noise series (represented by symbolism ‘‘7’’), and the simulation returns (represented by
symbolism ‘‘1, 2, 3, 4, 5, 6’’ respectively).

3. Stylized facts exploration

In this section, some stylized facts of real return series, Gaussian noise series and the model return series with different
infection intensities λ (λ = 3,5) and jump intensities γ (γ = 5, 10, 20) are focused, including fat-tail property and power-
law scaling behavior. The empirical results in literature indicate that the probability density functions (PDF) of returns series
exhibit the fat-tail compared with the Gaussian time series. Specifically, the distributions of returns have excess kurtosis
and fatter tails than the normal distribution [61,64]. Moreover, the tail distributions of returns can be approximatively fitted
by the functional form f (x) = β1 x−α1 , which is called the power-law scaling behavior. In general, the power-law exponent
α1 ∈ (2.5, 3.5). These stylized facts are able to describe the price fluctuations from different perspectives and distinguish
the differences between financial time series and other time series, such as Gauss noise time series [65–67].
Fig. 2 displays the box plots of real return series, Gaussian noise series, and the simulation return series, which is a
method to describe the distribution of values through their quartiles and judge the degree of dispersion for time series
by the ranges between the distinct parts of the box [68]. The results show that the second quartiles (or the median) of all
return series are approximately equal to 0, while the lower quartiles (Q 1) and the upper quartiles (Q 3) of real returns and
simulation returns series are dramatically less than the values of Gaussian noise series. More specifically, for the simulation
data and the real data, the values of Q 1 and Q 3 are basically within the intervals [−0.009, −0.007] and [0.008, 0.011]
respectively, but the corresponding values of Gaussian noise series are approximately −0.020 and 0.018. There exist similar
results for interquartile ranges (IQR), the upper extremes and the lower extremes, where IQR= Q 3 − Q 1. It indicates that
the distributions and dispersion degree of the simulation data are similar to those of the real return series, but are obviously
deviated from the Gaussian noise series.
Then, the logarithmic plots of PDF for all time series are shown in Fig. 3. It is clearly observed that three real returns and
six simulation returns exhibit fatter tails and more sharp-peak of distribution than Gaussian noise series, which is consistent
102 G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113

Fig. 3. Logarithmic plots of PDF for all real return series, Gaussian noise series, and the simulation return series.

Fig. 4. Plots and log–log plots of cumulative distributions for all real return series, Gaussian noise series, and the simulation return series.

Table 1
Estimates of probability distributions of returns.
data α1 data r(t)
SSE 2.9758 SZSE 3.3150
HS300 3.1742 Gaussian noise 5.2907
γ =5 2.8142 γ =5 3.1425
λ=3 γ = 10 2.9472 λ=5 γ = 10 3.1708
γ = 20 3.0121 γ = 20 3.3450

with the fat-tail phenomenon of the returns in previous empirical researches. Moreover, we focus on the power-law scaling
behaviors of real returns, Gaussian noise series, and the simulation returns. Fig. 4 exhibits the plots and log–log plots of
cumulative distributions. As it is shown, the simulation returns have the similar tracks of the cumulative distributions and
tails with real returns, which also clearly deviate from the Gaussian noise series. The specific estimated results of power-law
exponent α1 of each series are shown in Table 1 according to fitting the last 15% data. The values of α1 of simulation returns
series are within the ranges [2.8142, 3.3450], which are all around 3 and very close to those of real return data, but far from
the value of Gaussian noise series 5.2907. It indicates that the real returns and the simulation returns of the model have the
similar power-law scaling behaviors.
G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113 103

4. Composite distance fuzzy entropy analysis

In this section, a new and efficient entropy-based measure method named composite distance fuzzy entropy (C-FuzzyEn)
is developed, inspired by fuzzy entropy method and complexity-invariant distance, to measure the complexity behaviors of
real return series, Gaussian noise series and the simulation return series with different infection intensities λ (λ = 3,5) and
jump intensities γ (γ = 5, 10, 20), in order to verify the effectiveness of the proposed method.

4.1. Description of composite distance fuzzy entropy method

The fuzzy entropy (FuzzyEn) is proposed by Chen et al. [56,57], where the complexity of time series mainly reflects
regularity and self occurrences of the internal patterns. It is believed that the larger the complexity behavior is in FuzzyEn
method, the less the regularity and the stronger randomness the time series has. For a time series {xt , t ∈ {1, 2, . . . , N }}, the
FuzzyEn can be measured as follows. Firstly, a m-dimension embedding vector is constructed

Xim = {xi , xi+1 , . . . , xi+m−1 } − x̄i , 1≤i≤N −m+1 (10)


1
∑m−1 m
where m represents the embedding dimension, and x̄i = m j=0 xi+j . Then, the distance di,j between two given embedding
vectors Xim and Xjm can be described by the maximum absolute difference as follows

i,j = d[Xi , Xj ] = max{|xi+k − x̄i − (xj+k − x̄j )|, 0 ≤ k ≤ m − 1}.


dm m m
(11)

Next, the similarity degree Dm m


i,j of the distance di,j can be calculated as
{ }
i,j = µ(di,j , n, r) = exp −(di,j ) /r
Dm m m n
(12)

where µ(dm i,j , n, r) represents the fuzzy function, n is the gradient of boundary of fuzzy function, and r is the width of fuzzy
function (or called tolerance value). For all embedding vectors {Xim , 1 ≤ i ≤ N − m + 1}, we can calculate the probability
C m (n, r) by the mean values of Dm i ,j

N −m N −m
1 ∑( 1 ∑ )
C m (n, r) = i,j .
Dm (13)
N −m N −m−1
i=1 j=1,j̸ =i

Similarly, the probability C m+1 (n, r) can be obtained for the embedding dimension m + 1. Finally, the FuzzyEn of time series
is estimated by
( C m (n, r) )
FuzzyEn(m, n, r) = ln . (14)
C m+1 (n, r)
Generally speaking, in order to ensure the accuracy of the fuzzy function and remain more detailed information, the
embedding dimension m and gradient n generally are set as two small constant values. The choice of tolerance value r is
much freer compared with parameters n and m, because the changes of r will not cause the drastic deviation in result of
complexity measure. Moreover, it is worth noting that r is usually multiplied by the standard deviation (SD) of time series
(r × SD) [56,57].
The complexity-invariant distance (CID) is a special and parameter-free correction method for existing distance measures,
which can capture more information between two time series to significantly improve the accuracy of classification
in different domains, such as medicine, engineering, and signal processing [69,70]. For two time series Y = {yt , t ∈
{1, 2, . . . , N }} and Z = {zt , t ∈ {1, 2, . . . , N }}, the concrete form of CID is defined as
CID(Y , Z ) = ED(Y , Z ) × CF(Y , Z ) (15)
√∑
N
where ED represents the Euclidean distance ED(Y , Z ) = i=1 (yi − zi ), and CF is a correction factor described as
max(CE(Y ), CE(Z ))
CF(Y , Z ) = (16)
min(CE(Y ), CE(Z ))
√∑
N −1
CE is a complexity estimate CE(Y ) = i=1 (yi − yi+1 ).
Inspired by the advantage of complexity-invariant distance, a novel generalized methods of FuzzyEn, called composite
distance fuzzy entropy (C-FuzzyEn), is proposed, where the CID is used to calculate the distance CIDmi,j = CID(Xi , Xj )
m m

between two given embedding vectors Xim and Xjm replacing dm i,j derived from the maximum absolute difference. Corre-
m
spondingly, the similarity degree Di,j is described as
m
{ }
i,j ) /r .
Di,j = exp −(CIDm n
(17)
104 G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113

Fig. 5. (a) Plots of FuzzyEn and C-FuzzyEn for white noise and 1/f noise with data length N = 3000. (b) Plots of CVs in FuzzyEn and C-FuzzyEn for white
noise and 1/f noise with data length N = 3000.

m
The probability C (n, r) for the embedding dimension m can be calculated respectively as
N −m N −m
m 1 ∑( 1 ∑ m
)
C (n, r) = Di,j (18)
N −m N −m−1
i=1 j=1,j̸ =i

m+1
Similarly, we can calculate C (n, r) for the embedding dimension m + 1. Thus, the C-FuzzyEn of time series {xt , t ∈
{1, 2, . . . , N }} is computed by
m
( )
C (n, r)
C-FuzzyEn(m, n, r) = ln m+1
. (19)
C (n, r)
In this work, we set parameters m = 2 and n = 2, and calculate the values of C-FuzzyEn at different tolerance levels to
consider the complexity behaviors of time series by means of varying the parameters r.

4.2. Comparison analysis of FuzzyEn and C-FuzzyEn

In this section, in order to clearly verify the effectiveness of C-FuzzyEn method compared with FuzzyEn method, we
firstly apply two common synthetic noise signals - white noise and 1/f noise to test the accuracy of methods, where white
noise represents Gaussian noise comprising uncorrelated data and 1/f noise can be deemed to represent the most complex
fluctuation patterns in physical systems [71]. Then, two representative stock indexes in global financial markets (SSE and
DJIA) are applied to evaluate robustness to noise for FuzzyEn method and C-FuzzyEn method by adding different degrees of
Gaussian noise to original series.

4.2.1. Accuracy test by white noise and 1/f noise


In this paper, we set lengths of white noise series and 1/f noise series N= 3000, where 30 independent samples are
simulated in order to improve accuracy. Fig. 5(a) presents the error bars of FuzzyEn and C-FuzzyEn for white noise time
series and 1/f noise time series at different tolerance levels r × SD. It is obvious that the curves of FuzzyEn and C-FuzzyEn
decrease as tolerance levels r × SD increases, which indicates a decreasing complexity behaviors for white noise and 1/f
noise. However, the decay rates for C-FuzzyEn are obviously weaker than those of FuzzyEn, that is, the change of tolerance
levels has a less impact on C-FuzzyEn method. Therefore, compared with FuzzyEn method, it is easier to choose the proper
tolerance levels in C-FuzzyEn method. Moreover, In order to make a clearer comparison for accuracy of FuzzyEn and C-
FuzzyEn, we plot the values of coefficients of variation (CVs) for white noise and 1/f noise in Fig. 5(b), which are also exhibited
in Table 2. In probability and statistics theory, the coefficient of variation (CV) is an important standardized measure to
compare variation degree or discrete degree of different data series, even if there exist drastically different means for these
data series. The coefficient of variation (CV) is defined as CV = σ /µ, where σ is the standard deviation and µ is the mean. It
is observed that the CVs of FuzzyEn dramatically increase as r increases for 1/f noise, while the CVs of C-FuzzyEn basically
remain approximately small constant, which also can verify the above conclusion. Meanwhile, the CVs of C-FuzzyEn are
completely smaller than those of FuzzyEn for both analyzed noise time series. It indicates that the proposed C-FuzzyEn
has the smaller discrete degree and the greater estimation accuracy than the FuzzyEn, which is an important advantage in
measuring the complexity behaviors of time series.
G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113 105

Table 2
CVs of FuzzyEn and C-FuzzyEn for white noise and 1/f noise.
data r = 0.25 r = 0.5 r = 1.0 r = 2.0 r = 3.0 r = 4.0
white noise 0.0066 0.0089 0.0122 0.0155 0.0170 0.0177
FuzzyEn
1/f noise 0.0639 0.0956 0.1345 0.1606 0.1678 0.1706
white noise 0.0050 0.0049 0.0056 0.0081 0.0105 0.0123
C-FuzzyEn
1/f noise 0.0180 0.0225 0.0324 0.0346 0.0305 0.0270

Fig. 6. Performances of FuzzyEn and C-FuzzyEn on measuring SSE and DJIA returns at different degrees of noise.

4.2.2. Robustness to noise


To better compare the robustness to noise for FuzzyEn method and C-FuzzyEn method, we add the different degrees of
Gaussian noise ranging from 1% to 10% to original return series with length N = 3000 derived from SSE and DJIA, where
the tolerance level is fixed as 0.2 × SD and 50 independent samples of Gaussian noise are added to each return series
respectively. Fig. 6 shows the performances of two entropy-based measure methods on measuring SSE and DJIA returns at
different degrees of noise. It is obvious that the curves of SSE and DJIA for same method are very close, and all curves increase
when degree of noise is small, and rapidly up to a certain saturation point when degree of noise is large enough. However, the
stable value representing the degree of noise corresponding to the saturation point in C-FuzzyEn is approximately 4%, which
is smaller than the stable value 5% in FuzzyEn. Moreover, the ranges of entropy values change are significantly different,
where the C-FuzzyEn is obviously less than the FuzzyEn. The above phenomena indicate that the C-FuzzyEn method shows
stronger robustness to noise than the FuzzyEn method.

4.3. C-FuzzyEn analysis for financial price dynamics

The C-FuzzyEn method is used to measure the complexity behaviors of the proposed model with different infection
intensities λ = 3,5 and jump intensities γ = 5, 10, 20, compared with real return series (SSE, SZSE and HS300) and Gaussian
noise series with same data length. Fig. 7 shows the plots of C-FuzzyEn for these time series with respect to parameter r. It
can be found that the corresponding curves of all time series have the similar downward trends as the r increases, which
suggests the decreasing complexity and irregular. Moreover, the curves of simulation returns series are very close to these
of SSE, SZSE and HS300 while substantially far more below than the curves of Gaussian noise series. It indicates that the
simulation returns series show the similar complexity behaviors to real financial markets, which are more regular and less
complex than Gaussian noise series. The same conclusion also can be obtained from Table 3. Further, it is noted that, for each
fixed infection intensity λ, the C-FuzzyEn values of returns are proportional to the values of jump intensities γ , that is, the
simulation return series corresponding to the larger γ have the larger C-FuzzyEn values and show the stronger complexity
behaviors. For instance, fixing r = 0.5 and λ = 3, C-FuzzyEn values of returns with jump intensities γ = 5, 10, 20 are 1.9525,
1.9698, 1.9938 respectively, and the simulation return series with γ = 20 obviously have the largest values. Because the
γ represents Poisson jump rates, which can influence strength and frequency of undue and rare large price fluctuations in
financial markets. To be more specific, the price fluctuations show more significant with the larger γ . In addition, according
to Fig. 7 and Table 3, the C-FuzzyEn values of returns with λ = 5 are larger than the values of λ = 3 for each fixed γ , which
shows that there exists a positive correlation between the infection intensities λ and the complexity behaviors of returns,
namely, the complexity and randomness increase when the λ increases. The reason is that the transmission speed of the
106 G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113

Fig. 7. Plots of C-FuzzyEn for all real return series, Gaussian noise series, and the simulation return series.

Table 3
C-FuzzyEn of returns with different parameter values of r.
data r = 0.1 r = 0.3 r = 0.5 r = 0.7 r = 0.9
SSE 2.9515 2.2632 1.9713 1.7654 1.6046
SZSE 3.0445 2.3233 2.0230 1.8165 1.6537
HS300 3.0203 2.3049 2.0033 1.7973 1.6359
Gaussian noise 3.3105 2.5059 2.2007 2.0070 1.8559
γ =5 2.9527 2.2569 1.9525 1.7362 1.5642
λ=3 γ = 10 2.9640 2.2767 1.9698 1.7561 1.5906
γ = 20 2.9641 2.2934 1.9938 1.7889 1.6281
γ =5 3.0194 2.3172 2.0147 1.8068 1.6434
λ=5 γ = 10 3.0161 2.3369 2.0339 1.8248 1.6618
γ = 20 3.0575 2.3473 2.0505 1.8488 1.6895

information among investors becomes faster with λ increasing, which leads to more active behaviors for the investors in
financial markets and more frequent price fluctuations. Hence the proposed model will have larger complexity behavior as
the λ increases.
In order to investigate whether the complexity behaviors depend on returns series’ temporal order, the C-FuzzyEn is
applied to calculate the values of a set of random shuffled time series by randomly intermixing orders of the above original
return series, where the shuffled process is repeated 30 times to ensure that shuffled data have representative characters.
Fig. 8(a) shows the error bar plots of shuffled time series generated from SSE, SZSE, HS300 and Gaussian noise series in
comparison with corresponding original series. It is observed that the curves of the shuffled time series decrease as the
r increases, which are similar with their original series. But C-FuzzyEn values of shuffled returns series derived from SSE,
SZSE, HS300 are larger than those of the original series, but are still smaller than the values of shuffled Gaussian noise
series and original Gaussian noise series. It can be seen that the values of shuffled Gaussian noise series remain about the
same compared with original series. The results indicate that the temporal orders of returns have a significant impact on
their complexity behaviors, and random shuffling process can destroy the regularity of intrinsic structures, which leads to
the increasing complexity behaviors and randomness. Conversely, there basically does not exist the regularity in Gaussian
noise series, that is to say, the data points of Gaussian noise series are basically unconnected. Moreover, the C-FuzzyEn
values of shuffled data for the model are shown in Fig. 8(b). It is obvious that, compared with shuffled Gaussian noise series,
the shuffled simulation returns are close to the shuffled time series derived from SSE, SZSE, HS300, which suggests the
similarity between the simulation returns and the real returns in terms of regularity and intrinsic structures. Moreover, for
different infection intensities λ and jump intensities γ , there exists same change law of complexity behaviors for the shuffled
simulation returns similar to the original series. It indicates that the regularity of the simulation data is still not completely
destroyed by randomly intermixing orders.
Furthermore, we consider the C-FuzzyEn values of absolute return series with different power exponents q = 0.25, 0.5,
0.75, 1, 1.5, 2, 3, described as |r(t)|q , to measure the complexity behaviors of different volatility behaviors levels for return
series. Corresponding results for each real returns and the simulation returns are shown in Fig. 9(a)–(i) respectively. It is
observed that the values of C-FuzzyEn for all absolute returns decreases as r increases, which are similar to the trends of
return series. Moreover, it is worth noting that the C-FuzzyEn value for fixed r becomes larger when the power exponent q
decreases, indicating that the complexity behaviors of absolute returns series are inversely proportional to the values of q.
In particular, we plot the C-FuzzyEn of |r(t)| for SSE, SZSE, HS300 and simulation returns in Fig. 9(j), which shows that the
G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113 107

Fig. 8. (a) Plots of shuffled time series generated from SSE, SZSE, HS300 and Gaussian noise in comparison with corresponding original series. (b) Plots of
shuffled time series generated from the simulation returns, SSE, SZSE, HS300 and Gaussian noise.

Table 4
C-FuzzyEn of |r(t)| with different parameter values of r.
data r = 0.1 r = 0.3 r = 0.5 r = 0.7 r = 0.9
SSE 2.6681 2.0262 1.7557 1.5734 1.4365
SZSE 2.7758 2.0854 1.7951 1.6084 1.4672
HS300 2.7279 2.0517 1.7700 1.5887 1.4529
γ =5 2.6217 2.0029 1.7153 1.5202 1.3732
λ=3 γ = 10 2.6629 2.0282 1.7395 1.5409 1.3885
γ = 20 2.7145 2.0652 1.7769 1.5739 1.4175
γ =5 2.7349 2.0744 1.7804 1.5803 1.4292
λ=5 γ = 10 2.7760 2.1017 1.8090 1.6097 1.4583
γ = 20 2.7987 2.1214 1.8351 1.6490 1.5071

absolute returns with a larger λ or γ for the proposed model have a larger C-FuzzyEn value, namely, it have the stronger
complexity behaviors, which can be explained by the similar reasons as before. Moreover, compared with corresponding
r(t), there exist the smaller C-FuzzyEn values of |r(t)| for all returns, which indicates that the absolute returns series have
stronger regularity than the return series, and the fluctuation behaviors show more complex phenomenon compared with
the volatility behaviors. The corresponding results are also shown in Table 4.

5. Matching energy analysis

In this section, another complexity measure called matching energy (ME) is applied to investigate the complexity
behaviors of real return series, Gaussian noise series and the simulation returns with λ = 3,5 and γ = 5, 10, 20 from a
different perspective.

5.1. Description of matching energy method

Matching energy (ME), as a novel complexity measure proposed by Fouda [72] in recent years, can estimate the
complexity behaviors by comparing the matching coordinates between the sorted time series and the initial time series.
In general, the more complex dynamics system should have the larger ME value, corresponding to the stronger randomness
and larger complexity behaviors. Empirical research shows that the ME method has low cost and strong robustness to noise
in complexity behaviors analysis of real-world data. It is believed that the ME method is useful for complexity analysis of
return series in financial markets.
For a time series {xt , t ∈ {1, 2, . . . , N }} of length N, the ME can be calculated as follows [72]. Firstly, the differential
dynamical quantization (DDQ) method is applied to normalize the time series {xt }, where the detailed procedures of the
algorithm are described in [73]. Two reorder time series {pt , t ∈ {1, 2, . . . , N }} and {qt , t ∈ {1, 2, . . . , N }} are obtained
by rearranging normalized {xt } by ascending order and descending order respectively. Then, constructing m-dimensional
vectors {Xi }, {Pi } and {Qi } corresponding to {xt }, {pt } and {qt } through

Xi = {xiτ , xiτ +1 , . . . , xiτ +m−1 }, Pi = {piτ , piτ +1 , . . . , piτ +m−1 } (20)

Qi = {qiτ , qiτ +1 , . . . , qiτ +m−1 } (21)


108 G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113

Fig. 9. (a)–(i) Plots of C-FuzzyEn of |r(t)|q with power exponent q respectively for SSE, SZSE, HS300 and the simulation data. (j) Plots of C-FuzzyEn of |r(t)|
for SSE, SZSE, HS300 and the simulation data.

where τ is the delay time of the embedding vectors, 1 ≤ i ≤ N − m + 1 and 1 ≤ τ ≤ m. The dimensional vector
{Xi , i ∈ {1, 2, . . . , N − m + 1}} is rearranged by ascending order and descending order to produce two time series
{Yi , i ∈ {1, 2, . . . , N − m + 1}} and {Zi , i ∈ {1, 2, . . . , N − m + 1}}.
Suppose V (1) and V (2) are the vector spaces of matching coordinates and non-matching coordinates of {Xi }, {Pi } and {Qi },
and m1 and m2 are the number of matching coordinates between {Yi } and {Pi }, and between {Zi } and {Qi }, respectively.
G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113 109

Table 5
ME values of returns r(t) and absolute returns |r(t)|.
data r(t) |r(t)| data r(t) |r(t)|
SSE 0.3568 0.1894 SZSE 0.4095 0.2159
HS300 0.3625 0.2022 Gaussian noise 0.6280 0.4134
γ =5 0.3444 0.1663 γ =5 0.3773 0.2090
λ=3 γ = 10 0.3579 0.1736 λ=5 γ = 10 0.3836 0.2131
γ = 20 0.3713 0.2026 γ = 20 0.4010 0.2365

Therefore, the dimensional of V (1) and V (2) are m1 and m − m1 for {Pi } respectively, similarly, m2 and m − m2 for {Qi }. Next,
the partial energy Ei (m) for the ith embedding vectors of {Xi }, {Pi } and {Qi } is defined as
0,
{
if m1 + m2 ̸ = 0
Ei (m) = (22)
Ei+ + di+ , if m1 + m2 = 0
(2)
where Ei+ is energy of {Xi }, corresponding to {Xi } in vector space V (2) , described as

 m−m1
 1 ∑( (2) (2)
.
)
Ei+ =√ Xi (k) − Xi (23)
m − m1
k=1

The Xt is the mean value of Xt . Moreover, di+ represents the distance of non-matching coordinates, defined as
1 ( (p) (q)
)
di+ = di+ + di+ (24)
2
where
 
 m−m1  m−m2
(p)
 1 ∑( (2) (2) (q)
 1 ∑( (2) (2)
, .
) )
di+ =√ Yi (k) − Pi di+ =√ Zi (k) − Qi (25)
m − m1 m − m2
k=1 k=1

Finally, the matching energy ME is calculated by the average value of partial energies Ei
T −1
1∑
ME = Ei (m). (26)
T
i=0

Empirical researches show that there exists correlation between the choice of delay time τ and embedding dimension m.
In general, the larger τ is, the smaller m should be selected, which can reduce the computational time [72]. In this paper, m
and τ are set to 20 and 10, respectively.

5.2. Matching energy analysis for financial price dynamics

We apply the ME method to analyze the complexity behaviors of r(t) and |r(t)| for above time series. The corresponding
results are shown in Table 5. It can be found that ME values of r(t) for the simulation returns are around 0.37 and
within the interval [0.3444, 0.4010], which are close to the values of real returns (MESSE = 0.3568, MESZSE = 0.4095,
MEHS300 = 0.3625) but far less than that of Gaussian noise series, i.e., 0.6280. Similarly, the ME values of |r(t)| for all time
series have the similar phenomenon, but those are obviously smaller than those of corresponding r(t). For example, r(t) and
|r(t)| of SSE are 0.3568 and 0.1894. Therefore, there exist similar complexity behaviors between the simulation returns and
the real returns, which have weaker randomness and simpler structure than Gaussian noise series. Moreover, the return
series have stronger complexity and randomness compared with the absolute return series, indicating that the volatility
behaviors show more regularity than the fluctuation behaviors. Furthermore, considering r(t) and |r(t)| for the simulation
data, the ME values increase with the increase of infection intensity λ and jump intensity γ , which shows that λ or γ has
the positive correlation with complexity behaviors of the volatility behaviors and the fluctuation behaviors, namely, the
complexity and randomness of the proposed model increase when λ or γ increases.

6. Multifractal property for financial dynamics

In this section, the MF-DFA method is utilized to study the multifractal properties of the real return series and the
simulation returns series with λ = 3,5 and γ = 5, 10, 20.
110 G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113

6.1. Description of MF-DFA method

Multifractal detrended fluctuation analysis, as one of the most common methods for analyzing the multifractal properties
of time series, is proposed by Kantelhardt et al. [62]. For a time series {xt , t ∈ {1, 2, . . . , N }}, the MF-DFA algorithm can be
sketched below [19,49,62]. Firstly, a profile from the original series is calculated by
i

Y (i) = [xt − ⟨x⟩], i = 1, . . . , N (27)
t =1

where ⟨·⟩ denotes the average of the entire time series of length N. We divide the profile Y (i) into Ns = int(N /s) disjoint
segments of equal length s starting from the beginning of the time series. Similarly, the same process is repeated starting
from the end of the time series. Then, the local trend for each of the 2Ns segments is calculated by

Ys (i) = Y [(Ns − ν )s + i] − yν (i) (28)


for each segment ν = 1, . . . , Ns , and

Ys (i) = Y [N − (ν − Ns )s + i] − yν (i) (29)


for ν = Ns + 1, . . . , 2Ns . The variance of the local trend is calculated by
s
1∑
F 2 (ν, s) = [Ys (i)]2 . (30)
s
i=1

Moreover, qth order fluctuation function by averaging all the segments


2Ns
{ 1 ∑ } 1q
q
Fq (s) = [F 2 (ν, s)] 2 (31)
2Ns
ν=1

where the index variable q is real value and the MF-DFA method reverts to the standard DFA method for q = 2.
Finally, the scaling behavior of the fluctuation is determined by analyzing Fq (s) log–log plots versus different scales s for
each index variable q. If the time series {xt , t ∈ {1, 2, . . . , N }} is long-range power-law correlated, the fluctuation function
Fq (s) versus s reveals the power law dependence for large values of s

Fq (s) ∼ sh(q) (32)


where h(q) represents the generalized Hurst exponent. In general, the time series is monofractal when h(q) is a constant for
all q, otherwise the time series is multifractal. In addition, the multifractal properties can also be measured by the multifractal
spectrum f (α ) = q[α − h(q)] + 1, where α = h(q) + qh′ (q). It is known that the multifractal spectrum of multifractal time
series basically shows shape of an inverted parabola, whose width can measure the degree of multifractality. On the contrary,
the multifractal spectrum of monofractal time series is a point [19,49,62].

6.2. MF-DFA for return series

We use the MF-DFA method to study the multifractal properties of above returns series. Fig. 10 shows the log–log plots of
fluctuation functions Fq (s) versus s, where the ranges of order q are set from −5 to 5. It is observed that log Fq (s) has a linear
relationship with log s for both the simulated data and the real data, indicating that these returns series are long-range power
correlated. Moreover, the plots of generalized Hurst exponent h(q) and multifractal spectrum f (α ) are shown in Fig. 11(a) and
Fig. 11(b) respectively. It is seen that the curve of h(q) basically decreases when q increases for each returns series, namely,
the value of h(q) is not a constant for all q. Each curve of f (α ) is similar to an inverted parabola but not a point. It indicates
that the simulation returns series with λ = 3,5 and γ = 5, 10, 20 have the similar multifractal properties with the real
return series. In general, the price model is able to reproduce the multifractal property of financial markets.

7. Conclusion

In this work, in order to study the fluctuation dynamic mechanism of price fluctuations in financial markets, a novel
complex financial stochastic interacting price model is introduced based on the theory of voter system and compound
Poisson process, where the dynamical mechanism of voter system is used to simulate the change of investors’ investment
attitudes towards the information of financial markets leading to the frequent small price fluctuations, and the compound
Poisson process can simulate the undue and rare large price fluctuations in financial markets derived from the sudden
economic events and political events. In order to verify the rationality of the proposed model, we compare the fluctuation
natures of the proposed model with different infection intensities λ = 3,5 and jump intensities γ = 5, 10, 20 with the real
return series (SSE, SZSE, HS300), Gaussian noise series of same data length. By empirical probability density function analysis,
the simulation return series show the similar stylized facts, including fat-tail property and power-law scaling behavior with
G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113 111

Fig. 10. Log–log plots of Fq (s) versus s of simulation returns and real returns.

Fig. 11. (a) Hurst exponent h(q) of simulation returns and real returns. (b) Multifractal spectra f (α ) of simulation returns and real returns.

real return series but are obviously different from Gaussian noise series. Then, in order to better measure the complexity
behaviors of the proposed price model, a new entropy-based approach called composite distance fuzzy entropy (C-FuzzyEn)
is proposed based on fuzzy entropy method and complexity-invariant distance. The effectiveness analyses including accuracy
test and robustness to noise are investigated to compare with the effectiveness of C-FuzzyEn method and FuzzyEn method.
The empirical results show the proposed C-FuzzyEn exhibits the smaller discrete degree and the greater estimation accuracy
for white noise and 1/f noise than the FuzzyEn. Similarly, it has the stronger robustness to noise for two representative stock
indexes (SSE and DJIA). Moreover, we apply the C-FuzzyEn to study complexity behaviors of r(t) and |r(t)| for above time
series. Empirical results show that r(t) and |r(t)| of the simulation data have the similar complexity behaviors with the
real financial data, which are far less than the complexity of Gaussian noise series. However, the complexity behaviors
of |r(t)| for all time series are smaller than those of corresponding r(t), indicating that the fluctuation behaviors have
stronger randomness and complexity than the volatility behaviors. The research on complexity behaviors of random shuffled
time series reveals that there exists a similar impact between the complexity behaviors and the temporal orders of return
series, where random shuffling process can lead to the their increasing complexity behaviors and randomness. In addition,
112 G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113

it is noted that the complexities of returns and absolute returns increase when infection intensity λ or jump intensity γ
increases. Further, another approach of complexity measure named matching energy (ME) is firstly applied to investigate
the complexity behaviors of r(t) and |r(t)| for all time series. The corresponding empirical results suggest that there exist
similar conclusions with the above C-FuzzyEn method. Finally, the MF-DFA is used to study the multifractal properties of
the simulation returns and the real returns. The empirical results show that the simulation returns have the multifractal
properties, which are similar to the real returns. Through the above analysis and comparisons, the financial price model is
able to simulate some fluctuation natures of financial markets, indicating the certain rationality of the proposed price model
to some extent.

Acknowledgments

Authors were supported by National Natural Science Foundation of China Grant Nos. 11371050 and 71271026.

References

[1] F. Black, M. Scholes, The pricing of options and corporate liabilities, J. Polit. Econ. 81 (1973) 637–654.
[2] E. Ghysels, C. Gouriéroux, J. Jasiak, Stochastic volatility duration models, J. Econometrics 119 (2004) 413–433.
[3] G.F. Gu, W.X. Zhou, Emergence of long memory in stock volatility from a modified Mike-Farmer model, EPL (Europhys. Lett.) 86 (2009) 48002.
[4] G.F. Gu, W.X. Zhou, On the probability distribution of stock returns in the Mike-Farmer model, Eur. Phys. J. B 67 (2009) 585–592.
[5] H. Meng, F. Ren, G.F. Gu, X. Xiong, Y.J. Zhang, W.X. Zhou, W. Zhang, Effects of long memory in the order submission process on the properties of
recurrence intervals of large price fluctuations, EPL (Europhys. Lett.) 98 (2012) 38003.
[6] T.C. Mills, The Econometric Modeling of Financial Time Series, Cambridge University Press, Cambridge, UK, 1999.
[7] X.H. Ni, Z.Q. Jiang, W.X. Zhou, Degree distributions of the visibility graphs mapped from fractional Brownian motions and multifractal random walks,
Phys. Lett. A 373 (2009) 3822–3826.
[8] T. Preis, S. Golke, W. Paul, J.J. Schneider, Statistical analysis of financial returns for a multi-agent order book model of asset trading, Phys. Rev. E. 76
(2007) 016108.
[9] M.C. Qian, Z.Q. Jiang, W.X. Zhou, Universal and nonuniversal allometric scaling behaviors in the visibility graphs of world stock market indices, J. Phys.
A Math. Theor. 43 (2010) 335002.
[10] E. Samanidou, E. Zschischang, D. Stauffer, T. Lux, Agent-based models of financial markets, Rep. Progr. Phys. 70 (2007) 409–450.
[11] A. Svorenčík, F. Slanina, Interacting gaps model, dynamics of order book, and stock-market fluctuations, Eur. Phy. J. B. 57 (2007) 453–462.
[12] A.B. Tavares, J.D. Curto, G.N. Tavares, Modelling heavy tails and asymmetry using ARCH-type models with stable Paretian distributions, Nonlinear
Dynam. 51 (2008) 231–243.
[13] R.D. Willmann, G.M. Schütz, D. Challet, Exact Hurst exponent and crossover behavior in a limit order market model, Physica A 316 (2002) 430–440.
[14] Z. Kostanjčar, S. Begušić, H.E. Stanley, B. Podobnik, Estimating tipping points feedback-driven financial networks, IEEE J. Sel. Top. Sign. Proces. 10
(2016) 1040–1052.
[15] Stjepan Begušić, Zvonko Kostanjčar, Dejan Kovač, H. Eugene Stanley, Boris Podobnik, Information feedback in temporal networks as a predictor of
market crashes, Complexity 2018 (2018) 2834680.
[16] A. Yasutomi, The emergence and collapse of money, Physica D 82 (1995) 180–194.
[17] J.H. Zhang, P. McBurney, K. Musial, Convergence of trading strategies in continuous double auction markets with boundedly-rational networked
traders, Rev. Quant. Finan. Acc. 50 (2018) 301–352.
[18] J.H. Zhang, Influence of individual rationality on continuous double auction markets with networked traders, Physica A 495 (2018) 353–392.
[19] P. Oświecimka, S. Droz̄dz̄, R. Gebarowski, A.Z. Górski, J. Kwapien, Multiscaling edge effects in an agent based money emergence model, Acta Phys.
Polon. B 46 (2015) 1579–1591.
[20] P. Clifford, A. Sudbury, A model for spatial conflict, Biometrika 60 (1973) 581–588.
[21] R. Durrett, Lecture Notes on Particle Systems and Percolation, Wadsworth & Brooks, California, 1988.
[22] A.Z. Górski, S. Droz̄dz̄, P. Oświecimka, Modelling emergence of money, Acta Phys. Pol. A 117 (2010) 676–680.
[23] R.A. Holley, T.M. Liggett, Ergodic theorems for weakly interacting infinite systems and the voter model, Ann. Probab. 3 (1975) 643–663.
[24] G. Iori, A microsimulation of traders activity in the stock market: The role of heterogeneity, agents’ interaction and trade frictions, J. Econ. Behav.
Organ. 49 (2002) 269–285.
[25] T. Kaizoji, Speculative bubbles and crashes in stock markets: An interacting-agent model of speculative activity, Physica A 287 (2000) 493–506.
[26] A. Krawiecki, Microscopic spin model for the stock market with attractor bubbling and heterogeneous agents, Internat. J. Modern Phys. C 16 (2005)
549–559.
[27] S.A. Kiselev, A. Phillips, I. Gabitov, Long scale evolution of a nonlinear stochastic dynamic system for modeling market price bubbles, Phys. Lett. A 272
(2000) 130–142.
[28] T.M. Liggett, Stochastic Interacting Systems: Contact, Voter, and Exclusion Processes, Springer, 1999.
[29] T. Lux, M. Marchesi, Scaling and criticality in a stochastic multiagent model of a financial market, Nature 397 (1999) 498–500.
[30] T. Lux, Financial Power Laws: Empirical Evidence, Models and Mechanisms, Cambridge University Press, Cambridge, UK, 2008.
[31] D. Mollison, Spatial contact models for ecological and epidemic spread, J. R. Stat. Soc. 39 (1977) 283–326.
[32] D. Stauffer, T.J.P. Penna, Crossover in the Cont-Bouchaud percolation model for market fluctuation, Physica A 256 (1998) 284–290.
[33] Y.D. Wang, S.Z. Zheng, W. Zhang, J. Wang, G.C. Wang, Modeling and complexity of stochastic interacting Levy type financial price dynamics, Physica
A 499 (2018) 498–511.
[34] Y. Yu, J. Wang, Lattice oriented percolation system applied to volatility behavior of stock market, J. Appl. Stat. 39 (2012) 785–797.
[35] J.H. Zhang, J. Wang, Modeling and simulation of the market fluctuations by the finite range contact systems, Simul. Model. Pract. Th. 18 (2010) 910–925.
[36] Y.F. Lu, J. Wang, H.L. Niu, Agent-based financial dynamics model from stochastic interacting epidemic system and complexity analysis, Phys. Lett. A
379 (2015) 1023–1031.
[37] D. Sornette, W.X. Zhou, Importance of positive feedbacks and overconfidence in a self-fulfilling Ising model of financial markets, Physica A 370 (2006)
704–726.
[38] C. Aghamohammadi, M. Ebrahimian, H. Tahmooresi, Permutation approach, high frequency trading and variety of micro patterns in financial time
series, Physica A 413 (2014) 25–30.
[39] L. Calvet, A. Fisher, Multifractal Volatility: Theory, Forecasting, and Pricing, Academic Press, New York, USA, 2008.
[40] L. Feng, B. Li, B. Podobnik, T. Preis, H.E. Stanley, Linking agent-based models and stochastic models of financial markets, Proc. Natl. Acad. Sci. USA 109
(2012) 8388–8393.
G. Wang, S. Zheng and J. Wang / Physica A 517 (2019) 97–113 113

[41] G.F. Gu, W. Chen, W.X. Zhou, Empirical distributions of Chinese stock returns at different microscopic timescales, Physica A 387 (2008) 495–502.
[42] Z.Q. Jiang, W.X. Zhou, Scale invariant distribution and multifractality of volatility multipliers in stock markets, Physica A 381 (2007) 343–350.
[43] Z.Q. Jiang, W.X. Zhou, Multifractality in stock indexes: Fact or fiction? Physica A 387 (2008) 3605–3614.
[44] R.N. Mantegna, H.E. Stanley, An Introduction To Econophysics: Correlations and Complexity in Finance, Cambridge University Press, Cambridge, 1999.
[45] H.L. Niu, J. Wang, Volatility clustering and long memory of financial time series and financial price model, Digit. Signal Process. 23 (2013) 489–498.
[46] B. Podobnik, H.E. Stanley, Detrended cross-correlation analysis: A new method for analyzing two nonstationary time series, Phys. Rev. Lett. 100 (2008)
084102.
[47] B. Podobnik, I. Grosse, D. Horvatic, S. Ilic, P.C. Ivanov, H.E. Stanley, Quantifying cross-correlations using local and global detrending approaches, Eur.
Phys. J. B 71 (2009) 243–250.
[48] B. Podobnik, D. Horvatic, A.M. Petersen, H.E. Stanley, Cross-correlations between volume change and price change, Proc. Natl. Acad. Sci. USA 106
(2009) 22079–22084.
[49] S.A.R. Rizvi, G. Dewandaru, O.I. Bacha, M. Masih, An analysis of stock market efficiency: Developed vs Islamic stock markets using MF-DFA, Physica A
407 (2014) 86–99.
[50] W. Zhang, J. Wang, Nonlinear stochastic exclusion financial dynamics modeling and time-dependent intrinsic detrended cross-correlation, Physica A
482 (2017) 29–41.
[51] L. Zunino, D.G. Pérez, M.T. Martín, M. Garavaglia, A. Plastino, O.A. Rosso, Permutation entropy of fractional brownian motion and fractional gaussian
noise, Phys. Lett. A 372 (2008) 4768–4774.
[52] L. Zunino, M. Zanin, B.M. Tabak, D.G. Pérez, O.A. Rosso, Forbidden patterns, permutation entropy and stock market inefficiency, Physica A 388 (2009)
2854–2864.
[53] L. Zunino, M. Zanin, B.M. Tabak, D.G. Pérez, O.A. Rosso, Complexity-entropy causality plane: A useful approach to quantify the stock market inefficiency,
Physica A 389 (2010) 1891–1901.
[54] C. Bandt, B. Pompe, Permutation entropy: A natural complexity measure for time series, Phys. Rev. Lett. 88 (2002) 174102.
[55] S.M. Pincus, D.L. Keefe, Quantification of hormone pulsatility via an approximate entropy algorithm, Am. J. Physiol. 262 (1992) 741–754.
[56] W.T. Chen, Z.Z. Wang, H. Xie, W. Yu, Characterization of surface EMG signal based on fuzzy entropy, IEEE Trans. Neural Syst. Rehabil. Eng. 15 (2007)
266–272.
[57] W.T. Chen, J. Zhuang, W.X. Yu, Z.Z. Wang, Measuring complexity using FuzzyEn, ApEn, and SampEn, Med. Eng. Phys. 31 (2009) 61–68.
[58] J.F. Valencia, A. Porta, M. Vallverdú, F. Clarià, R. Baranowski, E. Orowska-Baranowska, P. Caminal, Refined multiscale entropy: Application to 24-h
Holter recordings of heart period variability in healthy and aortic stenosis subjects, IEEE Trans. Biomed. Eng. 56 (2009) 2202–2213.
[59] A. Lempel, J. Ziv, On the complexity of finite sequences, IEEE Trans. Inform. Theory 22 (1976) 75–81.
[60] J.S. Shiner, M. Davison, P.T. Landsberg, Simple measure for complexity, Phys. Rev. E 59 (1999) 1459–1464.
[61] J. Kwapień, S. Drożdż, Physical approach to complex systems, Phys. Rep. 515 (2012) 115–226.
[62] J.W. Kantelhardt, S.A. Zschiegner, E. Koscielny-Bunde, S. Havlin, A. Bunde, H.E. Stanley, Multifractal detrended fluctuation analysis of nonstationary
time series, Physica A 316 (2002) 87–114.
[63] R.L. Schilling, Financial modelling with jump processes, Publ. Am. Stat. Assoc. 101 (2005) 1315–1316.
[64] R. Cont, Empirical properties of asset returns: Stylized facts and statistical issues, Quant. Finance 1 (2001) 223–236.
[65] A. Clauset, C.R. Shalizi, M.E.J. Newman, Power-law distributions in empirical data, SIAM Rev. 51 (2009) 661–703.
[66] V. Plerou, P. Gopikrishnan, B. Rosenow, L. Amaral, H.E. Stanley, Econophysics: Financial time series from a statistical physics point of view, Physica A
279 (2000) 443–456.
[67] R.N. Mantegna, H.E. Stanley, Scaling behaviour in the dynamics of an economic index, Nature 376 (2002) 46–49.
[68] D.C. Hoaglin, Some implementations of the boxplot, Am. Stat. 43 (1989) 50–54.
[69] G.E.A.P.A. Batista, X.Y. Wang, E.J. Keogh, A complexity-invariant distance measure for time series, SDM 11 (2011) 699–710.
[70] G.E.A.P.A. Batista, E.J. Keogh, O.M. Tataw, V.M.A. de Souza, CID: An efficient complexity-invariant distance for time series, Data Min. Knowl. Discov.
28 (2014) 634–669.
[71] M.T. Lo, Y.C. Chang, C. Lin, H.W. Young, Y.H. Lin, Y.L. Ho, C.K. Peng, K. Hu, Outlier-resilient complexity analysis of heartbeat dynamics, Sci. Rep. 5 (2015)
8836.
[72] J.S.A.E. Fouda, The matching energy: A novel approach for measuring complexity in time series, Nonlinear Dynam. 86 (2016) 2049–2060.
[73] J.S.A.E. Fouda, W. Koepf, Detecting regular dynamics from time series using permutations slopes, Commun. Nonlinear Sci. Numer. Simul. 27 (2015)
216–227.

You might also like