Analysis of Shannon-Fisher information plane in time series based on information

entropy
Yuanyuan Wang, and Pengjian Shang

Citation: Chaos 28, 103107 (2018); doi: 10.1063/1.5023031

View online: https://doi.org/10.1063/1.5023031
View Table of Contents: http://aip.scitation.org/toc/cha/28/10
Published by the American Institute of Physics

Articles you may be interested in

Statistics of inverse interspike intervals: The instantaneous firing rate revisited
Chaos: An Interdisciplinary Journal of Nonlinear Science 28, 106305 (2018); 10.1063/1.5036831

Harnessing stochasticity: How do organisms make choices?

Chaos: An Interdisciplinary Journal of Nonlinear Science 28, 106309 (2018); 10.1063/1.5039668

Recurrence-based information processing in gene regulatory networks

Chaos: An Interdisciplinary Journal of Nonlinear Science 28, 106313 (2018); 10.1063/1.5039861

Differentiating resting brain states using ordinal symbolic analysis

Chaos: An Interdisciplinary Journal of Nonlinear Science 28, 106307 (2018); 10.1063/1.5036959

Two-dimensional spatiotemporal complexity in dual-delayed nonlinear feedback systems: Chimeras and

dissipative solitons
Chaos: An Interdisciplinary Journal of Nonlinear Science 28, 103106 (2018); 10.1063/1.5043391

Efficient determination of synchronization domains from observations of asynchronous dynamics

Chaos: An Interdisciplinary Journal of Nonlinear Science 28, 106301 (2018); 10.1063/1.5037012
 CHAOS 28, 103107 (2018)

Analysis of Shannon-Fisher information plane in time series based on

information entropy
Yuanyuan Wanga) and Pengjian Shangb)
Department of Mathematics, School of Science, Beijing Jiaotong University, Beijing 100044,
People’s Republic of China
(Received 20 January 2018; accepted 26 September 2018; published online 11 October 2018)
In this paper, we propose a Shannon-Fisher information plane based on the information entropy to
analyze financial stock markets. In order to evaluate the effectiveness of this method, we apply this
method to two types of artificial time series: Autoregressive Fractionally Integrated Moving Average
models and Chebyshev map model. The results show that with the embedding dimension m and the
number of possible states of the system M increasing, the normalized Shannon entropy increases, and
the Fisher information measure (FIM) decreases. When the parameter M is not so big, the embedding
dimension m plays a leading role in determining the FIM. In addition, compared with the classical
Shannon-Fisher information through permutation entropy, we conclude that the proposed approach
can give us more accurate information on the classification of financial stock markets. Published by
AIP Publishing. https://doi.org/10.1063/1.5023031

The paper proposes a Shannon-Fisher information plane
based on the information entropy to analyze the financial
time series of U.S., Australian, and Asian stock mar-
kets. As we know, the probability distribution P mainly
depends on the number of possible states of the system M.
We find that with the increase of the embedding dimen-
sion m and the parameter M, the normalized Shannon
entropy increases, and the Fisher information measure
(FIM) decreases, indicating that as m and M increase,
the financial time series is more complex, thus produc-
ing large amounts of normalized Shannon entropy and its
corresponding FIM is smaller. In addition, applying this
method into analyzing the financial stock markets, we also
conclude that the ten stock markets are organized into
three groups according to their own characteristics. The
three Asian stock markets TWII, SSEC, and CTSP300
are grouped together. The second group consists of the
American stock markets: DJI, NDX, NYA, and S&P500.
The rest Australian stock markets AFLI, AORD, and
AXJO are grouped together. We observe that the FIM
values of the four American stock markets DJI, NDX,
NYA, and S&P500 are bigger than those the other stock
markets. These suggest that these four American stock
markets are less active than the Asian stock markets and
the Australian stock markets.
I. INTRODUCTION
Time series of measurements are the basic elements for
investigating natural phenomena. From time series, we can
extract much information on dynamical systems. Recently,
information-based measures of randomness (or “regularity”)
of time series have attracted more and more attention in vari-
ous branches of science.1–11 As is well-known, the Shannon
entropy is the important magnitude to quantify the degree
a)
Author to whom correspondence should be addressed:
17121633@bjtu.edu.cn
b)
pjshang@bjtu.edu.cn
1054-1500/2018/28(10)/103107/9/$30.00
of disorder in dynamical systems. It is a standard measure
for the order state of sequences and has been applied pre-
viously to DNA sequences. The Shannon entropy quantifies
the probability density function of the distribution of values.
In addition, it can also be used to represent the degree of
uncertainty involved in predicting the output of a probabilis-
tic event.12,13 For discrete distributions, the probability will
be a maximum value if one predicts the result exactly before
it happens; as a consequence, the Shannon entropy will be a
minimum. If one is absolutely capable of predicting the result
of an incident, the Shannon entropy will be zero. As for con-
tinuous distributions, that is not the case. For instance, over
the real line, Shannon entropy can reach any arbitrary value,
positive or negative. In practice, some studies have analyzed
the physical and chemical properties of atomic and molecular
systems from an information theoretical point of view.14–25
Nowadays, the Fisher Information Measure (FIM),26
which was introduced by Fisher in 1925 in the context of
statistical estimation, is also playing an increasing role in sci-
entific analysis. It is a powerful tool to detect complex and
non-stationary signals. The Frieden extreme physical infor-
mation principle27 utilizes FIM to derive important laws of
chemistry and physics,28 such as the equations of the nonrel-
ativistic quantum mechanics29 or relevant results in density
functional theory.30,31 Since FIM allows one to accurately
describe the behaviors of dynamic systems,32 its application to
the characterization of complex signals issued from these sys-
tems appears quite natural. Martin et al. have employed this
approach to characterize the dynamics of electroencephalo-
gram (EEG) signals.33 They have shown the informative
content of FIM in investigating significant changes in the
behavior of nonlinear dynamical systems34 disclosing. There-
fore, FIM is an important quantity involved in many aspects
of the theoretical and observational description of natural
phenomena. Moreover, one of the interesting results of their
study is that FIM can detect some non-stationary behav-
iors in situations where the Shannon entropy shows limited
dynamics.
28, 103107-1 Published by AIP Publishing.
103107-2 Y. Wang and P. Shang Chaos 28, 103107 (2018)

Motivated by the work, some researchers define a a key role in information theory. From Eq. (2), Martin et al.
Shannon-Fisher information plane as a measure of complex- introduce the integrand, which is the scalar product of two
ity. In the previous studies, Romera and Dehesa35 introduced vectors, is independent of the reference frame.34
the concept of Shannon-Fisher information plane as a specific Let us pay attention to the one-dimensional case. Con-
correlation measure. In addition, the Shannon-Fisher plane sider a random variable x whose probability density function
has been applied successfully into all kinds of fields, such as is represented as f (x). Its associated FIM27,37 is given by
the analysis of signals,32 the study of electron correlation35  +∞  
within the He-isoelectronic series, and so on. The Shannon- 1 df (x) 2
I= dx
Fisher information plane allows one to quantify the global −∞ f (x) dx
versus local characteristic of the time series generated by  +∞  
dψ(x) 2
the dynamical process. This means that it incorporates a =4 dx, (3)
−∞ dx
global measure of uncertain (Shannon) and a local measure
of accuracy (Fisher). It is known to us that the Shannon-  +∞
where x ∈ R and −∞ f (x)dx = 1. By Eq. (3), we can see that
Fisher information plane is commonly based on permutation the division by f (x) is not convenient if f (x) → 0 at certain
entropy. However, in this paper, we propose a Shannon-Fisher x-values. In order to avoid this happening, some researchers
information plane based on the information entropy to ana- use real probability amplitudes f (x) = ψ 2 (x),27,37 which is a
lyze financial time series, which can give us more detailed, simpler form (no divisors), and present that FIM simply mea-
accurate, and clearer information on the classification of finan- sures the gradient content in ψ(x). Thus, this measure is called
cial stock markets. The information entropy is on the basis a “local” one. The Shannon entropy12 as a measure of “global
of novel complete quantification of the information, where character” is defined as
the information is accepted as an interpretation of the spatial  +∞
structures (which are different from the temporal structures H =− f (x) log f (x)dx. (4)
applied in the multiscale entropy). In the Shannon-Fisher −∞
information plane, since the FIM quantifies the amount of
For convenience, the alternative concept of entropy power12
organization or order in a system, while the Shannon entropy
can be used:
measures its degree of uncertainty or disorder, it is obvious
1 2H
that the larger the FIM is, the smaller the Shannon entropy is. N= e ; (5)
As a result, applying the Shannon-Fisher information plane 2π e
into financial time series is very effective, and we also gain this satisfies the so-called “isoperimetric inequality” IN ≥
insight of the inner dynamics of a system. For those complex D,35,38–40 and D represents the dimension of the space of the
systems, there might be still the possibility of increasing noise variable x.
and order in the sense of more and more intricate patterns in The probability density function f (x) can be estimated
the presence of increasing noise in some ways. according to the kernel density estimator technique,41,42 which
The rest part of this paper is arranged as follows. In approximates the density function as Eq. (6), and these forms
Sec. II, we present the details of the Shannon-Fisher infor- are shown in Refs. 43 and 44.
mation plane based on information entropy. Section III shows
M  
the results with two types of artificial time series: Autore- ˆfM (x) = 1 x − xi
K , (6)
gressive Fractionally Integrated Moving Average (ARFIMA) Mb i=1 b
models36 and Chebyshev map model. Application to financial
time series is illustrated in Sec. IV. Finally, we summarize and where b is the bandwidth, and the most successful method
make a conclusion. among all the current bandwidth selection methods, both
empirically and theoretically, is the solve-the-equation plug-
II. METHODOLOGY in method.45 M is the number of data. In addition, K(u) is the
kernel function; it should satisfy Eq. (7)
Consider the relevant Fisher- and Shannon-associated  +∞
quantities.34 Let f = ψ 2 be a probability density in K(u) ≥ 0 and K(u)du = 1. (7)
Rd (d ≥ 1). Fisher’s information associated with f (or to the −∞
probability amplitude ψ) is expressed as the (possible infinite) In the estimation procedure, kernel used is the Gaussian of
non-negative number I zero mean and unit variance. As a result, Eq. (6) can also be
 shown as
|∇f |2
I(f ) = dx , (1)
Rd f 1 M
−
(x−xi )2
fˆM (x) = √ e 2b2 . (8)
or in the light of the amplitudes M 2π b2 i=1

I(ψ) = dx(∇ψ · ∇ψ), (2) In our study, we let the probability distribution P = {pi : i =
Rd 1, 2, . . . , M }. In this case, the Shannon entropy is given by
where ∇ is the differential operator. This formula defines a 
M
convex, isotropic functional I, which was proposed by Fisher H =− pi · ln pi . (9)
(1925) in the context of statistical estimation,26 and it plays i=1
103107-3 Y. Wang and P. Shang Chaos 28, 103107 (2018)

Moreover, the “normalized Shannon entropy” is calculated as Here, m represents the embedding dimension, and the embed-
M
ding dimension m is strictly determined by using Takens
H 1 
S= = − pi · ln pi . (10) theorem.54
Hmax Hmax i=1 Step 2: Distance matrix construction: Define the dis-
tance matrix D = {dij } among vectors X (i) and X (j) for all
It is obvious that Hmax = ln(M ) is obtained by a uniform
1 ≤ i, j ≤ N − m + 1. The distance dij is shown by
probability distribution Pe = {pi = 1/M , ∀ i = 1, 2, . . . , M }.
In practice, discretizing the equivalent continuous expres- dij = max{|u(i + k) − u(j + k)|, 0 ≤ k ≤ m − 1}. (13)
sion in Eqs. (1) and (3), we obtain discrete forms of Fisher Step 3: Probability density estimation: The distribution
information that do not have the same value. Here, Eqs. (1) characteristics of all dij for 1 ≤ i, j ≤ N − m + 1 should be
and (3) can be used as a FIM-starting point in the discrete complete quantification of the information underlying the
case. The proposal of Frieden27 and of Ferri et al.46 take, at distance matrix D. In this study, the histogram approach is
starting point, Eq. (1). However, the concomitant discretiza- applied to estimating the empirical probability mass function
tions can be considered just approximations to what results of D. In the histogram approach, the interval of histogram
when using Eq. (3) as the starting point. We emphasize that graph is [max(dij ) − min(dij )]/M , where M represents the
a discrete normalized FIM, convenient for our purposes, is number of bin of histogram graph. This means that the range
defined as of the first histogram is {min(dij ), min(dij ) + [max(dij ) −

M −1 min(dij )]/M }, and the last histogram is in the range of
I = I0 [(pi+1 )1/2 − (pi )1/2 ]2 . (11) {max(dij ) − [max(dij ) − min(dij )]/M , max(dij )}. Then, we let
i=1 pi (i = 1, 2, . . . , M ) be the probability of each bin when
It has been extensively discussed that this discretization is the histogram has M bins. Therefore, the probability pi is
well behaved in a discrete environment,47 and this method shown as
for discrete distribution has been employed to do some the number of dij of the ith histogram
research.48,49 Here, the normalization constant I0 is given by pi = . (14)
(N − m + 1)2
⎧
⎪
⎨1 if pi∗ = 1 for i∗ = 1 or i∗ = M and Step 4: Calculation: For obtaining the probability dis-
I0 = pi = 0 ∀ i
= i∗ , (12) tribution P, we should calculate the normalized Shannon
⎪
⎩ entropy and FIM via Eqs. (10) and (11), respectively.
1/2 otherwise.
The use of Shannon-Fisher information plane was initially III. NUMERICAL RESULTS FOR ARTIFICIAL TIME
proposed by Vignat and Bercher,32 and it is applied into many SERIES
works.50–52 In this plane, we can find that if our system lies in To evaluate the effectiveness of the Shannon-Fisher infor-
a very ordered state, the normalized Shannon entropy S ∼ 0 mation plane method, two types of artificial time series gen-
and the FIM I ∼ 1. On the contrary, when the system is erated by ARFIMA models36 and Chebyshev map model are
shown by a very disordered state, we obtain that S ∼ 1 and tested, respectively.
I ∼ 0. The Shannon entropy (and better the Shannon length
e(S[f ]) , which is always non negative and has dimension of A. ARFIMA models
length)53 is a more appropriate measure of the global spread-
ing because it does not depend on any specific point of the The power-law auto-correlations in stochastic variables
interval definition of the distribution, and it gives a more sim- can be expressed by an ARFIMA process:55
ilar weight to all points of that interval. On the contrary, the ∞

local sensitivity of FIM for discrete probability distribution xt = an (d)xt−n + t , (15)
functions is reflected in the fact that the specific “i-ordering” n=1
of the discrete values pi must be seriously taken into account where t ∼ N(0, 1), d ∈ (0, 0.5) is a memory parameter corre-
in evaluating the sum in Eq. (11). The summands can be lated with Hurst exponent as hxx = 0.5 + d,56 and the weight
regarded as a kind of “distance” between two contiguous an (d) is given by
probabilities. For a different ordering of the pertinent sum-
an (d) = d(n − d)/[(1 − d)(n + 1)], (16)
mands would be lead to a different FIM-value, hereby its
local nature. Thus, the Shannon-Fisher information plane is where  is the Gamma function. The two-component
an effective tool to contrast global and local characteristic of ARFIMA process36 is defined as follows:
the probability distribution in research.
xt = [WXt + (1 − W )Yt ] + t , (17)
A. Algorithm yt = [(1 − W )Xt + WYt ] + νt , (18)
In order to put the Shannon-Fisher information plane ∞

based on information entropy into use, the method we have Xt = an (d1 )Xt−n , (19)
proposed above can be divided into four steps. n=1
Step 1: State-space reconstruction: Consider a original ∞

series {u(i), 1 ≤ i ≤ N}, form (N − m + 1) vectors X (i) by Yt = an (d2 )Yt−n , (20)
X (i) = {u(i), u(i + 1), . . . , u(i + m − 1)}, 1 ≤ i ≤ N − m + 1. n=1
103107-4 Y. Wang and P. Shang
FIG. 1. Shannon-Fisher information plane with ARFIMA time series as a
function of m from 6 to 15 for (M = 32, 64, 128, 256, 512), respectively.
where t and νt are two different i.i.d. Gaussian variables
with mean μ = 0 and variance σ 2 = 0. an (d1 ) and an (d2 )
are the weights represented by Eq. (15) through the scaling
parameters d1 and d2 ; d1 and d2 are in the range of (0, 0.5).
The parameter W ∈ [0.5, 1] quantifies the coupling strength
between the two processes xt and yt . When W = 1, xt and yt
show the total absence of correlation. The cross-correlation
between xt and yt decreases when W increases from 0.5 to 1.
In our simulations, we set as d1 = 0.1, d2 = 0.4, and W =
0.8. Thus, we can obtain two time series xt and yt with the
length of n = 6000 separately. Here, we apply the time series
xt into this work.
As shown in Fig. 1, we find that the FIM decreases, but
the normalized Shannon entropy increases with the increase of
the embedding dimension m for a given M on the whole. The
parameter M represents the number of possible states of the
system under study. The choice of m is determined by using
the false nearest-neighbor method.54 In order to make the
results more precise, we choose M = 2i for (i = 5, 6, 7, 8, 9).
This property is totally in accord with the characteristics of the
FIG. 2. Shannon-Fisher information plane with ARFIMA time series on various embedding dimensions m as a function of (M = 25 , 26 , 27 , 28 , 29 , 210 ).
Chaos 28, 103107 (2018)
Shannon-Fisher information plane. The FIM is an effective
tool to quantify the amount of organization or order in a sys-
tem, while the Shannon entropy is used to measure the degree
of disorder in dynamical systems. In addition, it is clear that
the bigger the embedding dimension m is, the more complex
the time series is, thus producing large amounts of Shannon
entropy. This means that the normalized Shannon entropy is
larger, the FIM is smaller.
Figures 2(a) and 2(b) depict the results of our proposed
method on various embedding dimensions m as a function
of M . That is to say, for every curve, it indicates that the
results obtained by M = 2i (i = 5, 6, 7, 8, 9, 10) separately for
a specific embedding dimension m. We find that the normal-
ized Shannon entropy becomes larger and larger, but the FIM
becomes smaller and smaller as the parameter M increases. As
we know, the probability distribution depends on the param-
eter M , and the bigger the parameter M is, the more chaotic
the time series is. As a result, large amounts of normalized
Shannon entropy can be obtained; on the contrary, the FIM
tends to be smaller. In practice, this finding also conforms
to the theory of entropy. Moreover, curves generally move to
the lower right. However, the curve may not be as regular as
results given by Fig. 2(b). From the picture of Fig. 2, with the
increase of M , we see that the normalized Shannon entropy
changes generally with the same amplitude, while the changes
of FIM have been smaller and smaller. As M = 29 (512) and
M = 210 (1024), the FIM changes little. Perhaps this may be
explained by the fact that the embedding dimension m has
much influence on the FIM when M is not big enough. This
means that the embedding dimension m is the main factor to
determine FIM instead of the parameter.
B. Chebyshev map
In this subsection, we utilize Chebyshev map model to
generate the artificial time series with the length of n = 6000.
It is defined as
xn+1 = cos(4 cos−1 xn ); (21)
we here set as the initial value x1 = 0.3.
103107-5 Y. Wang and P. Shang
FIG. 3. Shannon-Fisher information plane with Chebyshev map as a function
of m from 6 to 15 for (M = 32, 64, 128, 256, 512), respectively.
First of all, we analyze the results of Shannon-Fisher
information plane as a function of the embedding dimension
m by Fig. 3. In this paper, we choose the parameter m from the
point of view of phase space reconstruction. For a specific M ,
the result is obtained by m from 6 to 15. As above, we say, let
M be the number of possible states of the system. Similarly,
we choose M = 2i (i = 5, 6, 7, 8, 9), respectively. By observ-
ing Fig. 3, it shows that when the embedding dimension m is
from 6 to 15, the normalized Shannon entropy increases, and
the FIM decreases. This feature is absolutely in accord with
the characteristics of Shannon-Fisher information plane. The
FIM can measure the degree of order in dynamical systems,
but the Shannon entropy quantifies the degree of disorder in
dynamical systems. In some ways, FIM may have an opposite
relation to Shannon entropy.
In Fig. 4, it plots the Shannon-Fisher information plane
on various embedding dimensions as a function of M . We set
to M = 2i for (i = 5, 6, 7, 8, 9, 10). In general, the value of
m is determined by using the false nearest-neighbor method.
FIG. 4. Shannon-Fisher information plane with Chebyshev map on various embedding dimensions m as a function of (M = 25 , 26 , 27 , 28 , 29 , 210 ).
Chaos 28, 103107 (2018)
From the picture of Figs. 4(a) and 4(b), we conclude that
with the increase of the parameter M , the normalized Shannon
entropy increases, and the FIM decreases. It totally conforms
to the theory of entropy. Since M represents the number of
possible states of the system under study, when M increases,
the time series becomes more chaotic; thus, the normalized
Shannon entropy is bigger, and then the FIM is correspond-
ing to be smaller. This finding is also similar to the property
obtained by Fig. 2. In addition, when M = 25 (32) and M =
26 (64), we find that the FIM changes a lot as the embed-
ding dimension m increases. However, when M = 29 (512)
and M = 210 (1024), the FIM shows little change. It may sug-
gest that the embedding dimension m plays a leading role in
measuring the FIM when M is not big enough.
In short, applying the Shannon-Fisher information plane
based on information entropy into two types of artificial time
series, we find this method is very useful and available in
the study. Moreover, we can extract much information about
artificial time series from the Shannon-Fisher information
plane.
IV. NUMERICAL RESULTS FOR FINANCIAL TIME
SERIES
In this section, our proposed method is applied to the
analysis of the financial stock markets to verify the validity
of the method in real world data. Daily closing prices of ten
financial stock markets from October 28, 2002 to October 27,
2017 will serve as financial time series data sets.57 The ten
stock markets including four American markets, three Asian
markets, and three Australian markets are listed in Table I.
In this paper, we apply the Shannon-Fisher informa-
tion plane algorithm based on information entropy into the
log return series r(t) of financial time series to study the
relationship with financial stock markets.
r(t) = log[x(t)] − log[x(t − 1)], (22)
where x(t) indicates the closing price on day t.
It is known to us that M represents the number of possible
states of the system. Applying the artificial time series into the
103107-6 Y. Wang and P. Shang
TABLE I. Ten stock markets.
i Stock market index Abbreviation
1 Dow Jones Industrial Average DJI
2 NASDAQ 100 NDX
3 New York Stock Exchange NYA
4 Standard & Poor’s S&P500
5 S&P/CITIC300 CTSP300
6 Shanghai Stock Exchange SSEC
7 Taiwan Stock Exchange TWII
8 S&P/ASX 50 AFLI
9 ASX All Ordinaries AORD
10 S&P/ASX 200 AXJO
proposed method, we conclude that the embedding dimension
m has much influence on the result of FIM when M is not big
enough. This means that the FIM is mainly determined by the
embedding dimension m. Therefore, Fig. 5 shows the result of
Shannon-Fisher information plane obtained by M = 26 (64).
For each curve, we find that when m is from 6 to 15, the nor-
malized Shannon entropy of financial stock market increases,
and the corresponding FIM always decreases. This finding is
conformed with the results obtained by using artificial time
series. Since the embedding dimension m is increasing, the
financial time series tends to be more complex; thus, it can
produce bigger normalized Shannon entropy. In general, FIM
is used to measure the degree of order in dynamical systems,
so the FIM becomes smaller. This characteristic is completely
conformed to the properties of Shannon-Fisher information
plane. Interestingly, we observe that the financial stock mar-
kets seem to be divided into three groups. One group includes
TWII, SSEC, and CTSP300. It is obvious that these three
stock markets are from Asian. The second group consists of
three Australian stock markets: AFLI, AORD, and AXJO. We
know that the trends of daily closing price are very simi-
lar for these three stock markets. Thus, their properties may
be identical in some ways. The last group are DJI, NDX,
NYA, and S&P500; they belong to the American stock mar-
kets. As a result, applying the proposed method to classifying
m!, where m
FIG. 5. Shannon-Fisher information plane with financial time series for
M = 64 as a function of m from 6 to 15.
Chaos 28, 103107 (2018)
Country
USA
USA
USA
USA
China
China
Taiwan
Australia
Australia
Australia
FIG. 6. Shannon-Fisher information plane with financial time series for
m = 12 as a function of (M = 25 , 26 , 27 , 28 , 29 , 210 ).
the financial stock markets is very useful and effective. In
addition, we also see that the FIM of Asian stock markets is
smaller than the FIM obtained by the American stock markets.
It may suggest that DJI, NDX, NYA, and S&P500 are more
stable than the three Asian stock markets: TWII, SSEC, and
CTSP300. The stability of Australian stock markets AFLI,
AORD, and AXJO is between American stock markets and
Asian stock markets.
By observing Figs. 2 and 4, we find that the results
obtained by m from 11 to 15 are better than the results
obtained by m from 6 to 10. As a consequence, in order to
investigate the relationship between the parameter M and the
Shannon-Fisher information plane in the financial time series,
we let m be 12. The value of m is determined by using Tak-
ens’ embedding theorem.54 As we see in Fig. 6, the result
shows a trend of decline for a given embedding dimension m.
It demonstrates that with the increase of the parameter M , the
normalized Shannon entropy increases, and the corresponding
FIM decreases. This property is in accord with the character-
istics of the theory of entropy. Moreover, the results of TWII,
SSEC, and CTSP300 are grouped together. They all belong
to the Asian stock markets. The three Australian stock mar-
kets AFLI, AORD, and AXJO nearly coincide. The other four
stock markets DJI, NDX, NYA, and S&P500 are from Amer-
ica, and their normalized Shannon entropy values are smaller
than the others for a given M . Perhaps this may be explained
by the fact that the Australian stock markets and Asian stock
markets are more active than the American stock markets.58
In general, the classical Shannon-Fisher information
plane is based on the permutation entropy. However, in
this study, we propose the Shannon-Fisher information plane
through information entropy. It is well-known that when com-
puting the probability distribution P by using the permutation
entropy, it should satisfy the number of data N
represents the embedding dimension. There are almost 3900
closing daily prices for every stock market. Thus, m = 5 and
m = 6 are applied into the classical information plane based
on permutation entropy, respectively. The results are shown in
Fig. 7. As we see, Fig. 8 describes the result of our proposed
103107-7 Y. Wang and P. Shang
FIG. 7. The classical Shannon-Fisher information plane of financial time series with m = 5 and m = 6, respectively.
method. Considering the properties obtained by Figs. 5 and 6,
we set m = 12 and M = 26 (64). For a specific embedding
dimension m and M , we can calculate the normalized Shan-
non entropy and its FIM for every stock market. By Fig. 8, we
find that the financial stock markets are partitioned into three
groups. This finding is also found in Figs. 5 and 6. It is obvi-
ous that four American stock markets are grouped together.
The second group includes three Australian stock markets,
AFLI, AORD, and AXJO; their FIM values are nearly iden-
tical. The other Asian stock markets are grouped together. In
addition, we observe that American stock markets have the
biggest FIM, and Asian stock markets have the smallest FIM.
It indicates that the American stock markets is less active than
the other stock markets, including three Australian stock mar-
kets and three Asian stock markets. However, applying the
classical Shannon-Fisher information plane into financial time
series, whatever m = 5 or m = 6, it is hard to classify the
financial stock markets.
In summary, the proposed method is very useful and
effective to analyze the financial time series. Compared with
FIG. 8. Shannon-Fisher information plane based on information entropy with
financial time series for m = 12 and M = 64.
Chaos 28, 103107 (2018)
the classical Shannon-Fisher information plane, it can pro-
vide us more detailed, accurate, and clearer information on
the classification of financial stock markets to some extent.
V. CONCLUSION
In this paper, we propose a Shannon-Fisher information
plane based on the information entropy. In our simulation
experiments, for a given M , which represents the number of
possible states of the system, we find that with the increase of
the embedding dimension m, the normalized Shannon entropy
increases, and the FIM decreases. The choice of m is deter-
mined by using the false nearest-neighbor method. We know
that the probability distribution P mainly depends on the
parameter M . When analyzing the results of the Shannon-
Fisher information plane as a function of M , it is obvious that
as the parameter M increases, the curve shows the trend of
decline. In addition, when M is not big enough, we observe
that the FIM changes a lot. It may suggest that the embed-
ding dimension m plays a leading role in determining the FIM
when M is not so big. As m is from 11 to 15, the curve gen-
erally moves to lower right; however, the results obtained
by m from 6 to 10 may not be as regular. In the real series
experiments, the similar properties are obtained, that is, with
the parameters m and M increasing, the normalized Shannon
entropy increases, and the FIM tends to be smaller. Perhaps
this may be explained by the fact that as the parameters m and
M increase, the financial time series is more complex, thus
producing large amounts of normalized Shannon entropy, and
its corresponding FIM is smaller. As we know, the Shannon
entropy is a tool to measure a degree of disorder in dynamical
systems; on the contrary, the FIM can quantify the degree of
order in dynamical systems. These properties are totally con-
formed to the characteristics of Shannon-Fisher information
plane. Moreover, applying the proposed method into finan-
cial stock markets, we conclude that the ten stock markets are
accurately divided into three groups. The FIM values of four
American stock markets DJI, NDX, NYA, and S&P500 are
bigger than those of the other stock markets. These show that
these four American stock markets may be less active than
103107-8 Y. Wang and P. Shang
the Asian stock markets TWII, SSEC, and CTSP300 and the
Australian stock markets AFLI, AORD, and AXJO.
In summary, compared with the classical Shannon-Fisher
information plane based on the permutation entropy, our pro-
posed method can give us more exact information on the
classification of financial stock markets. We hope that this
method could be of great possibility and potentiality to be
used to study other fields.
ACKNOWLEDGMENTS
The financial support from the funds of the Fundamen-
tal Research Funds for the Central Universities (2018YJS178,
2018JBZ104), the China National Science (61771035), and
the Beijing National Science (4162047) is gratefully acknowl-
edged.
1 quantum mechanics using the principle of minimum Fisher information
J. F. Bercher and C. Vignat, “On minimum Fisher information distributions
with restricted support and fixed variance,” Inform. Sci. 179, 3832–3842
(2009).
2 ture,” Chem. Phys. Lett. 372, 28–34 (2003).
A. L. Berger, V. J. Della-Pietra, and S. A. Della Pietra, “A maximum
entropy approach to natural language processing,” Comput. Linguist. 22,
39–71 (1996).
3 32
S. A. Della-Pietra, V. J. Della-Pietra, and J. Lafferty, “Inducing features of
random fields,” IEEE Trans. Pattern Anal. Mach. Int. 19, 380–393 (1997).
4 33
B. R. Frieden, Physics from Fisher Information (Cambridge University
Press, 1998), Vol. 33, pp. 327–343.
5 34
Y. Y. Wang and P. J. Shang, “Analysis of financial stock markets through
multidimensional scaling based on information measures,” Nonlinear Dyn.
89, 1827–1844 (2017).
6 plane, an electron correlation tool,” J. Chem. Phys. 120, 8906–8912
H. Xiong and P. J. Shang, “Weighted multifractal cross-correlation analysis
based on Shannon entropy,” Commun. Nonlinear Sci. Numer. Simulat. 30,
268–283 (2016).
7 “Modeling long-range cross-correlations in two-component ARFIMA and
J. S. A. E. Fouda and W. Koepf, “Detecting regular dynamics from
time series using permutations slopes,” Commun. Nonlinear Sci. Numer.
Simulat. 27, 216–227 (2015).
8 Philos. Trans. R. Soc. Lond. 222, 309–368 (1922).
A. M. Lopes and J. A. T. Machado, “Analysis of temperature time-series:
Embedding dynamics into the MDS method,” Commun. Nonlinear Sci.
Numer. Simulat. 19, 851–871 (2014).
9 39
J. N. Xia and P. J. Shang, “Multiscale entropy analysis of financial time
series,” Fluct. Noise Lett. 11, 1250033 (2012).
10
Y. Yin and P. J. Shang, “Comparison of multiscale methods in the
stock markets for detrended cross-correlation analysis and cross-sample
entropy,” Fluct. Noise Lett. 13, 1450023 (2014).
11
Q. Tian, P. J. Shang, and G. C. Feng, “Financial time series analysis based
on information categorization method,” Physica A 416, 183–191 (2014).
12
C. E. Shannon, “A mathematical theory of communication,” Bell Syst.
Tech. J. 27, 379–423 (1948).
13
R. C. Hilborn, Chaos and Nonlinear Dynamics (Oxford University Press,
2000).
14
S. B. Sears, R. G. Parr, and U. Dinur, “On the quantum-mechanical kinetic
energy as a measure of the information in a distribution,” Israel J. Chem.
19, 165–173 (2013).
15
S. R. Gadre, “Information entropy and Thomas-Fermi theory,” Phys. Rev.
A 30, 620–621 (1984).
16
S. R. Gadre, S. B. Sears, S. J. Chakravorty, and R. D. Bendale, “Some
novel characteristics of atomic information entropies,” Phys. Rev. A 32,
2602–2606 (1985).
17
R. J. YáÌĹez, W. V. Assche, and J. S. Dehesa, “Position and momen-
tum information entropies of the D-dimensional harmonic oscillator and
hydrogen atom,” Phys. Rev. A 32, 3065–3079 (1994).
18
M. Hô, R. P. Sagar, V. H. Smith, Jr., and R. O. Esquivel, “Atomic infor-
mation entropies beyond the Hartree-Fock limit,” J. Phys. B 27, 5149
(1994).
19
M. Hô, R. P. Sagar, J. M. Pérez-Jordá, V. H. Smith, Jr., and R. O. Esquivel,
“A numerical study of molecular information entropies,” Chem. Phys. Lett.
219, 15–20 (1994).
20
Á. Nagy and R. G. Parr, “Information entropy as a measure of the quality
of an approximate electronic wave function,” Int. J. Quantum Chem. 58,
323–327 (1996).
Chaos 28, 103107 (2018)
21
N. L. Guevara, R. P. Sagar, and R. O. Esquivel, “Shannon-information
entropy sum as a correlation measure in atomic systems,” Phys. Rev. A
67, 5149–5153 (2003).
22
N. L. Guevara, R. P. Sagar, and R. O. Esquivel, “Information uncertainty-
type inequalities in atomic systems,” J. Chem. Phys. 119, 7030–7036
(2003).
23
N. L. Guevara, R. P. Sagar, and R. O. Esquivel, “Local correlation measures
in atomic systems,” J. Chem. Phys. 122, 084101 (2005).
24
C. C. Moustakidis and S. E. Massen, “The dependence of information
entropy of uniform Fermi systems on correlations and thermal effects,”
Phys. Rev. B 71, 045102 (2005).
25
K. D. Sen, “Characteristic features of Shannon information entropy of
confined atoms,” J. Chem. Phys. 123, 074110 (2005).
26
R. A. Fisher, “Theory of statistical estimation,” Proc. Camb. Phil. Soc. 22,
700–725 (1925).
27
B. R. Frieden, Science from Fisher Information (Cambridge University
Press, 2004).
28
B. R. Frieden and B. H. Soffer, “Lagrangians of physics and the
game of Fisher-information transfer,” Phys. Rev. E 52, 2274–2286
(1995).
29
M. Reginatto, “Erratum: Derivation of the equations of nonrelativistic
[Phys. Rev. A 58, 1775 (1998)],” Phys. Rev. A 60, 1730 (1999).
30
R. Nalewajski, “Information principles in the theory of electronic struc-
31
Á. Nagy, “Fisher information in density functional theory,” J. Chem. Phys.
119, 9401–9405 (2003).
C. Vignat and J. F. Bercher, “Analysis of signals in the Fisher-Shannon
information plane,” Phys. Lett. A 312, 27–33 (2003).
M. T. Martin, F. Pennini, and A. Plastino, “Fisher’s information and the
analysis of complex signals,” Phys. Lett. A 256, 173–180 (1999).
M. T. Martin, J. Perez, and A. Plastino, “Fisher information and nonlinear
dynamics,” Physica A 291, 523–532 (2001).
35
E. Romera and J. S. Dehesa, “The Fisher-Shannon information
(2004).
36
B. Podobnik, D. Horvatic, A. L. Ng, H. E. Stanley, and P. C. Ivanov,
FIARCH processes,” Physica A 387, 3954–3959 (2008).
37
R. A. Fisher, “On the mathematical foundations of theoretical statistics,”
38
A. Dembo, T. A. Cover, and J. A. Thomas, “Information theoretic inequal-
ities,” IEEE Trans. Inform. Theory 37, 1501–1518 (1991).
J. C. Angulo, J. Antolin, and K. D. Sen, “Fisher-Shannon plane
and statistical complexity of atoms,” Phys. Lett. A 372, 670–674
(2008).
40
R. O. Esquivel, J. C. Angulo, J. Antolin, J. S. Dehesa, S. Lopez-Rosa,
and N. Flores-Gallegos, “Analysis of complexity measures and informa-
tion planes of selected molecules in position and momentum spaces,” Phys.
Chem. Chem. Phys. 12, 7108–7116 (2010).
41
L. Devroye, A Course on Density Estimation (Birkhauser, 1987).
42
A. Janicki and A. Weron, Simulation and Chaotic Behavior of Stable
Stochastic Processes (Marcel Dekker, 1994).
43
L. Telesca, M. Lovallo, H. L. Hsu, and C. C. Chen, “Analysis of dynamics
in magnetotelluric data by using the Fisher–Shannon method,” Physica A
390, 1350–1355 (2011).
44
L. Telesca, M. Lovallo, A. Chamoli, V. P. Dimri, and K. Srivastava,
“Fisher–Shannon analysis of seismograms of tsunamigenic and non-
tsunamigenic earthquakes,” Physica A 392, 3424–3429 (2013).
45
V. C. Raykar and R. Duraiswami, “Fast optimal bandwidth selec-
tion for kernel density estimation,” in Proceedings of the SIAM Inter-
national Conference on Data Mining (Bethesda, Maryland, 2006),
pp. 524–528.
46
G. L. Ferri, F. Pennini, and A. Plastino, “LMC-complexity and various
chaotic regimes,” Phys. Lett. A 373, 2210–2214 (2009).
47
P. Sánchez-Moreno, R. J. YáÌĹez, and J. S. Dehesa, “Discrete densities and
Fisher information,” in Proceedings of the 14th International Conference
on Difference Equations and Applications (Bahçeşehir University Press,
2009), pp. 291–298.
48
B. A. Gonçalves, L. Carpi, and O. A. Rosso, “Time series characterization
via horizontal visibility graph and Information Theory,” Physica A 464,
93–102 (2016).
49
O. A. Rosso, F. Olivares, and A. Plastino, “Noise versus chaos in a causal
Fisher-Shannon plane,” Physics 7, 070006 (2015).
103107-9 Y. Wang and P. Shang Chaos 28, 103107 (2018)

50 54
M. G. Ravetti, L. C. Carpi, B. A. Gonçalves, A. C. Frery, and O. A. Rosso, F. Black and M. Scholes, “The pricing of options and corporate liabilities,”
“Distinguishing noise from chaos: Objective versus subjective criteria J. Polit. Econ. 81, 637–654 (1973).
55
using horizontal visibility graph,” PLoS One 9, e108004 (2014). J. R. M. Hosking, “Fractional differencing,” Biometrika 68, 165–176
51
O. A. Rosso, F. Olivares, L. Zunino, L. D. Micco, A. Aquino, A. Plastino, (1981).
56
and H. Larrondo, “Characterization of chaotic maps using the permuta- B. Podobnik and H. E. Stanley, “Detrended cross-correlation analysis:
tion Bandt-Pompe probability distribution,” Eur. Phys. J. B 86, 116–129 A new method for analyzing two nonstationary time series,” Phys. Rev.
(2013). Lett. 100, 084102 (2008).
52 57
F. Olivares, A. Plastino, and O. A. Rosso, “Contrasting chaos with noise via See http://finance.yahoo.com for daily closing prices of financial stock
local versus global information quantifiers,” Phys. Lett. A 376, 1577–1583 markets.
58
(2012). S. H. Bian and P. J. Shang, “Refined two-index entropy and multiscale
53
C. E. Shannon and W. Weaver, The Mathematical Theory of Communica- analysis for complex system,” Commun. Nonlinear Sci. Numer. Simul. 39,
tion (University of Illinois Press, Urbana, 1947). 233–247 (2016).