Physica A 444 (2016) 722–731

Contents lists available at ScienceDirect

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

Multifractal detrended fluctuation analysis based on fractal

fitting: The long-range correlation detection method for
highway volume data
Meifeng Dai ∗ , Jie Hou, Dandan Ye
Nonlinear Scientific Research Center, Faculty of Science, Jiangsu University, Zhenjiang, Jiangsu, 212013, PR China

highlights
• Investigating the traffic time series for volume data observed on the Guangshen highway.
• Proposing multifractal detrended fluctuation analysis based on fractal fitting (MFDFA-FF) to detect long-range correlations of time
series.
• Fractal fitting can get a better detrend effect than polynomial fitting.
• MFDFA-FF is a more appropriate method for time series to detect the long-range correlation and multifractality.

article info abstract

Article history: In this paper, we investigate the traffic time series for volume data observed on the
Received 20 April 2015 Guangshen highway. We introduce a multifractal detrended fluctuation analysis based
Received in revised form 9 September 2015 on fractal fitting (MFDFA-FF), which is one of the most effective methods to detect long-
Available online 23 October 2015
range correlations of time series. Through effective detecting of long-range correlations,
highway volume can be predicted more accurately. In order to get a better detrend effect,
Keywords:
we use fractal fitting to replace polynomial fitting in detrend process, the result shows
Fractal interpolation
Fractal fitting
that fractal fitting can get a better detrend effect than polynomial fitting and the MFDFA-
MFDFA-FF FF method can achieve a more accurate research result. Then we introduce the Legendre
Legendre spectrum spectrum to detect the multifractal property characterized by the long-range correlation
Long-range correlation and multifractality of Guangshen highway volume data.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction

Fractal was proposed by Mandelbrot in 1975 [1]. So far, the mathematical definition of fractal is not strict, which only
has a descriptive definition. Fractal is a rough or fragmented geometrical shape that can be subdivided into parts, each of
which is a reduced-size copy of the whole, a fractal system is usually described by a scale invariant parameter called fractal
dimension [2]. Although there is not a unified, strict definition about ‘‘fractal’’, it has received extensive attention of many
disciplines as a new analysis method. Relevant empirical studies have made many valuable results. Using fractal theory to
study the nature of the complex system has become one of the most widely discussed fields recently.
Many fractals arising in nature have a far more complex scaling relation and require a set of parameters to describe
characteristics of complex systems that are known as multifractals. It is well-known that the formation, occurrence and

∗ Corresponding author.
E-mail address: daimf@mail.ujs.edu.cn (M. Dai).

http://dx.doi.org/10.1016/j.physa.2015.10.073
0378-4371/© 2015 Elsevier B.V. All rights reserved.
 M. Dai et al. / Physica A 444 (2016) 722–731 723

development process of many natural phenomena are not strictly random, but have correlation and predictability. Due to
the difficulty of analysis in achieving long-range correlations of nonstationary series, Peng et al. [3] proposed an approach
to analyze the DNA sequences which are known as detrended fluctuation analysis (DFA). Stanley et al. [4–6] researched
deeply on detrended fluctuation analysis (DFA). DFA method has been applied to many fields, such as DNA structure [7],
records of the heart [8] and meteorological data [9]. Although the DFA method is widely used to achieve monofractal
scaling properties, it cannot properly describe multifractal of time series. Based on the DFA method, Kantelhardt et al. [10]
introduced the multifractal detrended fluctuation analysis (MFDFA) for the multifractal characterization of nonstationary
time series. MFDFA can be used to establish a scientific and reasonable prediction, forecasting model. The fractals have been
applied to studying traffic time series in the last several years [11–13]. As a powerful technique, MFDFA has so far been
applied to the time series analysis in traffic fields [14]. For traffic time series, to better estimate the long-range correlations
and its multifractal properties, MFDFA based on fractal fitting called MFDFA-FF is proposed.
Another fundamental concern is to study the possible fractal structure of long-range correlation, the Legendre spectrum
was extensively studied. Dai et al. [15] used Legendre spectrum and Hölder exponent to analyze multifractal and singularity
of Guangshen highway volume data. In this paper, we use MFDFA-FF method to achieve the Hurst exponent h(q) which
related to the classical multifractal scaling exponent τ (q). MFDFA-FF is a more appropriate method for time series to
detect the long-range correlation and multifractality. Note that unlike the previous paper, based on MFDFA-FF method,
we can achieve the value of Hurst exponent more accurately to detect the long-range correlation and generate the Legendre
spectrum more precisely to analyze the multifractal property.
The organization of this paper is as follows. Firstly, we introduce the fractal interpolation and fractal fitting in Section 2
and multifractal detrended fluctuation analysis based on fractal fitting (MFDFA-FF) in Section 3. Secondly, the detrended
ability analysis with MFDFA-FF shows that MFDFA-FF can achieve a better detrend effect than MFDFA in Section 4. Thirdly,
we introduce the multifractal spectrum in Section 5. Then we analyze the long-range correlation and multifractal of
Guangshen highway volume data in Section 6. Finally, we make a conclusion.

2. The fractal interpolation and fractal fitting

Fractal interpolation has been developed as an alternative interpolation technique suitable for data with inherent fractal
structure. Given interpolation points {(xi , yi )}Ni=0 , fractal interpolation is based on the theory of iterated function systems
(IFS) that is convenient for fitting physical or experimental data [16,17]. In contrast to fractal interpolation, traditional
interpolation is built on elementary functions such as polynomials. Fractal interpolation function has been used in fitting
the unsmoothed curve which shows the superiority to other functions. Fractal interpolation function is becoming a research
field for approximation theory of functions.
A fractal interpolatory scheme can be introduced in the following way [16].
Let I = [a, b]. a = x0 < x1 < · · · < xN = b is a partition for I and N ≥ 2. y0 , y1 , . . . , yN are arbitrary real numbers. Let
K = I × R and Ii = [xi−1 , xi ], i = 1, 2, . . . , N. Li is a compression homeomorphism for I → Ii

Li (x0 ) = xi−1 , Li (xN ) = xi , (1)

and for any u1 , u2 ∈ I
|Li (u1 ) − Li (u2 )| ≤ li · |u1 − u2 |,
where 0 < li < 1.
Fi is a continuous function for K → R
Fi (x0 , y0 ) = yi−1 , Fi (xN , yN ) = yi , (2)
and for any u ∈ I , v1 , v2 ∈ R
|Fi (u, v1 ) − Fi (u, v2 )| ≤ qi · |v1 − v2 |,
where 0 ≤ qi < 1.
wi is a mapping K → K
Li (x)
   
x
wi = , i = 1, 2, . . . , N . (3)
y Fi (x, y)

So far, IFS {K ; wi , i = 1, 2, . . . , N } has been constructed. Specially, when Li (x) and Fi (x, y) are linear functions, Eq. (3) is as
follows,
      
x ai 0 x ei
wi = + , i = 1, 2, . . . , N , (4)
y ci di y fi

satisfying
Li (x0 ) = xi−1 , Li (xN ) = xi , Fi (x0 , y0 ) = yi−1 , Fi (xN , yN ) = yi , (5)
724 M. Dai et al. / Physica A 444 (2016) 722–731

Fig. 1. (a) is the accumulated

 deviation curve on the Guangshen highway on 4 April 2008. (b) is the fractal fitting curve, where we achieve the minimum
(φ(xi ) − yi )2 when d1 = −0.0383, d2 = 0.2489, d3 = 0.0900, d4 = −0.1617.
N
value 638.6741 of i=0

for i = 1, 2, . . . , N. Solving Eqs. (1)–(5) we obtain

xi − xi−1 xN xi−1 − x0 xi
ai = , ei = ,
xN − x0 xN − x0
yi − yi−1 yN − y0 xN yi−1 − x0 yi xN y0 − x0 yN
ci = − di , fi = − di
xN − x0 xN − x0 xN − x0 xN − x0

where the real numbers ai , ei , ci , fi (i = 1, 2, . . . , N ) are completely determined by the interpolation points, and di (i =
1, 2, . . . , N ) are variables of the affine transformations satisfying |di | < 1, which guarantees that the IFS is hyperbolic with
respect to an appropriate metric. The transformations wi are shear transformations: line segments parallel to the y-axis are
mapped to line segments parallel to the y-axis contracted by the vector di . For this reason, di (i = 1, 2, . . . , N ) are called
i=1 ωi (G) of the aforementioned IFS is the graph of a
N
vertical scaling (or contractivity) factors. So that the attractor G =
continuous function f : [x0 , xN ] → R that passes through the interpolation points. This function is aforementioned fractal
interpolation function (FIF) corresponding to these points. For a set of points {(xj , yj )}M j=0 , take the appropriate points as
interpolation points {(xi , yi )}Ni=0 (N ≤ M ) and use fractal interpolation to obtain fractal interpolation function f , fractal
fitting is a process that adjusts the vertical scaling
factors di to change the value of function f , and to look for the most
(f (xi ) − yi )2 .
N
appropriate di to achieving the minimum value of i=0
We use the accumulated deviation data observed on the Guangshen highway over a period of one day, 4 April 2008, which
include 288 data points. The equation of accumulated deviation Y (k) is Y (k) ≡ i=1 [yi − ⟨y⟩] where k = 1, . . . , N. The
k
raw data we obtained are the number of cars passed the Meiling checkpoint in every five minutes. We use five interpolation
points (x1 , y1 ), (x73 , y73 ), (x145 , y145 ), (x217 , y217 ), (x288 , y288 ) and find four optimal vertical scaling vectors d1 , d2 , d3 , d4 . The
accumulated deviation data curve is showed in Fig. 1(a) and fractal fitting curve is showed in Fig. 1(b), where we achieve
(f (xi ) − yi )2 when d1 = −0.0383, d2 = 0.2489, d3 = 0.0900, d4 = −0.1617.
N
the minimum value 638.6741 of i =0
 M. Dai et al. / Physica A 444 (2016) 722–731 725

3. Multifractal detrended fluctuation analysis based on fractal fitting (MFDFA-FF)

Consider the time series x = {xi }i=1,2,...,N that is a non-stationary with its statistical moments (e.g. the mean
value, variance) may change with time. Non-stationary time series often produces unreliable results and leads to poor
understanding and forecasting. The solution to the problem is that we transform the time series from the non-stationary
to stationary. There are two methods, differencing and detrending procedures. On the one hand, if the non-stationary
process is a random walk with or without a drift, it is transformed to stationary process by differencing. On the other
hand, if time series exhibits a deterministic trend, the spurious results can be avoided by detrending. MFDFA-FF is one
of those techniques that use detrending procedure to deal with the non-stationary time series. To analyze large-sized time
series, a division–integration procedure is used by many methods, including the detrended fluctuation analysis, permutation
entropy, cross-sample entropy, superinformation, etc. [3,18–20]. MFDFA-FF, which is a practice of the division–integration
procedure, detects the long-range correlations in the non-stationary time series. The theoretical principles are constructed
on the random walk theory and the detrended variance. Next, we introduce MFDFA-FF.
(1) Let the profile of {xi } be as follows.
k

X (k) ≡ [xi − ⟨x⟩], (6)
i=1

where k = 1, . . . , N.
(2) The profiles X (k) (k = 1, 2, . . . , N ) are divided into Ns ≡ [N /s] non-overlapping segments of equal length s, where [·]
is a lower integer sign.
(3) For each segment v (v = 1, 2, . . . , Ns ) we fit the integrated time series by fractal fitting (FF) function xv to eliminate
fractal trend of the profile function. xv (j) means the value of fractal fitting in each segment v . The local detrended variance
function in each segments v (v = 1, . . . , Ns ) is calculated as
s
1
F 2 (s, v) = {X [(v − 1)s + j] − xv (j)}2 . (7)
s j =1

We give the following explanation. Traditional MFDFA method fits the integrated time series by polynomial functions,
the degree of fitting polynomial can be varied in order to eliminate linear, quadratic, cubic or higher order trends of the
profile function and accordingly the procedure is said to be the MFDFA1, MFDFA2, MFDFA3, MFDFAn analysis. MFDFA
based on fractal fitting is said to be MFDFA-FF analysis.
(4) Average over all segments to obtain the detrended variance function
 1/q
Ns
1  q/2
Fq (s) = [F (s, v)]
2
, (8)
Ns v=1

where q can be any real value.

(5) The above computation is repeated on different scale s to provide a relationship between Fq (s) and s. For considered
time series, Fq (s) increases with s asymptotically as

Fq (s) ∼ sh(q) , (9)

where the Hurst exponent h(q) characterizes the long-range correlation property of variable.
We give the following explanation. When q = 2, the Hurst exponent h(q) of DFA has the following properties [21]: if
h(q) = 0.5 or Fq (s) oscillates around zero, there is no correlation; if h(q) > 0.5, there is a positive power-law correlation,
which indicates that time series has a positive correlation; if h(q) < 0.5, time series is anti-correlation. If h(q) varies with
q, time series presents multifractal correlation. If h(q) is a constant, time series presents monofractal correlation.

4. MFDFA-FF detrended ability analysis

In this section, we compare MFDFA-FF with MFDFAn (n = 1, 2, . . .) to analyze the detrended effect. MFDFA is a very
famous method to obtain the Hurst exponent which characterizes the long-range correlation property of variable. MFDFA-FF
is improved from MFDFA. In order to explain MFDFA-FF is a more appropriate method than MFDFA for time series to research
the long-range correlation and multifractal property, we compared MFDFA-FF with MFDFAn from three respects. Figs. 2 and
3 are the fitting effect between MFDFA-FF and MFDFAn (n = 1, 2, . . .). In order to observe the fitting effect clearly, we
choose the Guangshen highway over a period of half day and one day, 4 April 2008, 144 points and 288 points, respectively.
From Figs. 2 and 3 we can observe that fractal fitting can get better effect than polynomial fitting, using measure criterion

i=0 (f (xi ) − yi ) . Obviously, the smaller V is, the stronger the fitting effect it has. The values of V obtained by linear,
N 2
V =
quadratic, cubic polynomial fitting are 3559.1, 1182.9, 792.5 on 144 data points. The value of V obtained by fractal fitting
726 M. Dai et al. / Physica A 444 (2016) 722–731

Fig. 2. Take the Guangshen highway over a period of half day, 4 April 2008, 144 points data. (a) shows the accumulated deviation curve of original data
and linear, quadratic, cubic polynomial fitting curve on the Guangshen highway over a period of half day. (b) shows the accumulated deviation curve of
original data and fractal fitting curve. We find that fractal fitting has the best fitting effect of all.

is 206.5. Similarly, the values of V obtained by linear, quadratic, cubic polynomial fitting are 4784.9, 3390.5, 1112.4 on 288
data points and the value of V obtained by fractal fitting is 638.7. According to the above-mentioned two comparisons, we
observe that the value of V obtained by fractal fitting is the smallest both on 144 and 288 data points and fractal fitting has
better detrend effect than polynomial fitting. Due to some traffic phenomena occur periodicity, we research the detrend
ability of cyclical trends between polynomial fitting and fractal fitting. Firstly, add a sine trend for time series, such as
X ′ (i) = X (i) + 10p sin(2π i/q), let p = 2.5, q = 64, Fig. 4 shows that there are some volatility in both polynomial fitting
and fractal fitting, but MFDFA-FF obtains smaller values of Fq (s) than MFDFAn (n = 1, 2, 3) in every length s, it implies
that MFDFA-FF has stronger detrend effect than MFDFAn (n = 1, 2, 3) for cyclical trends. In order to achieve more accurate
comparison, we fit the slope line of log–log plot for each method when q = 2. Fig. 5 is the fitting slope when q = 2 between
MFDFA-FF and MFDFAn (n = 1, 2, 3). The value of fitting slope is the Hurst exponent h(q) when q = 2, the hurst exponents
of MFDFA-FF, MFDFA1, MFDFA2, MFDFA3 are 0.8878, 0.5940, 0.6650 and 0.7638, respectively. With the larger degree of
fitting polynomial, the hurst exponent increased. MFDFA-FF achieves the relative maximum value. The degree of fitting
polynomial can be varied in order to eliminate linear, quadratic, cubic or higher order trends of the profile function. The
highway volume data have a complicated trend. So as the degree of fitting polynomial is getting larger, the value of hurst
exponent is more approaching to the real value. The advantage of MFDFA-FF is that it can achieve a more accurate value
no matter what trend the time series has. We use the standard deviation V of fitting to measure the volatility of MFDFA-FF
and MFDFAn (n = 1, 2, . . .), the values of V obtained by 4 methods are as follows, VMFDFA-FF = 0.2903, VMFDFA1 = 0.3537,
VMFDFA2 = 0.2878, VMFDFA3 = 0.3488. From the comparison of values, MFDFA-FF has a quite small value of V and the strongest
ability of detrend. In this additive trend we use sine function to add a cyclical trend, we find MFDFA2 has a low volatility.
But there are various trends in the actual situation and it is difficult to judge which trends it has, so MFDFA-FF can achieve
a better detrend effect no matter what trend the time series has.
Next we compare MFDFA-FF with Rescaled Range Analysis (R/S Analysis) to quantify long-range correlations. Rescaled
Range Analysis [22] is one of methods to detect long-range correlations. Based on a lot of researches, Hurst proposed
the R/S Analysis method. R/S Analysis detects long-range correlations by calculating hurst exponent. R/S Analysis method
can be expressed as (R/S )n = C · nH , where n is the number of observed samples, R is the rescaled range, S is the
standard deviation and H is the Hurst exponent. As a general rule, using the following equation as the research object:
 M. Dai et al. / Physica A 444 (2016) 722–731 727

Fig. 3. Take the Guangshen highway over a period of one day, 4 April 2008, 288 points data. (a) shows the accumulated deviation curve of original data
and linear, quadratic, cubic polynomial fitting curve on the Guangshen highway over a period of one day. (b) shows the accumulated deviation curve of
original data and fractal fitting curve. We find that fractal fitting also has the best fitting effect of all.

Fig. 4. The log–log plot of qth statistical moment Fq(S ) versus S based on MFDFA-FF, MFDFA1, MFDFA2, MFDFA3, respectively.
728 M. Dai et al. / Physica A 444 (2016) 722–731

Fig. 5. The slope fitting line of F2 (s) versus S based on MFDFA-FF, MFDFA1, MFDFA2, MFDFA3, respectively.

Fig. 6. The slope fitting line based on MFDFA-FF and R/S Analysis.

log(Rn /Sn ) = log(C ) + H · log(n), the value of slope H is the Hurst exponent. Fig. 6(b) shows that the log(R/S ) points around
the slope fitting line have a sense of drift especially on small scale but Fig. 6(a) shows that the points around the slope
fitting line have a comparatively stable distribution. In conclusion, MFDFA-FF is a more appropriate method for time series
to research the long-range correlation and multifractal property.

5. Multifractal spectrum

The power-law behavior of detrended variance function in MFDFA-FF method indicates the presence of scale-invariance
in the correlation of variable. The monofractal variable can be described by one scaling exponent. However, many variables
do not exhibit a simple monofractal scaling behavior. Thus several exponents or the exponent spectrum are required for
the full description of the scaling behavior of variables. Multifractal spectrum is an alternative candidate to describe the
multifractal structure of variables. A standard formalism considers each interval at a certain scale s and estimates the local
Hurst exponent in each interval for normalized variables [23,24],
 
Ns
1 
τ (q) = log (s)
qhv
, (10)
log(s) v=1
 M. Dai et al. / Physica A 444 (2016) 722–731 729

Fig. 7. The volume time series observed on the Guangshen highway over a period of three month, from 1 February in 2008 to 30 April in 2008.

where Ns ≡ [N /s], and [·] is a lower integer sign. Typically, if s ∼ 2−n when the whole interval [1, N ] is mapped to [0, 1],
2n
τ (q) ∼ −1/n ∗ log2 [ v=1 2−nqhv ]. hv represents the local Hurst exponent that is estimated in interval v as,

log[F 2 (s, v)]

hv = , (11)
2 log(s)
where F 2 (s, v) is defined in Eq. (7) as the local detrended variance. According to the standard multifractal formalism, the
classical multifractal scaling exponent τ (q) is related to the Hurst exponent via MFDFA-FF as
τ (q) = qh(q) − 1. (12)
The Legendre spectrum is defined as the Legendre transform of the function τ (q) = lim infN −→∞ τ (q). The singularity
strength or the Hölder exponent α and the singularity spectrum f (α) are related to τ (q) as Ref. [25],
α(q) = d(τ (q))/dq, (13)
and
f (α) = qα(q) − τ (q). (14)
We can easily infer that α(q) = h(q) + qh′ (q) and f (α) = q[α(q) − h(q)] + 1. By construction, the singularity spectrum
is always a concave function. A single Hölder exponent α denotes monofractality, while in the multifractal case, different
parts of the structure are characterized by different value of α , leading to the existence of the multifractal spectrum.

6. Long-range correlation and multifractal analysis of Guangshen highway volume data

In this section, we use MFDFA-FF method to investigate the existence of long-range correlation and multifractal in the
data observed on the Guangshen highway over a period of three month, from 1 February in 2008 to 30 April in 2008 (see
Fig. 7). The raw data we obtained are the number of cars passed the Meiling checkpoint in every five minutes.
We use several length s to divide N into Ns ≡ [N /s], in each segment v (v = 1, 2, . . . , Ns ), we take a point every s/4
interval from (v− 1)s + 1 to v s and we will have five points in each segment v , we use these five points to be the interpolation
points of fractal fitting and achieve the fractal interpolation functions (FIF). According to observed slope of the log–log plot
(see Fig. 8), we achieve the generalized hurst exponent h(q). Fig. 9 presents the relationship between q and hurst exponent
h(q), h(q) is a monotonously decreasing function of q and the corresponding series presents multifractal correlation. When
q = 2, the hurst exponent, h(q) = 0.8878, presents that the volume time series is a long-range correlated series and has a
positive power-law correlation. Fig. 10 displays the Legendre spectra f (α) for the volume time series from the Guangshen
highway. The spectrum displays the commonly observed bell shape, which means that the volume data exhibit multifractal
behavior. Define α0 as the value of α when f (α) takes the maximum, 1α = αmax − αmin and 1f = f (αmin ) − f (αmax )
[14,26], where αmin is the minimum value of α and αmax is the maximum one; 1α is the width of the multifractal spectrum
of traffic data, the larger the 1α is, the stronger the multifractality it has; 1f is the difference between the fractal dimension
of the maximum probability subset and that of the minimum probability subset, 1f > 0 represents that the chance of the
traffic data lying at higher site is more than that at lower site and vice versa. From Fig. 10, we calculate some characteristic
quantities: 1α ≈ 0.4946, 1f ≈ 0.2117 and α0 ≈ 0.9467. The range of α > α0 is larger than that of α < α0 , reflecting
that the traffic data lying at higher site has more chance than that at lower site on highway traffic, and the traffic flow of
Guangshen highway is low in the most time of one day.
730 M. Dai et al. / Physica A 444 (2016) 722–731

Fig. 8. The log–log plot of qth statistical moment F (q) versus q when q = −5, q = 0, q = 2, q = 5.

Fig. 9. The slope plot of the hurst exponent h(q) versus q.

Fig. 10. The Legendre spectrum f (α) for the volume time series observed on the Guangshen highway over a period of three month.

7. Conclusion

In this paper, we estimated the long-range correlations and existence of multifractal in Guangshen highway volume
data. According to the detrended ability comparison between polynomial fitting and fractal fitting, we can observe that
fractal fitting has better detrend effect than polynomial fitting, and MFDFA-FF has better effect to research the long-range
correlation characteristics and multifractal property. Researching the degree of precision of long-range correlation can
allow us to better forecast the future trends and avoid the estimation errors. We can replace MFDFA with MFDFA-FF to
detect long-range correlations in order to predict highway volume more accurately. We found that the Hurst exponent
h(q) = 0.8878, the volume time series is a long-range correlated series and has a positive power-law correlation. In the
multifractal spectrum analysis, the multifractality is characterized in volume time series. In this paper, we observed that
MFDFA-FF is a more appropriate method for time series to detect the long-range correlation and multifractality so we can
use this method on various time series.
 M. Dai et al. / Physica A 444 (2016) 722–731 731

Acknowledgment

Research is supported by the Humanistic and Social Science Foundation from Ministry of Education of China (Grant
14YJAZH012).

References

[1] B.B. Mandelbrot, Fractals: Form, Chance, and Dimension, W.H. Freeman and Company, San Francisco, 1977.
[2] B.B. Mandelbrot, The Fractal Geometry of Nature, Freeman, New York, 1982.
[3] C.K. Peng, S.V. Buldyrev, S. Havlin, M. Simons, H.E. Stanley, A.L. Goldberger, Mosaic organization of DNA nucleotides, Phys. Rev. E 49 (1994) 1685–1689.
[4] D. Horvatic, H.E. Stanley, B. Podobnik, Detrended cross-correlation analysis for non-stationary time series with periodic trends, Europhys. Lett. EPL 94
(2011) 18007–18012.
[5] Z. Chen, K. Hu, P. Carpena, P. Bernaola-Galvan, H.E. Stanley, P.Ch. Ivanov, Effect of nonlinear filters on detrended fluctuation analysis, Phys. Rev. E 71
(2005) 011104.
[6] Z. Chen, P.Ch. Ivanov, K. Hu, H.E. Stanley, Effect of nonstationarities on detrended fluctuation analysis, Phys. Rev. E 65 (2002) 041107.
[7] P. Allegrini, M. Barbi, P. Grigolini, B.J. West, Dynamical model for DNA sequences, Phys. Rev. E 52 (1995) 5281–5296.
[8] T.H. Makikallio, T. Seppanen, Dynamic analysis of heart rate may predict subsequent ventricular tachycardia after myocardial infraction, Am. J. Cardiol.
80 (1997) 779–783.
[9] R.B. Govindan, D. Vjushin, Long-range correlations and trends in global climate models: comparison with real data, Physica A 294 (2001) 239–248.
[10] J.W. Kantelhardt, S.A. Zschiegner, E. Koscielny-Bunde, S. Havlin, A. Bunde, H.E. Stanley, Multifractal detrended uctuation analysis of nonstationary
time series, Physica A 316 (2002) 87–114.
[11] P. Shang, Multi-fractal analysis of highway traffic data, Chin. Phys. 16 (2007) 365–373.
[12] X. Li, P. Shang, Multifractal classification of road traffic flows, Chaos Solitons Fractals 31 (2007) 1089–1094.
[13] P. Shang, S. Kanlae, The application of Holder exponent to traffic congestion warning, Physica A 370 (2006) 769–776.
[14] P. Shang, Y. Lu, S. Kamae, Detecting long-range correlations of traffic time series with multifractal detrended fluctuation analysis, Chaos Solitons
Fractals 36 (2008) 82–90.
[15] M. Dai, C. Zhang, D. Zhang, Multifractal and singularity analysis of highway volume data, Physica A 407 (2014) 332–340.
[16] M.F. Barnsley, Fractal functions and interpolation, Constr. Approx. 2 (1986) 67–92.
[17] M.F. Barnsley, Fractal Everywhere, Academic press, New York, 1988.
[18] C. Bandt, B. Pompe, Permutation entropy: A natural complexity measure for time series, Phys. Rev. Lett. 88 (2002) 174102.
[19] S. Pincus, B. Singer, Randomness and degrees of irregularity, Proc. Natl. Acad. Sci. 93 (1995) 2083–2088.
[20] R. Bose, S. Chouhan, Super-information: A novel measure of information useful for DNA sequences, Phys. Rev. E 83 (2011) 051918.
[21] J. Muzy, R. Baile, E. Bacry, Random cascade model in the limit of infinite integral scale as the exponential of a nonstationary 1/f noise: Application to
volatility fluctuations in stock markets, Phys. Rev. E 87 (2013) 042813.
[22] H.E. Hurst, Long-term storage capacity of reservoirs, Trans. Am. Soc. Eng. 116 (1951) 770–799.
[23] P. Loiseau, C. Médigue, P. Goncalves, N. Attia, S. Seuret, F. Cottin, D. Chemla, M. Sorine, J. Barral, Large deviations estimates for the multiscale analysis
of heart rate variability, Physica A 391 (2012) 5658–5671.
[24] J. Barral, P. Goncalves, On the estimation of the large deviations spectrum, J. Stat. Phys. 144 (2011) 1256–1283.
[25] J. Feder, Fractals, Plenum Press, New York, 1988.
[26] X. Sun, H. Chen, Z. Wu, Y. Yuan, Multifractal analysis of Hang Seng index in Hong Kong stock market, Physica A 291 (2001) 553–562.