Multi-scale correlation and causality of stock exchanges

Neeraj, Prasanta K. Panigrahi

Indian Institute of Science Education and Research Kolkata, Mohanpur, 741 246, India.

Abstract

We study the multi-scale temporal correlations and causality connections between the New York Stock Exchange
(NYSE) and Bombay Stock Exchange (BSE) monthly average closing price indices for a period of 300 months, encompassing the time period of the liberalisation of the Indian economy, leading to its gradual exposure to global markets.
Clearly indentifiable 1, 2 and 3 year non-stationary periodic modulations in NYSE have been found to drive commensurate changes in BSE price index. Interestingly, at 1 year time scale, the two exchanges are phase locked only during
the turbulent times, while at a scale of 3 year, in-phase nature is observed for a longer time frame, starting with a small
phase lag which gradually leads to complete phase locked behaviour. The 2 year time scale embodies the characteristics
of both these time scales, and thus acts as transition scale. Although, expectedly, NYSE is found to Granger cause BSE
in case of monthly averaged normalised data, with a lag of 9 months, but average behaviour of the two exhibited no sign
of causality. However, counter-intuitively the observed Granger causality of fluctuations at smaller scales, revealed BSE
behaviour getting reflected in the NYSE index fluctuations, after a time lag.
Keywords: Stock market, Mean-reversion, Probability density, Wavelet transform, Wavelet coherence and
Toda-Yamomoto Granger causality (TYGC)

1. Introduction
Stock markets can exhibit intricate inter-relationships
at different time scales, not easily discernible through traditional methods of analysis [1, 2]. Unraveling these connections have become even more complex with the advent
of globalization. The degree of intricacy can vary from pair
to pair, depending on the inter-dependencies arising from
the requirements of the two economies, global factors and
also possibly on the size of economies of the two countries.
For a relatively insular economy like that of India, the
nature of the stock markets correlations with other countries can be subtle. Generally, insularity is expected to
protect an economy from short term global trends, while
possibly impacting it in long run. Hence, it is of deep interest to analyse the correlations between open and insular
economies to identify their short and long-term behaviour.
This will throw light on the nature of interaction between
open and protected economies, leading to a better understanding of the global economy as a whole.
The fact that, the Indian economy progressively opened
itself post 1991 and it still has not been fully integrated
with the open economies, makes it particularly interesting
to study its correlation and causal connections with that of

Email addresses: neerajiiser@gmail.com ( Neeraj),

pprasanta@iiserkol.ac.in (Prasanta K. Panigrahi)
Preprint submitted to elsevier

US, the largest open economy. For this purpose, we carry

out a systematic analysis of New York Stock Exchange
(NYSE) and Bombay Stock Exchange (BSE) monthly average closing price indices, for identifying temporal correlations, through a local multi-scale approach, assuming
that the behaviour of stock markets provides a faithful indicator for the state of the two economies. It is generally
expected that the condition of the economy is well reflected
in the countrys market behaviour.
The period of study is chosen from 1986 to 2010, to include various booms and bursts in the American economy
as well as pre and post liberalisation periods of the Indian economy. Both of these time series are highly nonstationary, revealing Gaussian random behaviour at certain scales, while possessing well defined periodic modulations at certain other scales [1]. Keeping these multi-scale
variability in mind, we take recourse to discrete wavelets
to isolate small-scale high frequency fluctuations from local smooth trends to unveil their mutual temporal correlations at multiple scales [1, 3]. Continuous Morlet wavelet
is effectively used to identify time varying periodic modulations at longer time scales [4]. Recently Random Matrix
approach has shown linkage between BSE and US market
on a small time frame, market crisis in US led to strong
correlation among the companies of the Indian market [5].
Post-liberalisation study of integration of Indian market
with the international markets, on the basis of correlations between the BSE and international stock exchanges,
March 8, 2016

possess bi-modal character, whereas BSE distribution is

uni-modal with a rightly skewed fat tail.

frequency

45
40
35
frequency

55

40

50

45

20

40

10
0
4
2
0
2
4
mean subtracted logarithmic return

30
25

35

30
20
10
0
4 2
0
2
4
mean subtracted logarithmic return

30
25

20

20

15

15

10

10
5

5
0

40

50

30

frequency

55

frequency

has clearly revealed that reactions of BSE are in tandem

with those seen globally [6]. Our multi-scale approach exhibits diametrically opposite behaviour at small and large
time scales. In the latter case, BSE is found to be strongly
driven by NYSE with a small time lag. The precise nature
of the long term phase correlation is found to be different
at 1, 2 and 3 year time scales. We make use of TodaYamamoto Granger causality tests to study the temporal causality at multiple scales [2, 7, 8]. The multi-scale
causality unearths the impacts of short and long term developments of one on the other. Considering the vastly
bigger size of the NYSE, on the basis of its market capitalisation, and its observed periodic influence on the BSE
price index, it is inferred that, smaller stock exchanges
may provide a natural setting for simulating future shortterm market behaviour of much bigger exchanges, as the
former being smaller reacts and equilibrates much faster
than the latter one. At a small time scale, high frequency
market fluctuations clearly showed that BSE behaviour is
being reflected in NYSE after a time lag.

0.5

1.5
2
2.5
3
normalised stock value

3.5

0.5

(a) BSE

1.5
2
2.5
3
normalised stock value

3.5

(b) NYSE

Fig.2: Histograms of the normalised BSE and NYSE closing price

values, revealing significant differences. Inset shows the distribution
of returns of the two stock exchanges, depicting almost Gaussian
behaviour for BSE, different from NYSE

For a better understanding of the local variations in

the data, we consider the histograms of the normalised
logarithmic returns for comparison:

2. Theory and Methodology

= p R(n) hR(n)i .
R
hR(n)2 i hR(n)i2

2.1. Data
Here, we study the monthly average closing price indices of NYSE and BSE for a period of 300 months from
January 01, 1986 to December 31, 2010 [9, 10]. As mentioned earlier, this time period includes several booms and
bursts of both the economies and also the pre and post
liberalisation periods of Indian economy. The mean subtracted monthly price indices have been scaled by their
respective standard deviations for normalisation. These
two time series, as depicted in Fig.1, show global similarities.

(1)

Here, p
R(n) = log xn+1 log xn is the logarithmic return
and, hR(n)2 i hR(n)i2 is the normalising factor, also
known as volatility of returns.

amplitude

20

10

10

50

100

150
time(month)

200

250

300

50

100

150
time(month)

200

250

300

20

amplitude

5
SENSEX
NYSE

10

10

3
2

Fig.3: Plots of the cumulative sum of normalised mean subtracted

logarithmic price indices of BSE(top) and NYSE(bottom), showing
mean reverting behaviour

1
0
1984

1987

1990

1993

1995 1998 2001

time(month)

2004

2006

2009

2012

Fig.1: Normalised monthly average closing price indices of BSE and

NYSE, exhibiting strikingly similar behaviour

The histograms of the normalised variations, shown in

Fig.2, reveal characteristic differences, although the global
features in Fig.1 are similar. NYSE price index variations
2

The distribution of returns shown in the insets of Fig.2

reveal features different from the global variations (see supplement material) with BSE is being closer to normal distribution. Clustering for small return values as well as
presence of outliers are evident for NYSE, which is in sharp
contrast with near Gaussianity of BSE. Interestingly, the
cumulative sum of the normalised mean subtracted logarithmic returns of both the stock exchanges exhibit mean
reversion (see supplement material).

More precisely the Hurst exponents are found to show

anti-persistence nature [11, 12], having values 0.4261 and
0.4512 for BSE and NYSE respectively. It is to be noted
that closer the value of exponent to zero, stronger is the
anti-persistent behaviour and hence the mean reverting nature. It is evident in Fig.3, that the rate of mean reversion
or aversion of the mean subtracted normalised logarithmic
return is higher for NYSE in the first half than BSE, implying higher volatility for the returns of NYSE. However, in
the second half, reversion frequency is comparable for the
two stock exchanges. For BSE the mean reversion rate has
not changed significantly over the analysis period, while
for NYSE, it is higher in the first half than the second.
Stock prices are affected by a plethora of factors, the dynamics of which are not revealed in the time series plot of
the stock prices. In order to have a better understanding
of the market dynamics, phase-space analysis is carried out
for indentifying the character of the changes with abscissa
showing the stock prices and the Y-axis the corresponding returns [13]. The phase space plots corresponding to
BSE and NYSE are shown in Fig.4. The curved lines are
the trajectory of the system as it evolves through time,
over a span of 300 months. One observes intersection of
trajectories in the plots, it arises due to the fact that the
price function has been unfolded in two dimensions, rather
than three or more. The stability of a curve is dependent
on its distance from the horizontal axis and the same indicates the magnitude of the variations too. It can be
inferred from the plots that BSE is less stable than NYSE
for a longer period of time. The trajectories are moving in
clockwise direction in both the cases.

dex fluctuations and non-stationary periodic trends, respectively. The discrete wavelets are specifically chosen
to extract variations from possible local polynomial behaviour at different scales [14, 15, 16]. The periodic modulations present in the data are extracted through the Morlet wavelet, which uses a Gaussian window, with a sinusoidal sampling function [4].

3.1. Discrete Wavelet Transform

Discrete wavelets make use of two kernels, known as
the father (t) and mother wavelets (t), satisfying the
following admissibility conditions [3, 17, 18]:

Z
Z
dt < , dt = 0, dt = 0,
Z
Z
2
|| dt = ||2 dt = 1,

(2)
(3)

The father, mother and daughter wavelets form a complete

orthogonal set, the daughter wavelets being the scaled and
translated versions of the mother wavelet:
j,k (t) = 2j/2 (2j t k),

j Z+ .

k Z,

(4)

Here, j and k are the scaling and translation parameters [18] (see supplement material).
SENSEX
3
2.5

1500

normalised stock value

300

1.5
1
0.5

200

0
0.5

1000

100

1
1985

1990

1995

2000
time(scale)

2005

2010

(a) Normalised average monthly values of the BSE

500

100
200

fluctuation

average behaviour

0.2

0.2

0.5

300
400

500

500
1000
0

2000

4000

6000

8000 10000 12000 14000 16000 18000

X

(a) BSE

600
1000

2000

3000

4000

5000
X

6000

7000

8000

0.5

0.5

9000

(b) NYSE

Fig.4: Phase-space plots corresponding to normalised data of BSE

and NYSE showing periodic modulations and bifurcating behaviour

0.5

1
0

0
5
1985

1990

1995

2000

2005

2010

1
1985

1990

1995

2000

2005

2010

(b) Average behaviour for the (c) Fluctuations for the first
first four levels
four levels

For a more systematic multi-scale comparison of the

two price indices, we study their time-frequency localisation and ensuing correlations at various time scales.

Fig.5: Plots of (a) normalised BSE monthly average index and (b)
average behaviour, calculated after averaging over progressively bigger temporal domains (c) fluctuations for the first four levels of BSE
average monthly index, exhibiting strong fluctuations in the second
half

3. Wavelet Transform
We make use of both discrete and continuous wavelet
transforms to study the local behaviour of the price in3

NYSE
2.5

0.04

SENSEX
NYSE

normalised stock value

1.5

0.035

0.5
0

0.98

0.03

0.5

1.5
1985

1990

1995

2000
time(scale)

2005

variance amplitude

2010

(a) Normalised monthly average of the NYSE

fluctuation

average behaviour
5

0.2

0.2

0.5

0.02
0.015
0.01

0.5

0.5

0.5

2
1985

0.96

0.025

1990

1995

2000

2005

2010

1
1985

0.005
0

1990

1995

2000

2005

0.5

1.5

2
2.5
level of filtering

3.5

Fig.7: The variance of the fluctuations at different levels of filtering,

with the inset showing the the variances corresponding to the average
behaviour revealing considerable differences between levels 5 and 7

2010

(b) Average behaviour for the (c) Fluctuations for the first
first four levels
four levels
BSE

Fig.6: Plots of (a) normalised NYSE monthly average index and (b)
average behaviour at multiple scales (c) fluctuations for the first four
levels of NYSE, showing similar characteristics as those of BSE

1
0.5
0

The average behaviour over progressively longer data

length are captured in the local trends for the first four levels, as depicted in, Figs.5(b) and 6(b). The fact that, the
two time series reveal similar global behaviour at different
scales, is evident in the average behaviour. For testing of
normality of the average behaviour and of the fluctuations,
SHAPIRO-WILK (SW) and
KOLMOGOROV-SMIRNOV (KS) tests are conducted [19] [20].
Null Hypothesis is rejected for both SW and KS tests, revealing that fluctuations do not belong to the normal distribution [19, 21]. The results of the tests for the first 4
levels are shown in the Tables.4 and 5 (see supplement
material).
For a more systematic understanding of the nature of variation at different scales, we first carry out a quantitative
estimation of the deviations of the local variations (fluctuations) from the normal distribution, thus taking recourse
to Quantile distributions. Quantile-Quantile plot uses medians, as compared to use of mean in histograms, making
it more resistant to outliers [22]. Distribution of fluctuations of BSE and NYSE for the first four levels show a significant digression from normal behaviour, characterised
by the dotted line [23] (see supplement material). The
deviations are significant, both at lower and upper ends
of the plots, which are indicative of long-tail behaviour
on both sides. Deviation from normality is more for BSE
than NYSE, at all the levels. The variations show volatile
behaviour at small scales and structured variations at progressively higher scales, as seen in Figs.5(c) and 6(c). The
second half of the fluctuations are highly volatile for both
the stock exchanges.

0.5
1
1

NYSE
2
1
0
1
1

2
level of filtering

Fig.8: Box-plots to show the skewness and outliers of the average behaviour at different levels, for both the stock exchanges. The lower
and upper boundary of the notched boxes represent the 25th (q1 )
and 75th (q3 ) percentiles, respectively. There is a red line in between the boundaries of each box, it is 50th percentile i.e, median
(q2 ). If the notches in the box-plot do not overlap, then it can be
concluded with 95% confidence that the true medians differ. Here,
length of the whiskers have been specified as 1.5 times the length
of the inter-quartile range and points beyond that are considered as
outliers represented by + in the box-plots. The value 1.5 corresponds
to approximately +/2.7 and 99.3 coverage if the data are normally
distributed

BSE
0.1

3.1.1. Pearson and Spearman Correlation

For estimations of multi-scale linear and monotonic

correlations, between the two stock exchanges, we compute both the Pearson and Spearman correlations of their
average behaviour and fluctuations. It is evident from the
values tabulated in Tables.6 and 7 (see supplement material), that the average values of the two stock exchanges exhibit monotonic relationship as Spearman correlation coefficients are higher than Pearson for all levels of filtering.
This indicates that the two stock exchanges are moving
concurrently throughout, though the increment or decrement in the stocks value for the two may not be equal.
While in the case of fluctuations, Pearson coefficients values are higher than Spearman, at various scales, indicating
a linear relationship between the fluctuations.

0.1
1

NYSE
0.1
0
0.1
1

2
level of filtering

Fig.9: Notched box-plots representation of the fluctuations of (a)

BSE and (b) NYSE, stock exchanges. Labelling and qualitative results hold from the average behaviour box-plots of Fig.8

3.1.2. Multi-scale Probability Density

Probability density estimation of the average behaviour
and fluctuations are carried out, with the help of Kernel
smoothing (KS) operations. As is known, Kernel is a special type of Probability Density Function (PDF) with the
property of being non-negative, real valued, even and having definite integral equal to 1, over its support value:

0.01
SENSEX
NYSE

0.009
0.008

0.5

variance amplitude

0.007
0

0.006

0.005

K(u)du = 1

0.004

(5)

0.003
0.002
0.001
0

K(u) = K(u)
0

0.5

1.5

2
2.5
level of filtering

3.5

(6)

The first requirement ensures that the method of kernel

density estimation results in a probability density function, while the second ensures that the average of the corresponding distribution is equal to that of the sample used.
Kernel smoothing is a non-parametric method of estimating PDF, as it does not assume any underlying distribution
for a variable. At every data, a Kernel is created with the
data point at the centre to ensure that the Kernel is symmetric about it. The PDF is then estimated by adding all
these Kernel functions and dividing by the number of data.
Probability density distribution of fluctuations show a progressive diffusive spread (see supplement material). The
plots depict the transition of fluctuations from leptokurtic
to mesokurtic (normal ) and finally to platykurtic with the
advent of structured variations (see supplement material).

Fig.10: Plots of variance of fluctuations at different levels with inset

showing the distribution of variances for average behaviour, after
removal of outliers. Significant differences are there in the variances
of BSE and NYSE, in case of average behaviour. The sudden jump
in the variance

For the purpose of comparison of the two indices, as

well as to quantify the changes in their behaviour as a
function of scale, the variances of the average behaviour
and fluctuations for the first 4 levels, are calculated. The
variance of fluctuations and average behaviour of the two
exchanges show negligible differences, at various scales, as
shown in Fig.7. However, once the outliers are removed,
difference in the variances for the average behaviour become glaring, as shown in Figs.10, the variances of BSE decreased drastically while that of NYSE remain unchanged.
In case of fluctuations the two exchanges continue to exhibit what was observed hitherto. These findings entail
that variance at different scales in the monthly stock value
of BSE is mainly due to outliers.

Average behaviour of both the stock exchanges exhibit

bi-modal nature, it is symmetric for NYSE and it is asymmetric with a pronounced rightly skewed tail for BSE.
Skewness persists, with increments in the level of filtering for BSE, indicating that the average behaviour is a
5

BSE
-0.23088
-0.013232
0.49524
-0.17078

0.8
0.6
0.4

probability density

0.2
0.0

1.0
0.8
0.6
0.0

0.0

0.2

0.4

probability density

0.8
0.6
0.4
0.2

probability density

(b) level 2

1.0

(a) level 1

(c) level 3

Table 1:

level
1
2
3
4

0.6
0.0

(d) level 4

Fig.11: Juxtapositions of the probability densities of average behaviour of BSE (blue) and NYSE (red) for (a) level 1 (b) level 2 (e)
level 3 and (f) level 4, showing significantly different behaviour

Fluctuation
Skewness
NYSE
0.14554
-0.32914
-0.3104
-1.0385

0.4
0.2

probability density

0.8

The relationship that the two stock exchanges exhibited in variance and probability distribution plots, bear
resemblance. It can also be inferred that, higher variance
in the BSE is due to rightly skewed fat tail of the average
behaviour. In order to cross check and underpin our claim
about more number of outliers in BSE than NYSE, the
average behaviour and fluctuations were analysed, regarding outliers and skewness, using the Box-plot method as
shown in Figs.8 and 9 (see supplementary material). It is
found that average behaviour of BSE possesses outliers for
all the four levels, while in case of NYSE there are none.
Fluctuations corresponding to both the stock exchanges
have same number of outliers.
The change in the variances, once the outliers are removed,
are drastic for BSE, as shown in Fig.10. It is physically
tenable that a developing economy like India, has more
number of outliers than developed one and vice-versa in
the case of variance or volatility [24, 25]. The correlation
between the two stock exchanges remain unchanged for
average behaviour, while in case of fluctuations it showed
a sign of decline, after the removal of the outliers. This
was inferred on the basis of the Pearson and Spearman
correlations coefficients, tabulated in the Table. (see supplementary material).

1.0

1.0

mixture of more smaller stock values and outliers, where

as NYSE has homogeneity in its values.

Kurtosis
BSE
7.5069
10.041
7.2145
10.977

NYSE
4.4489
4.7613
5.746
8.9887

Table 2:

Average behaviour

level
1
2
3
4

Skewness
BSE
1.4119
1.4076
1.3822
1.3351

NYSE
0.18182
0.17737
0.1692
0.13364

Kurtosis
BSE
3.8667
3.8454
3.745
3.5765

NYSE
1.8049
1.8001
1.8003
1.7515

Skewness and Kurtosis coefficients have been tabulated

in Tables , from which it can be inferred that , in the case
of fluctuations the probability density is highly leptokurtic for BSE in comparison to that of NYSE. Stronger leptokurtic behaviour is bestowed with fatter tails and thus
extreme fluctuations, on both sides, are more probable in
BSE, this can be vindicated by the boxplots, as shown in
Fig. . However, the distribution is symmetrical for both
the stock exchanges, as the skewness values are close to
zero. For average behaviour, the distribution is rightly
skewed with a weak leptokurtic nature in case of BSE,
while for NYSE they are symmetrical and platykurtic, at
6

1.5

1.5

where s is the scale. The admissibility conditions are given

by,
r~
~k)|2 d k
|(
|~k|r
Rr

1.0
0.5
0.0

Density
0.5
0.0

Density

1.0

<

1
~k) =
where, (
(2)r
Z
and
(~x)dr ~x = 0.

(8)
Z
Rr

(10)

Rr
2

(a) level 1

Here r is the number of spatial dimensions. In our analysis,

we use the complex Morlet wavelet:

(b) level 2

/2

(11)

Density

1.0

where, n is the localized time index. (n) is a marginally

admissible function, it is made admissible by taking 0 =
6. The Fourier wavelength of (n), F , is given by,

0.5

F =

(c) level 3

4s
p
1.03s,
0 + 2 + 02

(12)

implying that the scale and the wavelength for the Morlet
wavelet are approximately equal. The psuedo-frequency
s , pseudo-period W (s) and scale s are related by,

0.0

0.0

0.5

Density

1.0

1.5

(n) = 1/4 e0 n en
1.5

(~x)ek~x dr ~x (9)

(d) level 4

s =
Fig.12: Probability densities of average behaviour of BSE (blue) and
NYSE (red), after removal of outliers, for (a) level 1 (b) level 2 (e)
level 3 and (f) level 4

c
s

W (s) =

1
,
s

(13)

where, c is the central frequency of the wavelet and

is the sampling period. The cone of influence (COI) [26],
is known as, the region beyond which convolution errors
make the wavelet coefficients unreliable for analysis [17].
The scalograms in Fig.13, clearly demonstrates existence
of well defined periodic modulations at different scales.

various scales. The risk of having extreme events due to

outliers are low for NYSE.
For quantifying the periodic modulations in these nonstationary time series, we proceed to continuous wavelet
transform.

Semi-log plots in Figs.14 and 15 reveals the periodic

modulations present in the parameters corresponding to
certain months.

3.2. Continuous Wavelet Transform

2.5

X: 177
Y: 2.163

X: 106
Y: 1.97

For studying the structured variations and their phase

and temporal correlations Morlet wavelet is used [18, 26,
27]. An integrable, well localized (in both the physical
and Fourier domains), zero mean function, the mother
wavelet (n), is used as the analyzing function. Given
a discrete data set X = {xn : n Z+ }, the wavelet coefficients are calculated by convolving the data, with the
scaled and translated (n). The wavelet coefficients are
obtained through equation,
Wn (s) =

N
1
X
n0 =0

xn0

n n0
s

Log10(sum(|w(t,s)|))

0.5
X: 11
Y: 0.5228

0.5

50

100

150
scale (month)

200

250

300

Fig.14: Semi-log plot of the wavelet power for BSE data, summed
over all time at different scales unveiling the periodicity of 11 and 26
months are present


,

X: 26
Y: 1.213

1.5

(7)

Table 4: Dominant periods of NYSE and its first difference

NYSE

period(months)
first difference

4
11

15
9

24
32

56
42

98
64

162

SENSEX
scale 11

2
0

(a) 3D scalogram plot of BSE

SENSEX
scale 26
scale 34

5
1985

1990

1995

2000

2005

2010

Fig.16: Plot of mean subtracted normalised stock value of BSE and

its CWT coefficients at scale of 11, 26 and 34 months, corresponding
to the dominant periods unveiled in the semi-log plot of 14

(b) 3D-scalogram plot of NYSE

NYSE
scale 15
scale 23

2
0
2

Fig.13: 3D scalograms of the CWT coefficients depict transient nature as well as periodic modulations of the two indices, at certain
scales

NYSE
scale 34
scale 56

2
0
2
X: 56
Y: 1.486

X: 98
Y: 1.568

1.5

4
1985

X: 162
Y: 1.632

1990

1995

2000

2005

2010

1
0.5
X: 15
Y: 0.3147
0

Fig.17: Plot of mean subtracted normalised stock value of NYSE and

its CWT coefficients at scale of 11, 23, 34 and 56 months, showing
different phase correlations at different time scales

X: 24
Y: 0.3771

0.5
1
X: 4
Y: 1.058

1.5

The CWT coefficients juxtaposed with the mean subtracted normalised index, as shown in Figs.16 and 17, reveal the degree of similarity between the coefficients corresponding to the periods shown in semi-log plots. The
time series of BSE displays lesser number of dominant periods than NYSE, though the lower and upper bound are
comparable for the two stock exchanges. Lesser number
of dominant periods entails that, the number of dominant
factors driving structured variations in BSE are fewer in
number than the NYSE, or pertaining number of smaller
factors determining the behaviour of BSE is more than
NYSE.

2
0

50

100

150

200

250

300

Fig.15: Semi-log plot of the wavelet power for NYSE data, summed
over all time at different scales. Periodicity of 4, 14, 23 and 56 months
are clearly seen

Peaks in the semi-log plots pin point the dominant

scales; these are 11 and 26 months for BSE and 4, 14,
23 and 56 month for NYSE, shown in Tables.3.2 and 4.
Table 3: Dominant periods of BSE price index and its first difference

BSE

period(months)
first difference

11
15

26
80

106
71

3.2.1. Comparing the coefficients of BSE and NYSE

177

A careful study of CWT coefficients, as shown in Fig.18(a),

(b) and (c), reveal that BSE and NYSE are moving in tandem except when one of them is non-periodic, revealing a
transition period of an unstable regime.
The two stock exchanges maneuvered in un-synchronised
fashion throughout except between 2008 to 2010, 1987 to
8

2006 and 1987 to 1997 at 1, 2 and 3 year scales, respectively. They show synchronised behaviour only during the
crisis period, i.e between 2008-2010, at 1 year time scale
and 2006-2010 at a scale of 2 years. The crisis centric interactions, between the two exchanges, are more visible at
approx 1 year scale, while 2 and 3 years time scales portray
the interactions of the two, once the crisis impacts have
diffused and got integrated with the movements at higher
scales. At 3 years time scale, the synchronised behaviour
appear post 1997 with a lag of 5-6 months and slowly the
lag became zero with time, as seen in Fig.18(c), exhibiting
the degree of integration of BSE. Multi-level synchronisation between coefficients of two stock exchanges is possible
only if the driving forces of both the stock exchanges are
of similar nature or one of them is driving the other, it
is clear that NYSE being the driving force and BSE got
aligned with a time lag.

sensex
nyse

phase angle(rad)

2
1985

1990

1995

2000
time(month)

2005

2010

2005

2010

(a) scale 11

phase angle

SENSEX
NYSE

1
0
1
2
1985

1990

1995

2000
time(month)

(b) scale 24
sensex
nyse

phase angle(rad)

2
1985

1990

1995

2000
time(month)

2005

2010

(c) scale 34

CWT coefficients plot

3
SENSEX
NYSE

amplitude

Fig.19: Phase plots of CWT coefficients of average behaviour data

at level 1 BSE (blue) and NYSE (red) at a time scale of (a) 11 (b)
24 and (c) 34 months, respectively

2
1985

1990

1995

2000
time(month)

2005

2010

(a) scale 11

They are in anti-phase mode between 1987-2002 at 1 year

scale, from 1997-2002 and again from 2007-2011 at 2 year
scale and during pre-liberalisation period at 3 years scale.
The anti-phase nature at 1 year scale vindicate the finding
that its a crisis centric scale. At 3 year scale, it is seen that
in-phase nature start to emerge after 1996, though with a
lag of 3-4 months which subsequently got relieved and the
two become phase locked, even during the crisis periods.
In-phase nature in the second half can be attributed to the
opening up of the Indian market, thus paving the way for
synchronised behaviour.

amplitude

SENSEX
NYSE

5
1985

1990

1995

2000
time(month)

2005

2010

2005

2010

(b) scale 24
CWT coefficients plot
4

SENSEX
NYSE

amplitude

2
1
0
1
2
3
4
1985

1990

1995

2000
time(month)

(c) scale 34

3.2.3. Cross Wavelet Spectrum

Fig.18: Plots of continuous wavelet transform (CWT) coefficients of
BSE (blue) and NYSE (red) depicting the movement of two stock
exchanges at scale (a) 11 (b) 24 (c) and 34 months respectively, showing out of phase at smaller scale and in-phase behvaiour at higher
scales

In Figs.20 and 21, common power region lies in the second half of the plots. In case of fluctuations, it is maximum
at around 2, 4, and 8 month scales corresponding to levels
1, 2 and 3 respectively. In case of average behaviour, as
seen in Fig.21, it is observed at a time scale of 32 months.
However, the nature of the maximum power share for fluctuations and average behaviour exhibit differences. Maximum power sharing in the former is dependent on level
of filtering, while it is independent for the later. Interestingly, the scale at which cross-correlation is prevalent is
unaffected by level of filtering.

3.2.2. Phase Plot

Underlying dynamics of a dynamical systems can be
better understood through the phase structure of periodic
modulations and the manner in which they are in synchronization with each other. Here, phase structures of average
behaviour (trend), corresponding to the dominant periods
of BSE and NYSE have been analysed. Figs.19 (a), (b)
and (c), depict the evolution of phase angle of the average
behaviour of BSE and NYSE with respect to scales. Inphase nature between the two stock exchanges is observed
between 2002-2011 at 1 year scale, from 2002-2007 at 2
years scale and between 1997-2011 at a scale of 3 years.
9

(a) level 1

(c) level 3

(b) level 2

(a) level 1

(d) level 4

(c) level 3

Fig.20: Cross wavelet transform plots of fluctuations at 95% confidence interval with cone of influence, shown in lighter shade, indicating the region affected by edge affects. The colour code for
power ranges from blue (low power) to red (high power). The nature of phase relationship between the two stock exchanges is represented by the direction of the arrows. They are directed toward right,
right-upward, right-downward, left, left-upward, left-downward, respectively showing that both the variables are in phase, NYSE is
lagging, NYSE is leading, both are out of phase, NYSE is leading
and NYSE is lagging. The share of power between the variables are
plotted for (a) level 1 (b) level 2 (c) level 3 and (d) level 4. It is more
in the second half of the plot, at around time scale of 2 months for
the level 1 and 6 months for level 2 and at around 12-16 months for
level 3

(a) level 1

(c) level 3

(b) level 2

(d) level 4

Fig.22: Wavelet coherence plots of fluctuations obtained at (a) level 1

(b) level 2 (c) level 3 (d) level 4, of BSE and NYSE at 95% confidence
interval. The cone of influence, shown in lighter shade. The colour
code representing power ranges from blue (low power) to red (high
power), hotter the color higher the power, the X-axis and Y-axis
represent the time (month) and specific scale, respectively.

3.2.4. Wavelet Coherence

Wavelet coherence is calculated with two continuous
wavelet transformation and embodies information about
local correlation between two stock indices, in time frequency space [28, 18]. It helps in unveiling locally phase
locked behaviour. Here, the significance level has been determined by invoking Monte Carlo simulation.
Sporadic blobs are present at 4 and 8 months scale, as
seen in Figs.22(a) and (b) respectively. With the advent
of more structured variations, a bridge like structure can
be seen at a scale of 32 months. Arrows are found to point
in the upward direction, as seen in Figs.22(c) and (d), indicating that BSE and NYSE are coherent with a lag of 90
degree around this scale.
In case of average behaviour, intermittent red blobs are
present at scales of 4 and 8 months, as shown in Fig.23
(a), (b) and (c). Most of them are in the second half of
the plots. Presence of a bridge like structure at around
32 to 50 months scale, depicts existence of a very strong
correlation between the two stock exchanges, which is unaffected by the level of filtering. This demonstrates that
the correlation structure is resistant to small scale fluctuations and carry information about interaction taking place
at large scales.

(b) level 2

(d) level 4

Fig.21: Cross wavelet transform (XWT) for average behaviour for

levels 1, 2, 3 and 4 filtered BSE and NYSE at 95% confidence
level. Labelling and qualitative results hold from the XWT plots of
fluctuation, Fig20

10

(a) level 1

(c) level 3

stationarity. To overcome this shortcoming, KwiatkowskiPhillipsS

(KPSS) test is conducted, it has (trend) stationairty as
null hypothesis.
Toda Yamammoto developed a methodology which estimates an augmented VAR even when their is cointegration
of different orders. This method works even if VAR is stationary (about a trend), integrated or co-integrated of an
arbitrary order [8]. The steps that we followed in order
to know about the causality between the two variables are
discussed below [31].

(b) level 2

Here, Yt and Xt are the variables to be tested and ut

and vt are the mutually uncorrelated white noise errors,
t stands for time period and p is the time lag. Testing
H0 : b1 = b2 = .... =bp =0, against HA : not H0 , is
a test that X does not Granger cause Y. Similarly, H0 :
d1 = d2 = .... =dp =0, against HA : not H0 , is a test that
Y does not Granger cause X. Steps involved in the test
are;

(d) level 4

Fig.23: Wavelet coherence plots of average behaviour for (a) level 1,

(b) level 2, (c) level 3 and (d) level 4, at 95% confidence interval.
Labelling and qualitative results hold from the fluctuations plots i.e
22

4. Multi-Scale Causality
To test the causal relationship between two variables,
the standard Granger causality test (GCT) is implemented.
According to GCT, if X causes Y then the forecast of Y is
better if the information in X is used than when it is not.
The test relies on VAR model represented by equations 14
and 15:
Yt = a0 + a1 Yt1 + a2 Yt2 + .... + ap Ytp +
b1 Xt1 + b2 Xt2 + .... + bp Xtp + ut

Xt = c0 + c1 Xt1 + c2 Xt2 + .... + cp Xtp +

d1 Yt1 + d2 Yt2 + .... + dp Ytp + vt

(14)

(15)

In each case, rejection of the null hypothesis means there

is Granger causality. Here, we have used Toda-Yamamoto
Granger causality (TYGC), instead of the traditional Granger
causality, as the typical Granger-causality can be problematic [29]. Using F-statistics, when one or both time series
are non-stationary or when they are not co-integrated, can
lead to spurious causality [8, 30]. This is taken care of in
Toda-Yamamoto version of Granger causality. So we do
not need to carry out tests to check for co-integration of
the series concerned.
There are several tests like Augmented Dickey Fuller
(ADF) and Phillips and Perron to check whether the time
series is integrated or not, but the power of these tests are
low in comparison to the alternative hypothesis of (trend)
11

1. The first step is to check for stationairty, so ADF

and KPSS tests are conducted.
ADF test, null hypothesis H0 : unit root exists, to
check whether signal is stationary or not [32] (supplement material).
KPSS test, null hypothesis of this test implies stationarity in the univariate time series, with alternative that it is non-stationary unit root process [33]
(supplement material). The test uses the model comprising of equations S4 and S5. If the time series
is non-stationary; then differencing it and applying
steps 1 and 2 on it, until it becomes stationary, unveils the order of integration. Let the value of integration is i, which is maximum of the integration
values of the two time series concerned.
2. Set up bivariate VAR model with any arbitrary lag
value, using the actual data [34, 35]. The execution
of the VAR model gives information about the appropriate lags to be considered for the WALD test.
Here, finding the correct lag value plays a very significant role, as Granger causality is very much susceptible to lag length being considered, inclusion of
extra lags or considering lesser lags than required,
leads to spurious causality results. To deal with it,
Akaike Information Criteria (AIC), Hannan-Quinn
(HQ), Schwartz Criterion (SC) and Final Prediction
Error (FPE) values are used to find the appropriate
lag [35], let the lag value found out to be p.
3. Portmanteau Test (asymptotic) is conducted on the
coefficients corresponding to the lag p, to find out
if they possesses serial correlation (The null hypothesis shows no serial correlation in the coefficients).
Acceptance of the null hypothesis is done after, they
are subjected to stability test. Coefficients need to
pass both the tests in order to be considered for the
WALD test. In the case of monthly averaged normalised data, lag 9 is more stable than lag 1.

Table 5: Results of the Toda-Yamamoto Granger causality test for

the average behaviour and fluctuations of the two stock exchanges

fluctuations
level
1
no causality
2
no causality
3
BSE causing
NYSE
4
BSE causing
NYSE

lag average
behaviour

no causality
no causality
no causality

no causality

lag

causes BSE at lag=9 and there is no causality at lag=1.

Causality results for the average behaviour and fluctuations are tabulated in the Table 5.

Fig.24: Result of VAR model for the normalised monthly averaged

value of the two stock exchanges, showing the optimum lag value to
be 1 and 9 on the basis of the AIC, HQ, SC and FPE values

4. Again the VAR model is set up, this time with lag
value equal to p+i, here p is lag value corresponding to which coefficients passed the tests mentioned
in step 3.
5. Conduct Wald test on the coefficients obtained in
step 4 [36, 35] (null hypothesis H0 : X does not
Granger cause Y). The WALD test results for monthly
averaged normalised data is;
when the null hypothesis is BSE does not Granger
cause NYSE . In this case
Chi-squared test: X2 = 8.6, df = 6, P(> X2) = 0.2
Null hypothesis cannot be rejected at 5 percent confidence interval, NYSE granger cause BSE and
when the null hypothesis is NYSE does not Granger
cause BSE
Chi-squared test: X2 = 16.2, df = 6, P(> X2) =
0.013
Null hypothesis stands rejected, BSE does not Granger
cause NYSE.
Toda-Yamamoto test was carried out for monthly averaged normalized data of BSE and NYSE as well as for
the average behaviour and fluctuations, of the two stock
exchanges. Granger causality of average behaviour data
could only be found till level 6, as the integration value for
levels higher than 6 is either 20 or more. Hence, it is above
the upper threshold value of 16, which is the maximum integration value of average behaviour or fluctuations upto
which the analysis can be done, the size of the data set is
a constraint here. In case of monthly averaged normalized
data of the two exchanges, the outcome is NYSE Granger
12

Multi-scale causal relations have been established between the two stock exchanges with a motive to understand their short and long term relationship [1, 37, 38], as
well as to infer about the developments driving the causality relations between the two economies. It will help in
framing policies and taking measures during the time of
economic slowdown.
According to the results tabulated in Table.5, in case of
average behaviour, NYSE is Granger causing BSE at 32
months time scale with a lag of 9 months, while for fluctuations, NYSE is Granger causing BSE for structured variations. It is other way round for high frequency variations.
The absence of NYSE Granger causing BSE at lower scales
can be attributed to the insularity of the Indian economy.
The finding that BSE Granger causing NYSE at smaller
scales may appear counter intuitive and spurious, because
of following factors;
1. The NYSE is more stable than BSE.
2. NYSE represents a developed economy, while BSE is
a developing one.
3. The market capitalisation of NYSE is significantly
higher than that of BSE, NYSE is being the worlds
largest stock exchange according to market capitalisation and trade value.
A plausible explanation for the spurious causality at lower
scales, as NYSE Granger causing BSE at higher scales, because of trickle down effect the fluctuations in BSE at lower
scales might get affected by the causality at higher scales,
thus bearing similarities to the fluctuations of NYSE at
lower scales. Hence, paving way for BSE to spuriously
Granger cause NYSE but at the same time NYSE failing
to Granger cause BSE at lower scales, due to insularity of
the Indian economy. ( ) And as stock market is a complex
dynamical system, and because fluctuations in BSE equilibrate in short span of time, it can be used to simulate
the NYSE variations.

5. Conclusion

References
[1] J. B. Ramsey, L. Camille, The decomposition of economic relationships by time scale using wavelets: Expenditure and income,
Studies in Nonlinear Dynamics Econometrics 3 (1) (1998) 122.
[2] J. B. Ramsey, The contribution of wavelets to the analysis of
economic and financial data 357 (1760) (1999) 25932606.
[3] I. Daubechies, Ten lectures on wavelets, 64 CBMS-NSF Regional Conference Series in Applied Mathematics, Society for
Industrial Mathematics, 1992.
[4] J. Lin, Feature extraction based on morlet wavelet and its application for mechanical fault diagnosis 234 (2000) 135148.
[5] V. Kulkarni, N. Deo, Correlation and volatility in an indian
stock market: A random matrix approach, The European Physical Journal B 60 (1) (2007) 101109.
[6] D. Mukherjee, Comparative analysis of indian stock market
with international markets, great Lakes Institute of Management 1.
[7] Some recent development in a concept of causality, Journal of
Econometrics 39 (12) (1988) 199 211.
[8] H. Y. Toda, T. Yamamoto, Statistical inference in vector autoregressions with possibly integrated processes, Journal of Econometrics 66 (1-2) (1995) 225250.
[9] Downloaded from Yahoo Finance.
[10] Downloaded from stooq.
[11] Time-varying hurst exponent for {US} stock markets, Physica
A: Statistical Mechanics and its Applications 387 (24) (2008)
6159 6169.
[12] E. E. Peters, Fractal structure in the capital markets, Financial
Analysts Journal 45 (4) (1989) pp. 3237.
[13] N. H. Packard, J. P. Crutchfield, J. D. Farmer, R. S. Shaw,
Geometry from a time series, Phys. Rev. Lett. 45 (1980) 712
716.
[14] J. Modi, S. Nanavati, A. Phadke, P. Panigrahi, Wavelet transforms: Applications to data analysis-i, Resonance 9 (11) (2004)
l022.
[15] J. K. Modi, S. P. Nanavati, A. S. Phadke, P. K. Panigrahi,
Wavelet transforms: Application to data analysis-ii, Resonance
9 (12) (2004) 813.
[16] S. Ghosh, P. Manimaran, P. K. Panigrahi, Characterizing multiscale self-similar behavior and non-statistical properties of fluctuations in financial time series, Physica A: Statistical Mechanics and its Applications 390 (23-24) (2011) 43044316.
[17] C. Torrence, G. Compo, A practical guide to wavelet analysis,
Bulletin of the American Meteorological Society 79 (1) (1998)
6178.
[18] S. Mallat, et al., A theory for multiresolution signal decomposition: The wavelet representation, IEEE transactions on pattern
analysis and machine intelligence 11 (7) (1989) 674693.
[19] S. S. SHAPIRO, M. B. WILK, An analysis of variance test for
normality (complete samples) 52 (3-4) (1965) 591611.
[20] F. J. J. Massey, The Kolmogorov-Smirnov Test for Goodness
of Fit, Journal of the American Statistical Association 46 (253)
(1951) 6878.
[21] I. Melo, B. Tom
asik, G. Torrieri, S. Vogel, M. Bleicher,
S. Kor
ony, M. c. v. Gintner, Kolmogorov-Smirnov test and its
use for the identification of fireball fragmentation, Phys. Rev.
C 80 (2) (2009) 024904.
[22] R. Koenker, K. F. Hallock, Quantile regression, Journal of Economic Perspectives 15 (4) (2001) 143156.
[23] A. Ghasemi, S. Zahediasl, Normality tests for statistical analysis: A guide for non-statisticians, Int J Endocrinol Metab 10 (2)
(2012) 486489.
[24] S. M. BARTRAM, G. BROWN, R. M. STULZ, Why are u.s.
stocks more volatile?, The Journal of Finance 67 (4).
[25] S. I. COHEN, Stock performance of emerging markets, The Developing Economies 39 (2).
[26] S. G. Mallat, A wavelet tour of signal processing, Academic
Press, 1999.
[27] A. K. Behera, A. S. Iyengar, P. K. Panigrahi, Non-stationary
dynamics in the bouncing ball: A wavelet perspective, Chaos:

In conclusion, we studied the movement of NYSE and

BSE price indices over a span of 300 months, for a multiscale understanding of how these two economies behaved
during the turbulent times and the kind of impact they
had on each other. Logarithmic returns of the two stock
exchanges showed stationarity and mean reverting nature,
the frequency of mean reversion or aversion for NYSE is
not fixed, while for BSE it remained unchanged over the
analysis time frame. The Hurst exponents was found to
be less than 0.5 for both the exchanges, entailing antipersistent nature.
Wavelets belonging to the Daubechies family (Db4) optimally removed the local linear trends from the two time series. Monotonic correlation for average behaviour and linear in case of fluctuations, corresponding to the two stock
exchanges, at multiple scales were established. Symmetric and assymetric bi-modal nature, with a right and left
skewed tail for BSE and NYSE, respectively, were found.
Skewness and variance in BSE has been attributed to the
higher frequency of outliers, a trademark of developing
economy.
Continuous Morlet wavelet transform localised the periodic components, linear and non-linear in nature. 1, 2 and
3 year non-stationary periodic modulations, with NYSE
effecting commensurate changes in BSE price indices, was
found. At 1 and 2 years time scales, the cross-correlations
were driven by turbulence, where as 3 year time scale was
ubiquitous and helped in discerning the nature of postliberalization interactions. It witnessed the bifurcation in
the variance and differences in the bi-modal nature. Crosscorrelations near these scales were strong and unaffected
by fluctuations. NYSE Granger caused BSE at this scale
only, thus it played a pivotal role in unravelling the picture of the underlying post-liberalisation interactions, on
a global scale. Multi-scale Granger causality analysis revealed NYSE Granger causes BSE at higher scales, both in
the case of fluctuations and average behaviour. Fluctuations of BSE at lower scales are acting as mirror images for
the fluctuations of NYSE, thus can be utilised to simulate
the lower fluctuations of NYSE.

13

[33]

[34]

[35]
[36]
[37]

[38]

40
35
30

50
25
40
20
30
15
20

10

10

0
6

0
8

(a) BSE

(b) NYSE

Fig.S1: Histograms of the normalised returns for BSE and NYSE ,

show deviations from Gaussian behaviour for the former at the same
time has been bestwoed with outliers too

6. Quantile plots

QQ Plot of Sample Data versus Standard Normal

QQ Plot of Sample Data versus Standard Normal

0.2

0.3

0.15

0.2

0.1
Quantiles of Input Sample

[31]
[32]

45

70

0.05
0
0.05
0.1

0.1

0.1

0.2

0.3

0.15
0.2
3

1
0
1
Standard Normal Quantiles

0.4
3

(a) level 1
QQ Plot of Sample Data versus Standard Normal

QQ Plot of Sample Data versus Standard Normal

0.5

0.4

0.8

0.3

0.6

0.2
0.1
0
0.1
0.2

0.4
0.2
0
0.2
0.4

0.3
0.4
3

1
0
1
Standard Normal Quantiles

(b) level 2

Quantiles of Input Sample

[30]

80

60

Quantiles of Input Sample

[29]

Quantiles of Input Sample

[28]

Correlation and Causality : Supplemental

Materials

An Interdisciplinary Journal of Nonlinear Science 24 (4) (2014)

043107.
A. Grinsted, J. C. Moore, S. Jevrejeva, Application of the cross
wavelet transform and wavelet coherence to geophysical time
series, Nonlinear Processes in Geophysics 11 (5/6) (2004) 561
566.
K. M. Zonglu he, On spurious granger causality, ECONOMICS
LETTERS 73 (3) (2001) 307313.
S. Johansen, Statistical analysis of cointegration vectors, Journal of Economic Dynamics and Control 12 (2-3) (1988) 231254.
Toda-Yamamoto implementation in R.
D. Dickey, W. A. Fuller, Likelihood ratio statistics for autoregressive time series with a unit root, Econometrica 49 (4) (1981)
105772.
D. Kwiatkowski, P. Phillips, P. Schmidt, Y. Shin, Testing the
null hypothesis of stationarity against the alternative of a unit
root: How sure are we that economic time series have a unit
root?, Journal of Econometrics 54 (1-3) (1992) 159178.
M. Watson, Vector autoregressions and cointegration, Working Paper Series, Macroeconomic Issues 93-14, Federal Reserve
Bank of Chicago (1993).
B. Pfaff, Var, svar and svec models: Implementation within r
package vars.
J. J. Dolado, H. L
utkepohl, Making wald tests work for cointegrated var systems, Econometric Reviews 15 (4) (1996) 369386.
A. O. B. Atilla Cifter A, Multi-scale causality between energy consumption and gnp in emerging markets: Evidence from
turkey.
C. Zhang, F. In, A. Farley, A multiscaling test of causality effects among international stock markets, Neural, Parallel Sci.
Comput. 12 (1) (2004) 91112.

0.6

1
0
1
Standard Normal Quantiles

(c) level 3

0.8
3

1
0
1
Standard Normal Quantiles

(d) level 4

Fig.S2: Comparison of the quantiles of fluctuations for BSE at various levels of filtering (a) level 1 (b) level 2 (c) level 5 (d) level 6,
with quantiles of normal distribution, showing deviations at both the
ends

QQ Plot of Sample Data versus Standard Normal

QQ Plot of Sample Data versus Standard Normal

0.2

0.15

15

0.15

0.1

15

0.15

1
0
1
Standard Normal Quantiles

0.25
3

1
0
1
Standard Normal Quantiles

10

0.2

0.2
3

0.1
0.15

probability density

0.1

0
0.05

0.05

10

0.05

probability density

Quantiles of Input Sample

Quantiles of Input Sample

0.1
0.05

(a) level 1

(b) level 2

QQ Plot of Sample Data versus Standard Normal

0.6

0.3

0.4

0.0

0.1

0.2

0.3

1
0
1
Standard Normal Quantiles

(d) level 4

Fig.S3: Comparison of the quantiles of fluctuations for NYSE at

various levels of filtering, (a) level 1 (b) level 2 (c) level 5 (d) level
6, with quantiles of normal distribution, showing deviations at both
the ends

0.4

0.0

0.2

4
3

0.4

0.5

0.0

0.5

(d) level 4

Fig.S6: Juxtaposing the probability density of fluctuations for first

6 levels of BSE (blue) and NYSE (red) (a) level 1 (b) level 2 (e) level
5 (f) level 6
level
level
level
level
level
level

15

0.2

(c) level 3

SENSEX high pass filtered data

1
2
3
4
5
6

10

7. ADF and KPSS tests results corresponding to

logarithmic returns

Density

(c) level 3

0.8
3

10

0.4

0.6

0.3

probability density

0.2

0.2

0.1

0.2

0.4 0.3 0.2 0.1

(b) level 2

0.1

1
0
1
Standard Normal Quantiles

0.1

(a) level 1

probability density

Quantiles of Input Sample

Quantiles of Input Sample

0.2

0.0

QQ Plot of Sample Data versus Standard Normal

0.4

0.4
3

0.1

0.1
N = 300

0.0

0.1

Bandwidth = 0.006629

4Yt = 1 Yt1 +

P
X

Fig.S4: Probability density plots of BSE fluctuations obtained for

level 1 to 6, depicting the behaviour of fluctuations at different scales

4Yt = 0 + 1 Yt1 +

(S1)

P
X

j Ytj + t

(S2)

j=1

NYSE high pass filtered data

15

j Ytj + t

j=1

level 1

Density

10

level
2
level
13
level
level 4
level
25
level
level 6
level 3
level 4
level 5
level 6

4Yt = 0 + 1 Yt1 + 2 t +

P
X

j Ytj + t

(S3)

j=1

t is white noise. The test are based on null hypothesis

(H0 ):Yt is not I(0). If the calculated statistic are lower
than the Fullers statistics then the null hypothesis is accepted and the series is non-stationary.

0.10

0.05
N = 300

0.00

0.05

0.10

0.15

Bandwidth = 0.007784

Fig.S5: Probability density plots of NYSE fluctuations obtained for

level 1 to 6, depicting the behaviour of fluctuations at different scales

Yt = ct + t + u1t
2

(S4)

ct = ct1 + u2t ,

(S5)

Table 2: p-values of KPSS test corresponding to logarithmic returns

of BSE and small.

Here, is a trend coefficient, u1t is a stationary process

and u2t is an iid process with mean 0 and variance 2 .

lag
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

Unlike unit root tests, KPSS provides a straightforward

test for stationarity against the alternative of a unit root.
Yt = t + (rt + ) + ut

(S6)

rt = rt1 + ut ,

(S7)

Where,
is a random walk, the initial value r0 = act as an intercept.
ut is independent identical distribution with mean zero
and a non-zero constant variance.
And t is the time.
The model without the trend part is also used to test for
stationarity, i.e = 0, the null hypothesis,
H0 : Yt is trend stationary,
and alternative ; H1 : Yt has a unit root
Table 1: p-values of ADF test corresponding to logarithmic returns
of BSE and NYSE

lag
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

BSE
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001

BSE
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1

NYSE
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1

Table 3: p-values of KPSS trend stationary test corresponding to

logarithmic returns of BSE and small

NYSE
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001

lags
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

BSE
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1

NYSE
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1

8. Shapiro-Wilk Test and KS Test

Table 6: Spearman coefficients between BSE and NYSE monthly

data for both average behaviour and fluctuations for various levels

Table 4: p-values corresponding to average behaviour and fluctuation data ofBSE for SHAPIRO-WILK and KOLMOGOROV-

SMIRNOV (KS) TESTS

SW-test

fluctuation
correlation coefficients

1
2
3
4

0.8969
0.9013
0.9073
0.9121

0.4092
0.3648
0.4473
0.4061

KS-test

level average
behaviour

fluctuation

average
behaviour

fluctuation

1
2
3
4
5
6
7
8

2.65E-15
3.51E-15
1.63E-14
2.62E-18
2.66E-22
9.44E-21
2.08E-06
1.88E-06

1.11E-16
2.22E-16
3.33E-16
0
0
0
3.15E-12
1.90E-06

0
0
0
0
0
0
0
0

1.79E-19
1.65E-19
1.55E-19
1.60E-19
9.12E-20
1.29E-19
4.71E-19
3.10E-10

level

Spearman Coefficients
average behaviour
correlation coefficients

Cov(x, y)
(x y )

(S9)

Pn

((xi x)(yi y)
rp = p Pn i=1
Pn
( i=1 ((xi x)2 )( i=1 ((yi y)2 )
where, x =

The null hypothesis is rejected for both SHAPIRO and

thus low pass as well as high-pass (variations)
obtained at various levels do not belong to normal distribution.

Pn

i=1

xi

and y =

Pn

i=1

(S10)

yi

KS-TEST

Table 7: Pearson correlation coefficients between BSE and NYSE

monthly data for both average behaviour and fluctuations at various
levels

Table 5: p-values corresponding to low and high pass data of NYSE

for SHAPIRO and KOLMOGOROV-SMIRNOV (KS) TESTS.

SW-test
level average
behaviour

fluctuation average
behaviour

fluctuation

1
2
3
4

0.000124
4.82E-07
4.82E-09
1.27E-14

0
0
0
0

3.06E-11
2.12E-11
1.46E-11
7.18E-12

level
1
2
3
4

KS -test

1.11E-16
2.22E-16
3.33E-16
0

Pearson Coefficients
average behaviour
correlation coefficients
0.7769
0.7777
0.7775
0.7807

fluctuation
correlation coefficients
0.4532
0.5025
0.6974
0.7341

10. ADF and KPSS test to check the stationarity

and non-stationarity of first 10 levels of fluctuation

Table 8: ADF and KPSS tests p-value, null hypothesis for ADF and
KPSs tests are H0: signal is non-stationary and signal is stationary
respectively.

9. Spearman and Pearson correlation coefficients

adf test
KPSS test
Pearson correlation coefficients help in establishing linlevel BSE
nyse
H0
BSE
nyse
H0
ear relationship between two variables, if any. It is defined
1
0.01
0.01
reject
0.1
0.1
accept
as the ratio of covariance of two variables to the product of
2
0.01
0.01
reject
0.1
0.1
accept
their respective standard deviation. Whereas, Spearmans
3
0.01
0.01
reject
0.1
0.1
accept
correlation coefficients are rank based version of Pearsons
4
0.01
0.01
reject
0.1
0.1
accept
correlation coefficients ( supplement material).
Pn
((rank(xi ) rank(x))(rank(yi ) rank(y)))
rs = qP i=1
Pn
n
11.2 Vector Auto Regressive (VAR)
2
i=1 ((rank(xi ) rank(x))
i=1 ((rank(yi ) rank(y))
(S8)
Here, rank(xi ) and rank(yi ) are the ranks of the data
The Vector Auto Regressive model is a generalization
points in the sample.
of uni-variate Auto Regression (AR) model, letting us ex4

plore the linear inter-dependencies between multiple variables. In this model evolution of a variable is explained
with the help of its own lag values as well as lags of other
variables.