Professional Documents
Culture Documents
Multi-Scale Correlation and Causality of Stock Exchanges
Multi-Scale Correlation and Causality of Stock Exchanges
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
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
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
= 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
1
0
1984
1987
1990
1993
2004
2006
2009
2012
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].
Z
Z
dt < , dt = 0, dt = 0,
Z
Z
2
|| dt = ||2 dt = 1,
(2)
(3)
j Z+ .
k Z,
(4)
Here, j and k are the scaling and translation parameters [18] (see supplement material).
SENSEX
3
2.5
1500
300
1.5
1
0.5
200
0
0.5
1000
100
1
1985
1990
1995
2000
time(scale)
2005
2010
500
100
200
fluctuation
average behaviour
0.2
0.2
0.5
300
400
500
500
1000
0
2000
4000
6000
(a) BSE
600
1000
2000
3000
4000
5000
X
6000
7000
8000
0.5
0.5
9000
(b) NYSE
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
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
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
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
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
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
0.1
1
NYSE
0.1
0
0.1
1
2
level of filtering
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)
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
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
1.5
1.5
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
(b) level 2
/2
(11)
Density
1.0
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)
2.5
X: 177
Y: 2.163
X: 106
Y: 1.97
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)
NYSE
period(months)
first difference
4
11
15
9
24
32
56
42
98
64
162
SENSEX
scale 11
2
0
SENSEX
scale 26
scale 34
5
1985
1990
1995
2000
2005
2010
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
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
BSE
period(months)
first difference
11
15
26
80
106
71
177
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
amplitude
2
1985
1990
1995
2000
time(month)
2005
2010
(a) scale 11
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
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.
(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
(b) level 2
(d) level 4
10
(a) level 1
(c) level 3
(b) level 2
(d) level 4
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
(14)
(15)
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
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:
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
6. Quantile plots
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
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
[30]
80
60
[29]
[28]
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
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
0.1
0.05
(a) level 1
(b) level 2
0.3
0.4
0.0
0.1
0.2
0.3
1
0
1
Standard Normal Quantiles
(d) level 4
0.4
0.0
0.2
4
3
0.4
0.5
0.0
0.5
(d) level 4
15
0.2
(c) level 3
10
Density
(c) level 3
0.8
3
10
0.4
0.6
0.3
probability density
0.2
0.2
0.1
0.2
(b) level 2
0.1
1
0
1
Standard Normal Quantiles
0.1
(a) level 1
probability density
0.2
0.0
0.4
0.4
3
0.1
0.1
N = 300
0.0
0.1
Bandwidth = 0.006629
4Yt = 1 Yt1 +
P
X
4Yt = 0 + 1 Yt1 +
(S1)
P
X
j Ytj + t
(S2)
j=1
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
0.10
0.05
N = 300
0.00
0.05
0.10
0.15
Bandwidth = 0.007784
Yt = ct + t + u1t
2
(S4)
ct = ct1 + u2t ,
(S5)
lag
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
(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
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
Table 4: p-values corresponding to average behaviour and fluctuation data ofBSE for SHAPIRO-WILK and KOLMOGOROV-
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 =
Pn
i=1
xi
and y =
Pn
i=1
(S10)
yi
KS-TEST
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
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.
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.