Clustering financial time series with variance

ratio statistics∗
João A. Bastos and Jorge Caiado†
CEMAPRE, ISEG, Technical University of Lisbon, 1200-781 Lisboa, Portugal

Forthcoming: Quantitative Finance

Abstract
This study introduces a new distance measure for clustering financial time series
based on variance ratio test statistics. The proposed metric attempts to assess the
level of interdependence of time series from the point of view of return predictabil-
ity. Simulation results show that this metric aggregates better time series according
to their serial dependence structure than a metric based on the sample autocorrela-
tions. An empirical application of this approach to international stock market returns
is presented. The results suggest that this metric discriminates reasonably well stock
markets according to size and level of development. Furthermore, despite the sub-
stantial evolution of individual variance ratio statistics, the clustering pattern remains
fairly stable across different time periods.

1 Introduction
Clustering of time series has become an important tool in many scientific domains, such
as finance and economics, engineering and life sciences. The procedure for clustering time
series typically involves the construction of a convenient similarity measure between the
series. A number of approaches for clustering time series data are available in the litera-
ture, such as autoregressive expansion-based distances (Piccolo, 1990; Maharaj, 1996, 1999,
∗
An earlier version of this paper was circulated under the title “Clustering global equity markets with
variance ratio tests”.
†
Corresponding author. Email: [email protected]

1
2000), autocorrelation-based distances (Galeano and Peña, 2000; Caiado et al., 2006), fit-
ted residuals-based distances (Tong and Dabas, 1990), cross-correlation coefficient distances
(Bohte et al., 1980), periodogram-based distances (Maharaj, 2002; Caiado et al., 2006, 2009),
spectral coherence-based dissimilarities (Maharaj and D’Urso, 2010), dynamic time warping
distances (Berndt and Clifford, 1996; Wang and Gasser, 1997), Markov-operator distances
(Gregorio and Iacus, 2008), short time series distances (Möller-Levet et al., 2003) and cepstral
coefficient-based distances (Kalpakis et al., 2001; Savvides et al., 2008).
Cluster analysis of time series is particularly important in finance, since practitioners are
interested in identifying similarities in financial assets for investment and risk management
purposes. This has motivated financial researchers to develop multivariate statistical meth-
ods to identify similar structural patterns in asset prices. For instance, Mantegna (1999) and
Bonanno et al. (2001) used a function of the Pearson correlation coefficient as a measure of
similarity between pairs of stock returns. In order to take into account the information about
the volatility structure of time series, Caiado and Crato (2010) introduced a Mahalanobis-
like distance between the dynamic features of two return series and employed a clustering
procedure to investigate similarities among stocks of the DJIA index.
In this paper, we introduce a new distance measure for clustering time series with similar
stochastic dependence structure. The proposed metric is based on the distance between
variance ratio statistics computed for individual series. Variance ratios tests are popular
tests of the hypothesis that a time series follows a random walk or a martingale difference
sequence, i.e., that its returns are uncorrelated at all leads and lags. Lo and MacKinlay (1988,
1989) provided the asymptotic sampling theory for both homoscedastic and heteroscedastic
random walks, and showed that these tests are more powerful than traditional tests, such
as serial correlation and unit root tests, against several alternative processes. Wright (2000)
proposed non-parametric variance ratio tests based on ranks and signs. Unlike conventional
variance ratio tests, rank- and sign-based tests are exact, with sampling distributions that
do not rely on asymptotic approximations. Wright (2000) showed that rank- and sign-based
tests improve substantially the power of variance ratio tests with little size distortions. In
recent years, many innovations and refinements of the variance ratio methodology have been
proposed in the literature. An extensive survey of these developments is provided by Charles
and Darné (2009).
To some extent, the proposed metric assesses the level of interdependence of time-series
from the point of view of return predictability. Therefore, a natural empirical application
for this metric is provided by international stock markets. In fact, the level of predictability

2
of global markets is of great importance for investors seeking the reduction of idiosyncratic
risk through international portfolio diversification, and an abundant research has been de-
voted to examining whether the prices of securities conform to a random walk behavior. A
comprehensive review of these empirical studies is provided in the recent survey by Lim and
Brooks (2011). In particular, several studies employed variance ratios to test the random
walk hypothesis in stock markets (see, e.g., Hoque et al., 2007; Kim and Shamsuddin, 2008;
Smith, 2009).
In this study, we analyze daily returns of free float-adjusted market capitalization equity
indices from 46 different countries, covering the period from 1995 to 2009. Two types of
multivariate interdependence techniques are considered for analyzing the clustering patterns
of these markets. First, we employ multidimensional scaling maps, which can be used to
identify similarities between features of different return series, and to construct distances in
a multidimensional space. Then, we perform cluster analysis, which is particularly suited
for defining groups of equity markets with maximal structure dependence within the groups
while also having minimum structure dependence between the groups.
The remainder of this paper is organized as follows. Section 2 presents a brief review
of the variance ratio statistics that are used as inputs for our metric. Section 3 introduces
the clustering procedure and evaluates its properties on simulated data. Section 4 shows the
results of an empirical application to international stock markets. Finally, Section 5 presents
some concluding remarks.

2 Variance ratio tests

In this section, we describe three alternative variance ratio tests of the random walk hypoth-
esis. If a return series conforms to a random walk behavior then it should be uncorrelated
at all leads and lags. Let pt , t = 0, 1, ..., T denote a time-series of asset prices and yt denote
the continuously compounded return at time t, yt = log(pt /pt−1 ), t = 1, ..., T . Given the
time series of asset returns yt = µ + ϵt , where µ is a drift parameter, we want to test the null
hypothesis that: i) ϵt are identical and independently distributed (iid), or ii) ϵt are indepen-
dent and conditional heteroscedastic, that is, the return series forms a martingale difference
sequence.

3
2.1 Conventional variance ratio tests
Lo and MacKinlay (1988) variance ratio tests are based on the property that, if returns
are i.i.d., the variance of the k-period return is k times the variance of the one-period
return. Therefore, if a return series is a random walk the ratio of 1/k times the variance of
log(pt /pt−k ) to the variance of log(pt /pt−1 ) should be close to 1. This variance ratio is given
by ∑T
t=k (yt + yt−1 + ... + yt−k+1 − k µ̂)
1 2
VR(k) = Tk
∑T , (1)
t=1 (yt − µ̂)
1 2
T
∑
where µ̂ = T −1 Tt=1 yt . Lo and MacKinlay (1988) showed that, if the returns are i.i.d. then
the test statistic
M1 (k) = (VR(k) − 1) ϕ(k)−1/2 , (2)

where
2(2k − 1)(k − 1)
ϕ(k) = , (3)
3kT
follows the standard normal distribution asymptotically, under the null hypothesis that
VR(k) = 1. This null asymptotic distribution does not hold if the returns are subject
to conditional heteroscedasticity. Therefore, Lo and MacKinlay (1988) proposed an alter-
native test statistic which is robust against the presence of conditional heteroscedasticity,
given by
[ k−1 [ ]−1/2
∑ 2(k − j) ]2
M2 (k) = (VR(k) − 1) δj , (4)
j=1
k

where ∑T
− µ̂)2 (yt−j − µ̂)2
t=j+1 (yt
δj = [∑ ]2 . (5)
T
(y
t=1 t − µ̂)2

The test statistic M2 (k) also follows the standard normal distribution asymptotically under
the null hypothesis that VR(k) = 1 and the conventional critical values for the standard
normal distribution hold for both tests.

2.2 Rank-based variance ratio tests

The finite-sample null distribution of Lo and MacKinlay (1988) tests can be rather asym-
metric and nonnormal, exhibiting bias and positive skewness. To overcome these problems,
Wright (2000) proposed non-parametric tests based on ranks and signs. These tests have

4
exact sampling distributions and do not recur to any asymptotic approximations. Denote by
r(yt ) the rank of yt among y1 , ..., yT . Under the null hypothesis that ϵt is i.i.d., the integer
sequence r(yt ), t = 1, ..., T , is a random permutation of the integers from 1 to T , where each
permutation has equal probability. Wright (2000) suggests two alternative standardizations
of the ranks,
r(yt ) − T +1
r1t = √ 2
(6)
T 2 −1
12

and ( )
−1 r(yt )
r2t = Φ , (7)
T +1
where Φ is the standard normal cumulative distribution function. The proposed test statistics
are given by
[ ∑T 2
]
1
(r 1t + r1t−1 + ... + r 1t−k+1 )
R1 (k) = Tk t=k
1
∑T 2 − 1 × ϕ(k)−1/2 , (8)
T t=1 r1t

and [ ∑T ]
1
(r2t + r2t−1 + ... + r2t−k+1 )2
R2 (k) = Tk t=k
1
∑T 2 − 1 × ϕ(k)−1/2 . (9)
T t=1 r2t

The exact sampling distribution of R1 (k) and R2 (k) may be derived from simulation to any
arbitrary degree of accuracy. Note that in the presence of conditional heteroscedasticity r(yt ),
t = 1, ..., T , no longer corresponds to a random permutation of the set 1, ..., T with equal
probability and R1 (k) and R2 (k) are not exact. However, through Monte Carlo simulations
Wright (2000) showed that these tests do not exhibit serious size distortions under conditional
heteroscedasticity.

2.3 Sign-based variance ratio tests

Wright (2000) also suggests a sign-based variance ratio test. Let
{
1 if yt > 0
st = (10)
−1 otherwise

5
If µ = 0, the series st is i.i.d. with mean 0 and variance 1. Also, each st is equal to 1 with
probability 21 and equal to −1 otherwise. The sign-based test statistic is given by
[ ∑T 2
]
1
(st + st−1 + ... + st−k+1 )
S1 (k) = Tk t=k
1
∑T 2 − 1 × ϕ(k)−1/2 . (11)
T t=1 st

Again, the exact sampling distribution of S1 (k) may be obtained by simulation. This test
is exact even in the presence of conditional heteroscedasticity. Because the assumption that
µ = 0 is restrictive, Wright (2000) suggested an alternative test in which this condition is
relaxed but his Monte Carlo simulations showed that the power of this test did not compare
well with S1 (k).

3 Cluster analysis with variance ratio statistics

3.1 Variance ratio-based metric
A fundamental task in cluster analysis is to obtain a relevant measure of similarity be-
tween each pair of time series. Here, we propose the Euclidean distance between vectors
of the variance ratio statistics M1 , M2 , R1 , R2 and S1 , introduced in Section 2. Further-
more, these vectors include variance ratios evaluated at several lags k in order to capture
the serial dependence of the returns. To eliminate any bias due to scale differences across
variables, the variance ratios are standardized before computing the distances. Denoting
by vx′ = [VR1x , VR2x , ..., VRpx ] and vy′ = [VR1y , VR2y , ..., VRpy ] the p-dimensional vectors
of standardized variance ratios for time series x and y, respectively, the distance measure
between these vectors is
v
u∑
u p
dVR (x, y) = t (VRjx − VRjy )2 (12)
j=1

Let n denote the number of time series under consideration. We compute dissimilarities
between every pair of series in the data set. The result of this computation is an Euclidean
distance matrix D with n(n − 1)/2 different pairs of time series.

6
3.2 Multidimensional scaling
Multidimensional scaling (MDS) is a multivariate statistical method that uses the informa-
tion about the similarities (or dissimilarities) between objects to construct a configuration of
n points in low-dimensional space (see, for instance, Johnson and Wichern, 2007). Let D be
the observed n × n matrix of Euclidean distances. By multidimensional scaling, the matrix
D yields a n × d configuration matrix T . The rows of T are the coordinates of the n points
in a d-dimensional representation of the observed dissimilarities (d < n). The d-dimensional
representation that best approximates the observed dissimilarity matrix is given by the d
eigenvectors of T T ′ corresponding to the d largest eigenvalues.

3.3 Hierarchical cluster analysis

Cluster analysis attempts to determine groups (or clusters) of objects in a multivariate data
set. The most commonly used clustering algorithm is based on the hierarchical classification
of the objects. This linkage algorithm is concerned with the partition of a set of objects into
groups or clusters, in such a way that objects in the same group are similar to one another
and objects in different clusters are as distinct as possible. We begin with each object being
considered as a separate cluster (n clusters). In the second stage, the closest two groups are
linked to form n − 1 clusters. The process continues until the last stage, in which all the
objects are in the same cluster (for further discussion, see Johnson and Wichern, 2007).
The dendrogram (also called ”cluster tree”) is a graphical representation of the results
of the hierarchical cluster analysis. The dendrogram shows how clusters are formed at each
stage of the procedure. At the bottom of the dendrogram, each object is considered its
own cluster. The objects continue to combine upwards. At the top, all objects are grouped
into a single cluster. In hierarchical clustering, partitions are obtained by cutting off the
dendrogram at an arbitrary point. The choice of the appropriate number of clusters in the
dendrogram is sometimes subjective and depends on the expert judgment of the researchers.
A formal method for finding the appropriate partition in the data set are the Duda-Hart
Je(2)/Je(1) indices (Duda and Hart, 1973). These indices (also called “stopping rules”) are
computed for each cluster solution in a hierarchical cluster analysis. Larger values of the
indices indicate more distinct clustering (for more details, see Everitt et al., 2001).

7
3.4 Simulation results
In order to understand the properties of the variance-ratio based distance dVR (x, y) in clus-
tering time series, we simulated 10 random series from each of the following six processes
(see Wright, 2000):

Model (a): yt = εt , where εt is iid N(0,1).

Model (b): yt = exp(ht /2)εt , where ht = 0.95ht−1 + ξt , with ξt iid N(0,0.1) and independent
of εt .

Model (c): yt = 0.1yt−1 + exp(ht /2)εt ;

Model (d): (1 − L)d yt = exp(ht /2)εt , where (1 − L)d is the fractional differencing operator
and d = 0.1;

Model (e): yt = (1 − 0.5L)−1 υt + exp(ht /2)εt , where υt is iid N(0,0.1), independent of εt ;

Model (f): yt = (1 − L)−d υt + exp(ht /2)εt , where (1 − L)d is the fractional differencing
operator and d = 0.3.

For each series we calculated the variance ratios M1 , M2 , R1 , R2 and S1 evaluated at

four distinct values of k: 2, 5, 10, 20. Then the Euclidean distance dVR (x, y) between all
pairs of 20-dimensional vectors was computed. We also considered three different series
lengths: N = 500, 1000 and 2000. We explore the existence of possible clusters among the
six models by the complete linkage dendrogram associated with the variance ratio distances.
The dendrograms from which clusters can be identified are shown in Figure 1. For large values
of N , the method tends to group the series according to serial dependence structure. For
example, one cluster contains almost all series generated from processes (c) and (d). These
two processes have both serial dependence and conditional heteroscedasticity. On the other
hand, the series generated from processes with no serial dependence (Models (a) and (b)), the
series generated from the sum of a Gaussian AR(1) process and conditional heteroscedastic
noise (Model (e)), and the series generated from the sum of Gaussian fractionally integrated
process with conditional heteroscedastic noise (Model (f)) are, to a large degree, randomly
distributed across multiple clusters.
For comparison, we also consider a well-known discrepancy statistic based on the esti-

8
 20
10 15
Distance
5
0

a e a a e e a e f f b b f f f f b b a e c e c c b c d d d d
a f d b c c b a a f e c e a f b c a d b e b c c e d f d d d
Model (N=500)
40
20 30
Distance
10
0

a e a b e c b e e a e e a f e f a b b f b f d c d c d d d d
a e a b f a f c c e b b b b f a f a e f f c c c c c d d d d
Model (N=1000)
40
20 30
Distance
10
0

a c c c b f e a b e b b e a a b b f b f e e a c c d d d d d
a c c a e b f f b f f b e a e a a f f f e e c c d d c d d d
Model (N=2000)

Figure 1: Complete linkage dendrogram for variance-ratio based distances between the gen-
erated series.

9
mated autocorrelations (Galeano and Peña, 2000):
v
u L
u∑
dACF (x, y) = t ρx,l − ρby,l )2 ,
(b (13)
l=1

where ρbx,l and ρby,l are the sample autocorrelation functions of time series x and y, and L is the
number of autocorrelation lags (in our simulation study, we set L = N/10, as recommended in
Caiado et al. (2006)). Figure 2 shows the complete linkage dendrogram based on metric (13).
From the results given by these simulations, it can be seen that the autocorrelation method
performs poorly in detecting serial dependence. In fact, this metric cannot distinguish models
that have serial dependence and some conditional heteroscedasticity from those that are iid
or mds.
In order to better assess the methods, we have explored other hierarchical (single linkage,
average linkage and Ward’s linkage) and non-hierarchical (k-means) clustering procedures.
Irrespectively of the clustering procedure, the variance ratio based metric provides better
cluster solutions than the ACF based method. We have also obtained multidimensional
scaling solutions for both variance ratio and ACF distances between generated series. Re-
sults were similar to the ones obtained by the hierarchical and non-hierarchical clustering
procedures and provide the same recommendations for the discrepancy statistic choice.

4 An empirical application
4.1 Data
The data employed in this analysis consists of free float-adjusted market capitalization equity
indices constructed and maintained by Morgan Stanley Capital International (MSCI). In
order to avoid effects due to exchange rates, all indices are specified in local currency. The
construction and maintenance of the MSCI index family follows a consistent methodology.
Securities included in the indices are subject to minimum requirements in terms of market
capitalization, free-float, liquidity, availability to foreign investors and length of trading.
The database includes the following 23 developed markets: Australia, Austria, Belgium,
Canada, Denmark, Finland, France, Germany, Greece, Hong Kong, Ireland, Italy, Japan,
Netherlands, New Zealand, Norway, Portugal, Singapore, Spain, Sweden, Switzerland, the
United Kingdom and the United States. A market is classified as developed if i) the coun-
try GNI per capita is 25% above the World Bank high income threshold for 3 consecutive

10
 .6
Distance
.2 0 .4

a c c d a a b a b b f f c a f e d d b d e b c f d b d f e e
a e e a f a b f e c e b e a a c c d c c e b d b d f f c d f
Model (N=500)
.6
Distance
.2 0 .4

a b a f a f b a a c b c c c d e b d d f e f f c b b e e d d
b d c f a a e e f e b f a c e e c d d a b a c c f e f b d d
Model (N=1000)
.5
.2 .3 .4
Distance
.1
0

a c b b a f e d f b f a d d b e c a c e f e f d c d f d d c
b a f a a b b e a f b e a a b c e b f c e e c c e f c d d d
Model (N=2000)

Figure 2: Complete linkage dendrogram for autocorrelation-based based distances between

the generated series.

11
years, ii) there is a minimum number of companies satisfying minimum size and liquidity
requirements, and iii) there is a very high openness to foreign ownership, ease of capital
inflows/outflows, efficiency of the operational framework and stability of institutional frame-
work. The database also includes the following 23 emerging markets: Argentina, Brazil,
Chile, China, Czech Republic, Colombia, Egypt, Hungary, India, Indonesia, Israel, Korea,
Malaysia, Mexico, Morocco, Peru, Philippines, Poland, Russia, South Africa, Taiwan, Thai-
land and Turkey. To be included in the emerging category, a market must satisfy size and
liquidity requirements, and market accessibility criteria that are less tight than those for
their counterparts in developed economies.1
Although for some developed markets the database includes observations that date back
to year 1979, only observations from year 1995 onwards are available for the complete set of
developed and emerging markets. The data employed in the analysis consists of daily index
prices from 1995:01 to 2009:12, corresponding to 3,914 observations.2 In the event of days
where there is a market holiday, the index construction methodology simply carries forward
the index value from the previous business day. In some countries the number of market
holidays can be quite large (for example, the data for Egypt includes 914 market holidays).
Therefore, repeated observations were removed from the data in order to remove potential
biases associated to nontrading days.
Summary statistics (size, annualized mean, annualized standard deviation, skewness and
kurtosis coefficients) of daily percentage rates of return, log(pt /pt−1 ) × 100, where pt is the
index price at time t, for developed and emerging markets are reported in Tables 1 and 2,
respectively. The mean return and the standard deviation of the returns for emerging markets
(10.07% and 29.33%) are higher than those for developed markets (4.29% and 22.95%),
reflecting the risk/return trade-off suggested by finance theory. Several markets (i.e., Ireland,
Japan, Norway, China, Philippines, Taiwan and Thailand) exhibit negative mean returns in
this period. Most developed markets exhibit negative skewness coefficients, indicating that
return distributions in these markets typically have longer negative than positive tails. On
the other hand, almost half of the emerging markets exhibit positive skewness. The returns
series for the developed markets of Austria, Belgium, Canada, Hong-Kong, Ireland, Norway,
New Zealand, Portugal and United States, and for the emerging markets of Brazil, Chile,
Colombia, Czech Republic, Hungary, Indonesia, Malaysia, Philippines, Russia and Thailand
are highly leptokurtic, which means that extreme events occur with increasing frequency.
1
For details see www.mscibarra.com
2
Note that this sample covers the period posterior to the movement towards financial liberalization
experienced by many emerging economies in the late 80’s and early 90’s (see Kim and Singal, 2000).

12
 Market size mean (%) stdev (%) skew kurt
Australia 3797 6.02 16.62 -0.418 9.071
Austria 3713 1.78 22.85 -0.349 13.344
Belgium 3803 1.75 21.44 -0.588 14.154
Canada 3776 8.12 20.08 -0.636 11.563
Denmark 3755 8.89 20.25 -0.332 9.028
Finland 3758 7.71 37.43 -0.360 8.969
France 3804 5.52 22.93 -0.063 7.706
Germany 3795 4.62 24.56 -0.101 7.383
Greece 3742 4.29 27.70 -0.118 7.074
Hong-Kong 3703 3.89 27.02 0.026 11.434
Ireland 3777 -3.18 25.55 -0.722 15.672
Italy 3796 2.85 22.68 -0.062 7.916
Japan 3686 -3.25 22.44 -0.133 8.445
Netherlands 3819 4.10 23.21 -0.182 7.998
Norway 3767 -0.92 18.14 -0.632 18.228
New Zealand 3763 5.83 25.02 -0.564 10.054
Portugal 3770 4.09 17.76 -0.278 11.423
Singapore 3765 1.83 22.80 0.030 8.460
Spain 3774 10.15 23.29 -0.132 7.834
Sweden 3762 8,67 26.66 0.091 6.398
Switzerland 3769 6.25 19.73 -0.102 8.561
United Kingdom 3792 3.60 19.31 -0.174 9.399
United States 3778 6.00 20.34 -0.214 11.125
Average 3768 4.29 22.95 -0.261 10.054

Table 1: Summary statistics of daily percentage rates of return for developed markets: num-
ber of observations (size), annualized mean (mean), annualized standard deviation (stdev),
skewness (skew) and kurtosis (kurt).

13
 Market size mean (%) stdev (%) skew kurt
Argentina 3721 12.58 37.11 -0.066 9.488
Brazil 3711 15.44 33.56 0.333 13.417
Chile 3741 6.62 18.07 0.331 13.081
China 3834 -0.56 34.22 0.035 8.074
Colombia 3637 16.94 22.98 0.178 14.630
Czech Republic 3741 8.52 25.18 -0.359 12.167
Egypt 3000 16.96 26.19 -0.251 7.940
Hungary 3744 16.76 31.95 -0.379 11.200
India 3683 10.42 27.65 -0.139 8.311
Indonesia 3659 11.65 33.97 -0.130 11.277
Israel 3890 8.70 23.62 -0.354 7.843
Korea 3703 6.50 34.07 0.020 6.590
Malaysia 3697 2.11 25.35 0.763 44.721
Mexico 3771 16.44 25.97 0.099 7.844
Morocco 3668 8.69 13.29 -0.063 9.031
Peru 3744 14.97 29.08 -0.136 9.618
Philippines 3706 -1.24 25.95 0.350 12.333
Poland 3755 6.60 29.39 -0.115 5.159
Russia 3841 13.40 52.17 -0.367 12.633
South Africa 3746 9.06 21.42 -0.413 7.811
Taiwan 3690 -0.11 26.92 -0.024 5.030
Thailand 3675 -4.10 37.75 0.661 13.321
Turkey 3731 35.36 43.72 0.034 7.532
Average 3699 10.07 29.33 0.000 11.265

Table 2: Summary statistics of daily percentage rates of return for emerging markets: num-
ber of observations (size), annualized mean (mean), annualized standard deviation (stdev),
skewness (skew) and kurtosis (kurt).

14
4.2 Testing individual markets
The test statistics M1 (k), M2 (k), R1 (k), R2 (k) and S1 (k) were computed using the daily
returns of the 46 indices. Four lags were considered, k = 2, 5, 10, 20, corresponding to two
days, one week, two weeks and one month calendar periods, respectively. Lo and MacKinlay
(1988) and Wright (2000) tests are individual tests in which the random walk hypothesis
is rejected if the test statistic is rejected for any of the pre-defined values of k. Chow and
Denning (1993) argue that performing individual tests using several values of k may lead to
an over rejection of the null-hypothesis above the nominal level of significance. To overcome
the size distortion of individual tests they suggest a joint version of Lo and MacKinlay
(1988) test in which the decision concerning the null hypothesis is taken on the basis of the
maximum absolute value of the vector of test statistics. For instance, if individual tests are
evaluated using m lags ki , i=1,...,m, then the joint test statistics are

M1′ = max |M1 (ki )|, (14)

i=1,...,m

M2′ = max |M2 (ki )|.

i=1,...,m

Chow and Denning (1993) show that M1′ and M2′ follow a studentized maximum modulus
distribution with m and T degrees of freedom. When T is large the critical value is obtained
from the [1 − (α∗ /2)]th percentile of the normal distribution, where α∗ = 1 − (1 − α)1/m and
α is the expected level of significance. In the spirit of Chow and Denning (1993), analogous
joint tests can be devised for rank- and sign-based variance ratio tests

R1′ = max |R1 (ki )|, (15)

i=1,...,m

R2′ = max |R2 (ki )|,

i=1,...,m

S1′ = max |S1 (ki )|.

i=1,...,m

The sampling distribution and critical values of the joint tests R1′ , R2′ and S1′ can be derived
from simulation in the same fashion as they are obtained for the individual tests R1 (k), R2 (k)
and S1 (k). Through Monte Carlo simulations Belaire-Franch and Contreras (2004) show that
these tests have good size and power properties against several stochastic processes. In this
study, the critical values for these tests were simulated through 10,000 replications.
In order to understand how the clustering pattern evolves in the period covered by our
data, we analyze three 5-years sub-samples covering the periods from 1995 to 1999, from

15
 1995-1999 2000-2004 2005-2009
M1′ M2′ R1′ R2′ S1′ M1′ M2′ R1′ R2′ S1′ M1′ M2′ R1′ R2′ S1′
Australia • • • •
Austria • • • •
Belgium • • • • • • •
Canada • • • • • • •
Denmark • • •
Finland • • • •
France • • • • •
Germany
Greece • • • • • • • • • • • •
Hong Kong •
Ireland • • • •
Italy
Japan • •
Netherlands
New Zealand • • • • •
Norway • • •
Portugal • • • • • • •
Singapore • • • • •
Spain • • • • •
Sweden • • •
Switzerland • •
U. Kingdom • • • • • • • • • •
U. States • • • • • •

Table 3: Results of the conventional, rank-based and sign-based variance ratio tests for devel-
oped markets, and for three different periods: 1995-1999, 2000-2004 and 2005-2009. A bullet
(•) indicates that the random walk hypothesis was rejected with a statistical significance of
5%.

16
 1995-1999 2000-2004 2005-2009
M1′ M2′ R1′ R2′ S1′ M1′ M2′ R1′ R2′ S1′ M1′ M2′ R1′ R2′ S1′
Argentina • • • • • • • •
Brazil • • • •
Chile • • • • • • • • • • • • • •
China • • • • • • •
Colombia • • • • • • • • • • • • • •
Czech Rep. • • • • • •
Egypt • • • • • • • • • • • • • • •
Hungary • • • • • • • •
India • • • • • • • • • • • • •
Indonesia • • • • • • • • • • • • • •
Israel • • • • • •
Korea • • • • •
Malaysia • • • • • • • • • • • •
Mexico • • • • • • • • • • • •
Morocco • • • • • • • • • • • • • • •
Peru • • • • • • •
Philippines • • • • • • • • • • • • • •
Poland • • • • •
Russia • • • • • • •
South Africa • • • • • • • • •
Taiwan
Thailand • • • • • • • • • •
Turkey • • • •

Table 4: Results of the conventional, rank-based and sign-based variance ratio tests for
emerging markets, and for three different periods: 1995-1999, 2000-2004 and 2005-2009. A
bullet (•) indicates that the random walk hypothesis was rejected with a statistical signifi-
cance of 5%.

17
2000 to 2004, and from 2005 to 2009.3 The results of testing the random walk hypothesis
with statistics M1′ , M2′ , R1′ , R2′ and S1′ , for developed and emerging markets are shown in
Tables 3 and 4, respectively. For a given test statistic and market, a bullet (•) indicates
that the null hypothesis that the return series follows a random walk is rejected with a
statistical significance of 5%. Tables 3 and 4 provide evidence for the conventional belief
that index returns in emerging markets are typically more predictable that those in developed
markets, as a result of the lower trading volumes and liquidity, and higher levels of regulatory
restrictions. On the other hand, these results do not substantiate Griffin et al. (2010). This
study found small differences between developed and emerging markets, for both stocks and
portfolios, using individual Lo and MacKinlay (1988) variance ratio statistics covering the
period from 1994 through 2005.
Nevertheless, some markets do not conform to this pattern. For instance, in the period
1995-1999 all test statistics reject the null hypothesis in the developed markets of Belgium,
Greece, Portugal, Singapore and Spain, providing strong evidence in favor of predictability
in this period. On the other hand, all tests fail to reject the null hypothesis for the market
Taiwan, which is included in the emerging markets group. Also, from 2000 to 2009, all tests
fail to reject the null for the market of Korea. This is no surprise since these markets have
developed past the emerging market phase, despite the classification as emerging by MSCI.
The evolution of the test results across the three periods suggests a decrease of the
predictability of returns in the past years. In most developed markets and many emerging
markets the number of tests that fail to reject the random walk hypothesis decreases over
time. The number of developed markets in which all tests fail to reject the null hypothesis
increased from 8 in the first period to 14 in the most recent period. In the emerging markets
group, this figure increased from 1 in the period 1995-1999 to 10 in the period 2005-2009.
Remarkably, in the most recent period several tests reject the null in mature markets such
as Australia, Canada, United Kingdom, and the United States.

4.3 Multidimensional scaling maps

Figure 3 shows 2-dimensional scaling maps for the three subperiods. In order to better visu-
alize and interpret similarities among equity markets, we removed outliers which exhibited
very large absolute values of the test statistics. In particular, we dropped the markets of
3
Because the sampling theory of variance ratio tests is based on asymptotic approximations, a minimum
number of observations in each subperiod is necessary. This division results in around 1250 observations in
each period, a sample size that guarantees reasonably high power of the joint variance ratio tests (Belaire-
Franch and Contreras, 2004).

18
Morocco and Colombia from the clustering analysis in subperiods 1995-1999, 2000-2004 and
2005-2009, the market of Chile in subperiods 2000-2004 and 2005-2009, and the markets of
Egypt and Malaysia in subperiod 2005-2009.
The plot on the top of Figure 3 shows the multidimensional scaling (MDS) map for the
first period: 1995-1999. On the right hand side of the map, standing at larger values on
the horizontal axis (Dimension 1), one can identify a cluster containing Canada and the
United States, several developed markets from western Europe (Austria, Denmark, Finland,
France, Germany, Italy, Netherlands, Norway, Spain, Sweden, Switzerland and the United
Kingdom) and the most developed markets from the Pacific Rim (Australia, Hong-Kong,
Korea, Japan, New Zealand and Taiwan). The comparison of this cluster with the leftmost
panels in Tables 3 and 4 provides an interesting result. While this cluster contains most
developed markets in which at least one test statistic failed to reject the null hypothesis that
the return series conforms to a random walk, it only contains the single emerging market
in which all test statistics failed to reject the null (i.e., Taiwan). The exceptions to this
pattern are the markets of Spain, for which all tests reject the null, and Ireland which stands
between the main cluster of developed markets and the remaining developed markets.
In contrast to the developed markets, many emerging markets in which at least one test
statistic fails to reject the null (i.e., Brazil, Hungary, Israel, Malaysia, Mexico, South Africa
and Turkey) are displaced from this cluster. On the other hand, while all test statistics
fail to reject the null in the developed market of Spain, nevertheless, it is grouped with
the cluster of developed markets. Therefore, it appears that this procedure can capture
similarities between markets with similar levels of development that single variance ratio
tests fail to detect. On the left of this cluster, scattered more or less evenly across lower
values of Dimension 1, one can find most emerging markets and four smaller developed
markets (Belgium, Greece, Portugal and Singapore). Larger values of Dimension 1 appear
to be directly related to lower levels of stock market predictability and, in fact, this dimension
explains 86.7% of the total variance of the scaled data.
The plot in the middle of Figure 3 shows the MDS map for the period 2000-2004. The
visual inspection of this map reveals a large cluster containing all developed markets with
the exception of Greece. Corroborating this observation, Table 3 shows that Greece is the
only developed country in which all test statistics failed to reject the null hypothesis that the
return series is a random walk. We can find eight emerging markets with Dimension 1 values
within the range of the developed markets (Brazil, China, Hungary, Korea, Poland, Russia,
Turkey and Taiwan). Interestingly, while exactly the same tests reject the null hypothesis

19
 1995−1999
2 EGY RUS TUR
TAI
Dimension 2 (6.0%)

MAL
1

CR ISR SWI
HUN FIN
HK GER
ITA AUST
DEN SWE
CHIL POR SPA NET
AUS US
0

BRA NZ FRA
SING IRE JAP
NOR
THA IND BEL CAN
CHIN POL KOR
SA MEX
PER
−1

PHI
INDO
GRE
ARG
UK
−2

−8 −6 −4 −2 0 2 4 6
Dimension 1 (86.7%)
2000−2004
2

RUS JAP
SA
MEX
SWE
KOR
Dimension 2 (6.9%)
1

GRE ISR
BEL
CHIN AUST
IND PHI NOR
CR BRA CAN FRA
INDO FIN NET UK
0

EGY TAI HK
DEN SWI
THA IRE SPA
PER NZ
HUN GER
MAL POR AUS
SING
−1

TUR US
POL ITA
ARG
−2

−8 −6 −4 −2 0 2 4 6
Dimension 1 (84.1%)
2005−2009
3

MEX
2

CHIN
Dimension 2 (9.9%)

IRE
NOR IND
JAP NETKOR
1

SPA ISR POR

SWI SA
GER CR PER
ARG
SWE FINBRA THA INDO
0

CAN ITA
UK AUS
FRA RUS GRE
US AUST PHIHUN
−1

NZ SING
DEN
HK TUR
TAI
−2

POL
BEL
−3

−8 −6 −4 −2 0 2 4 6
Dimension 1 (73.4%)

Figure 3: Two-dimensional maps of global stock markets by metric multidimensional scaling.

20
 1995−1999

15
Distance
10
5
0
HK TAI FIN SPA KOR AUS MEX DEN CAN ISR NETAUST JAP US POL PER IND SING BEL CHIL THA RUS
ITA SWE SWI ARG NZ IRE HUN BRA TUR MAL GER FRA NOR UK GRE PHI SA INDOCHIN CR POR EGY
Market
2000−2004
15
Distance
10
5
0

FIN KOR DEN TUR CAN BRA BEL RUS UK GER SPA HK NZ SING AUS ISR PER THA PHI MEX GRE MAL
AUSTHUN POL ITA SWI JAP SWE FRA US NET NOR CHIN IRE TAI CR POR ARG INDO SA EGY IND
Market
2005−2009
10
Distance
5
0

CHIN KOR SA ITA HK SPA NOR NET TAI RUS GRE PER INDO AUS ARG BEL UK CAN FIN SWI US
MEX ISR IRE SING JAP GER CR TUR DEN POL HUN IND PHI THA POR FRA AUST SWE BRA NZ
Market

Figure 4: Dendrograms for Euclidean distances between international stock markets.

in the markets of France and Argentina, the former is found in the cluster of developed
markets and the latter is found at a considerable distance from this cluster. Finally, the plot
in the bottom of Figure 3 shows the MDS map for the most recent period. Apart from a
reflection about the vertical axis, this map reveals a clustering pattern resembling those of
the previous periods. A cluster containing most developed markets can be found at negative
values of Dimension 1. As in the first period, the three European markets of Belgium,
Greece and Portugal stand in the cluster of emerging markets. Interestingly, the market of
the United States is the farthermost from the emerging markets cluster, despite having four
tests rejecting the null during this period. Among the cluster of developed markets one can
find the emerging markets of Brazil, Czech Republic and South Africa.

21
 Duda-Hart Clusters
Period #clusters Je(2)/Je(1) pseudo-t2 1 2 3 4 5
1995-1999 1 0.4108 60.24 44
2 0.3945 21.48 28 16
3 0.5650 20.02 28 9 7
4 0.5133 4.74 20 9 8 7
5 0.7229 6.90 20 9 8 6 1
2000-2004 1 0.4465 50.83 43
2 0.5276 17.01 22 21
3 0.5685 15.18 22 17 4
4 0.5829 10.74 17 15 7 4
5 0.6343 7.49 15 12 7 5 4
2005-2009 1 0.5448 32.59 41
2 0.6494 15.65 31 10
3 0.7792 5.10 20 11 10
4 0.6732 3.88 14 11 10 6
5 0.5844 8.53 14 11 9 6 1

Table 5: Cluster solutions for Duda-Hart Je(2)/Je(1) index. The values in the rightmost
cells are the number of markets for each cluster solution.

4.4 Dendrogram analysis

Figure 4 shows the dendrograms for periods 1995-1999, 2000-2004, and 2005-2009 obtained
by the complete linkage method, which minimizes the maximum distance between equity
markets in the same group. We have computed the Duda-Hart Je(2)/Je(1) indices and the
associated pseudo-t2 statistics to determine the optimal cluster solutions, as shown in Table 5.
The appropriate number of clusters is determined by the largest Duda-Hart Je(2)/Je(1)
values. The results in Table 5 suggest five clusters for periods 1995-1999 and 2000-2004, and
three clusters for period 2005-2009.
Table 6 shows how the stock markets are grouped in these clusters. In period 1995-1999,
cluster 1 includes 11 developed markets (Austria, Canada, Denmark, Finland, Hong-Kong,
Ireland, Italy, New Zealand, Spain, Sweden and Switzerland) and nine emerging markets
(Argentina, Brazil, Hungary, Israel, Korea, Malaysia, Mexico, Taiwan and Turkey). In
cluster 2 we can find seven emerging markets (India, Indonesia, Peru, Philippines, Poland
and South Africa) and three developed markets (Belgium, Greece and Singapore). Cluster 3
includes the indices with the largest market capitalizations, such as the United States, Japan,
United Kingdom, Germany and France, together with the Netherlands and Norway. Cluster
4 contains five emerging markets (Chile, China, Czech Republic, Russia and Thailand) and

22
 1995 - 1999
cluster 1 Austria Canada Denmark Finland Hong-Kong Ireland Italy New Zealand
Spain Sweden Switzerland Argentina Brazil Hungary Israel Korea
Malaysia Mexico Taiwan Turkey
cluster 2 Belgium Greece Singapore India Indonesia Peru Philippines Poland S.
Africa
cluster 3 France Germany Japan Netherlands Norway United Kingdom United
States
cluster 4 Portugal Chile China Czech R. Russia Thailand
cluster 5 Egypt
2000 - 2004
cluster 1 Australia Belgium Canada Denmark Finland Italy Japan Sweden Switzer-
land Brazil Hungary Korea Poland Russia Turkey
cluster 2 Austria Ireland Hong-Kong New Zealand Portugal Singapore Argentina
China Czech R. Israel Peru Taiwan
cluster 3 France Germany Norway Netherlands Spain United Kingdom United
States
cluster 4 Indonesia Mexico Philippines S. Africa Thailand
cluster 5 Greece Egypt India Malaysia
2005 - 2009
cluster 1 Australia Canada Finland France New Zealand Sweden Switzerland
United Kingdom United States Brazil
cluster 2 Austria Belgium Greece Portugal Argentina Hungary India Indonesia
Peru Philippines Thailand
cluster 3 Denmark Germany Hong Kong Ireland Italy Japan Netherlands Norway
Singapore Spain China Czech R. Israel Korea Mexico Poland S. Africa
Taiwan Turkey

Table 6: Distribution of stock markets in the cluster solutions identified with the Duda-Hart
stopping-rule index.

23
one developed market (Portugal). Finally, cluster 5 is formed by the market of Egypt. Note
that, in the MDS map corresponding to this period this market can be found at a large
distance from all other markets.
The clustering results over the period 2000-2004 suggest again five distinct groups. Clus-
ter 1 includes nine developed markets (Australia, Belgium, Canada, Denmark, Finland, Italy,
Japan, Sweden and Switzerland) and six emerging markets (Brazil, Hungary, Korea, Poland,
Russia and Turkey). In cluster 2 we can find six emerging markets (Argentina, Czech Repub-
lic, China, Israel, Peru, Taiwan), three developed European markets (Austria, Ireland and
Portugal) and three developed Pacific Rim markets (Hong-Kong, New Zealand and Singa-
pore). Again, cluster 3 groups many of the markets with the largest market capitalizations.
Interestingly, Japan is no longer included in this group. This is consistent with the corre-
sponding MDS map, in which the market of Japan is found at a reasonable distance from
other large capitalization markets. Cluster 4 only contains emerging markets (Indonesia,
Mexico, Philippines, South Africa and Thailand), and cluster 5 is formed by the markets
with lowest Dimension 1 values in the MDS map: Egypt, India, Malaysia and Greece.
In contrast to the previous periods, the cluster solution for period 2005-2009 indicates a
separation of equity markets into three clusters. One cluster is formed by most developed
markets (Australia, Canada, Finland, France, New Zealand, Sweden, Switzerland, United
Kingdom and United States) and Brazil. A second cluster is formed by seven emerging
markets (Argentina, Hungary, India, Indonesia, Peru, Philippines and Thailand) and four
developed European markets (Austria, Belgium, Greece and Portugal). A third cluster is
formed by the remaining developed and emerging markets.

5 Conclusions
In this paper, we introduced a new distance measure for clustering time series which is
based on variance ratio statistics. The proposed metric attempts to gauge the level of
interdependence of time series from the point of view of return predictability. Simulation
results show that a variance ratio-based metric aggregates better time series according to
their serial dependence structure than a metric based on the sample autocorrelations. An
empirical application of this approach to a large set of 46 international stock market returns
was presented. The examination of individual variance ratio statistics provided further
evidence for the conventional belief that the returns of emerging stocks markets are usually
more predictable than those of their developed counterparts. It was also found that in the

24
period covered by the data there was a trend towards greater unpredictability of returns in
both developed and emerging markets. Groups of markets were identified by visual inspection
of two-dimensional scaling maps and by cluster solutions given by the Duda-Hart stopping-
rule index. The propose metric clustered reasonably well the stock markets according to
their size and level of development. It also provided interesting insights that the analysis of
single variance ratios failed to capture.
While this analysis was conducted using specific variance ratio tests and pre-determined
lag orders, future studies can explore alternative variance ratio tests, such as those with
automatic selection of the optimal holding period (see, Kim, 2009; Charles et al., 2011).
Furthermore, future studies could attempt to relate our clustering metric to factors that
may affect market efficiency, such as regulatory restrictions, freedom of capital movements,
and trade openness (see, Lim and Kim, 2011).

Acknowledgments
This study was supported by a grant from FCT - Fundação para a Ciência e Tecnologia.

References
Belaire-Franch, J., Contreras, D., 2004. Ranks and signs-based multiple variance ratio tests.
Working paper. University of Valencia.

Berndt, D.J., Clifford, J., 1996. Finding patterns in time series: a dynamic programming
approach. In: Advances in Knowledge Discovery and Data Mining (Ed.: U.M. Fayyad, G.
Piatetsky-Shapiro, P. Smyth and R. Uthurusamy). AAAI Press/The MIT Press, Menlo
Park, 229–248.

Bohte, Z.D., Cepar, D., Kosmelu, K., 1980. Clustering of time series, COMPSTAT 80,
Physica-Verlag, 587-593.

Bonanno, G., Lillo, F., Mantegna, R., 2001. High-frequency cross-correlation in a set of
stocks. Quantitative Finance 1, 96-104.

Caiado, J., Crato, N., Peña, D. (2006). A periodogram-based metric for time series classifi-
cation. Computational Statistics & Data Analysis 50, 2668-2684.

25
Caiado, J., Crato, N., Peña, D. (2009). Comparison of time series with unequal length in
the frequency domain. Communications in Statistics - Simulation and Computation 38,
527-540.

Caiado, J., Crato, N., 2010. Identifying common dynamic features in stock returns. Quanti-
tative Finance 10, 797-807.

Charles, A., Darné, O., 2009. Variance-ratio tests of random walk: an overview. Journal of
Economic Surveys 23, 503527.

Charles, A., Darné, O., Kim, J.H., 2011. Small sample properties of alternative tests for
martingale difference hypothesis. Economics Letters 110, 151-154.

Chow, K.V., Denning, K.C., 1993. A simple multiple variance ratio test. Journal of Econo-
metrics 58, 385-401.

Duda, R.O., Hart, P.E., 1973. Pattern Classification and Scene Analysis. Wiley, New York.

Everitt, B., Landau, S., Leese, M., 2001. Cluster Analysis, 4th Ed. Edward Arnold, London.

Galeano, P., Peña, D. (2000). Multivariate analysis in vector time series. Resenhas 4, 383-404.

De Gregorio, A., Iacus, S.M., 2008. Clustering of discretely observed diffusion processes.
Computational Statistics & Data Analysis 54, 598-606.

Griffin, J.M., Kelly, P.J., Nardari, F., 2010. Do market efficiency measures yield correct
inferences? A comparison of developed and emerging markets. Review of Financial Studies
23, 32253277.

Johnson, R., Wichern, D., 2007. Applied Multivariate Statistical Analysis, 6th Ed. Prentice-
Hall, New Jersey.

Kalpakis, K., Gada, D., Puttagunta, V., 2001. Distance measures for effective clustering of
ARIMA time-series. Proceedings of the IEEE International Conference on Data Mining,
San Jose, 273-280.

Kim, J.H., 2009. Automatic variance ratio test under conditional heteroskedasticty. Finance
Research Letters 6, 179-185.

Kim, E.H., Singal, V., 2000. Stock market openings: experience of emerging economies. The
Journal of Business 73(1), 25-66.

26
J.H. Kim and A. Shamsuddin, 2008. Are Asian stock markets efficient Evidence from new
multiple variance ratio tests. Journal of Empirical Finance 15, 518-532.

Hoque, H.A.A.B., Kim, J.H., Pyun, C.S., 2007. A comparison of variance ratio tests of
random walk: A case of Asian emerging stock markets. International Review of Economics
and Finance 16, 488–502.

Lim, K.-P., Brooks, R., 2011. The evolution of stock market efficiency over time: a survey
of the empirical literature. Journal of Economic Surveys 25, 69108.

Lim, K.-P., Kim, J.H., 2011. Trade openness and the informational efficiency of emerging
stock markets. Economic Modelling 28, 2228-2238.

Lo, A.W., MacKinlay, A.C., 1988. Stock market prices do not follow random walks: Evidence
from a simple specification test. Review of Financial Studies 1(1), 41–66.

Lo, A.W., MacKinlay, A.C., 1989. The size and power of the variance ratio test in finite
samples: A Monte Carlo investigation. Journal of Econometrics 40(2), 203–238.

Maharaj, E.A., 1996. A significance test for classifying ARMA models. Journal of Statistical
Computing and Simulation 54, 305-331.

Maharaj, E.A., 1999. Comparison and classifying of stationary multivariate time series.
Pattern Recognition 32, 1129-1138.

Maharaj, E.A., 2000. Clusters of time series. Journal of Classification 17, 297-314.

Maharaj, E.A., 2002. Comparison of non-stationary time series in the frequency domain.
Computational Statistics & Data Analysis 40, 131-141.

Maharaj, E.A., D’Urso, P.A., 2010. A coherence-based approach for the pattern recognition
of time series. Physica A: Statistical mechanics and its Applications 389, 3516-3537.

Mantegna, R.N., 1999. Hierarchical structure in financial markets. The European Physical
Journal B, 11, 193-197.

Möller-Levet, C. S., Klawonn, F., Cho, K. and Wolkenhauer, O. (2003). Fuzzy clustering
of short time-series and unevenly distributed sampling points In: Advances in Intelligent
Data Analysis V, volume 2810 of Lecture Notes in Computer Science, Springer Verlag,
330-340.

27
Piccolo, D., 1990. A distance measure for classifying ARIMA models. Journal of Time Series
Analysis 11, 152-164.

Savvides, A., Promponas, V. and Fokianos, K., 2008. Clustering of biological time series by
cepstral coefficients based distances. Pattern Recognition 41, 2398-2412.

Smith, G., 2009. Martingales in European emerging stock markets: Size, liquidity and market
quality. The European Journal of Finance 15, 249-262.

Tong, H., Dabas, P., 1990. Cluster of time series models: an example. Journal of Applied
Statistics 17, 187-198.

Wang, K., Gasser, T., 1997. Alignment of curves by dynamic time warping. Annals of Statis-
tics 25, 1251-1276.

Wright, J.H., 2000. Alternative variance-ratio tests using ranks and signs. Journal of Business
and Economics Statistics 18(1), 1-9.

28