doi: 10.1111/j.1467-6419.2008.00570.

x
VARIANCE-RATIO TESTS OF RANDOM
WALK: AN OVERVIEW
Am elie Charles
Audencia Nantes, School of Management
Olivier Darn e
LEMNA, University of Nantes
Abstract. This paper reviews the recent developments in the field of the variance-
ratio (VR) tests of the random walk and martingale hypothesis. In particular,
we present the conventional individual and multiple VR tests as well as their
improved modifications based on power-transformed statistics, rank and sign
tests, subsampling and bootstrap methods, among others. We also re-examine the
weak-form efficiency for five emerging equity markets in Latin America.
Keywords. Random walk hypothesis; Stock market efficiency; Variance-ratio tests
1. Introduction
There exists a long tradition in the literature concerning the test of the random walk
and martingale hypothesis, both in macroeconomics and finance. For instance,
the random walk hypothesis (RWH) provides a means to test the weak-form
efficiency and hence, non-predictability of financial markets (Fama, 1970,
1991), and to measure the long-run effects of shocks on the path of real output in
macroeconomics (Campbell and Mankiw, 1987; Cochrane, 1988; Cogley, 1990).
Given a time series {y
t
}
T
t =1
, the RWH corresponds to = 1 in the first-order
autoregressive model
y
t
= +y
t 1
+
t
where is an unknown drift parameter and the error terms
t
are, in general,
neither independent nor identically distributed (i.i.d.).
1
Many statistical tests
2
were designed to test the RWH but a class of test, based on
the variance-ratio (VR) methodology, has gained tremendous popularity in recent
years (see, for example, Campbell and Mankiw, 1987; Cochrane, 1988; Lo and
MacKinlay, 1988; Poterba and Summers, 1988). The VR methodology consists
of testing the RWH against stationary alternatives, by exploiting the fact that
the variance of random walk increments is linear in all sampling intervals, i.e.
the sample variance of the k-period return (or k-period differences), y
t
y
t k
,
of the time series y
t
, is k times the sample variance of the one-period return (or the
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd, 9600 Garsington Road,
Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA.
504 CHARLES AND DARN

E
first difference), y
t
y
t 1
. The VR at lag k is then defined as the ratio between
1/k of the k-period return (or the kth difference) to the variance of the one-period
return (or the first difference). Hence, for a random walk process, the variance
computed at each individual lag interval k (k = 2, 3, . . .) should be equal to unity.
The use of the VR statistic can be advantageous when testing against several
interesting alternatives to the random walk model, most notably those hypotheses
associated with mean reversion. In fact, a number of authors (e.g. Lo and
MacKinlay, 1989; Richardson and Smith, 1991; Faust, 1992) found that the VR
statistic had optimal power against such alternative.
However, while the intuition behind the VR test is rather simple, conducting a
statistical inference using the VR test is less straightforward. What makes things
complicated is that the VR test typically uses overlapping data in computing the
variance of long-horizon returns. The use of overlapping data was suggested by
Lo and MacKinlay (1988) because it can potentially improve the power of the VR
test, but the use of overlapping data also adds to the difficulties of analysing the
exact distribution of the VR test statistic. However, virtually nothing is often known
about the exact distribution of the VR test statistic that uses overlapping data, and
not even its moments are known.
3
In practice, an asymptotic distribution instead of
exact distribution is often used for conducting statistical inference on the VR test,
for fixed k and the sample size T increasing to infinity.
Lo and MacKinlay (1988) proposed two statistics for testing an individual
VR estimate, which are robust under homoscedasticity and heteroscedasticity. In
practice, it is customary to examine the VR statistics for several k values. The null
is rejected if it is rejected for some k value. However, as stressed by Chow and
Denning (1993), this sequential procedure leads to an oversized testing strategy. In
this context, multiple VR tests have been suggested such as multiple comparison
tests (Chow and Denning, 1993) and Wald-type joint tests (Richardson and Smith,
1989; Cecchetti and Lam, 1994). Even though the individual LoMacKinlay and
multiple VR tests are quite powerful testing for homoscedastic or heteroscedastic
nulls, it is critical to note that these tests are asymptotic tests in that their sampling
distributions are approximated by their limiting distributions. Indeed, the sampling
distribution of the VR statistic can be far from normal in a finite sample, showing
severe bias and right skewness. These finite sample deficiencies may give rise to
serious size distortions or low power, which can lead to misleading inferences. This
is especially true when the sample size is not large enough to justify asymptotic
approximations (Cecchetti and Lam, 1994). To circumvent this problem, some
alternatives
4
have been proposed, such as Chen and Deo (2006) with a power-
transformed VR statistic, Wright (2000) with exact VR tests based on rank and
sign, Whang and Kim (2003) with a subsampling method and Kim (2006) with a
bootstrap method, among others. Therefore, because of the important literature on
the VR tests we propose an overview on this subject.
The rest of this paper is organized as follows. The individual and multiple
VR tests are presented in Sections 2 and 3, respectively. Section 4 discusses
bootstrapping VR tests. An empirical illustration is proposed in Section 5. Section 6
concludes.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 505
2. Individual Variance Ratio Tests
The VR test is often used (see Cochrane, 1988; Lo and MacKinlay, 1988; Poterba
and Summers, 1988; among others) to test the hypothesis that a given time series
or its first difference (or return), x
t
= y
t
y
t 1
, is a collection of i.i.d. observations
or that it follows a martingale difference sequence. Define the VR of the k-period
return as
V(k) =
var(x
t
+ x
t 1
+ + x
t k+1
)/k
var(x
t
)
=
var(y
t
y
t k
)/k
var(y
t
y
t 1
)
= 1 +2
k1

i =1
(k i )
k

i
where
i
is the ith lag autocorrelation coefficient of {x
t
}. V(k) is a particular linear
combination of the first k 1 autocorrelation coefficients, with linearly declining
weights. The central idea of the VR test is based on the observation that when
returns are uncorrelated over time, we should have var(x
t
+ + x
t k+1
) = k
var(x
t
), i.e. V(k) = 1. One can therefore think of the VR test as a specification test
of H
0
:
1
= =
k
= 0, i.e. returns are serially uncorrelated.
A test can be constructed by considering the statistic based on an estimator of
V(k)
VR(k) =

2
(k)

2
(1)
where
2
(1) is the unbiased estimator of the one-period return variance, using the
one-period returns x
t
, and is defined as

2
(1) = (T 1)
1
T

t =1
(x
t
)
2
= (T 1)
1
T

t =1
(y
t
y
t 1
)
2
(1)
where = T
1

T
t =1
x
t
is the estimated mean. There are many ways to obtain
the estimator of the k-period returns variance
2
(k), using k-period returns
(x
t
+ + x
t k+1
). Due to limited sample size and the desire to improve the
power of the test, this estimator is often performed using overlapping long-horizon
returns (k-period), as advocated by Lo and MacKinlay (1988),
5
and it is defined as

2
(k) = m
1
T

t =k
(x
t
+ x
t 1
+ + x
t k+1
k )
2
= m
1
T

t =k
(y
t
y
t k
k )
2
where m = k(T k + 1)(1 kT
1
). The value of m is chosen such that
2
(k) is an
unbiased estimator of the k-period return variance when
2
t
is constant over time.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
506 CHARLES AND DARN

E
Following Wright (2000), the VR statistic can be written as
VR(x; k) =
_
(Tk)
1
T

t =k
(x
t
+ + x
t k+1
k )
2
_

_
T
1
T

t =1
(x
t
)
2
_
(2)
Moreover,Cochrane (1988) showed that the estimator of V(k) can be interpreted
in terms of the frequency domain. This estimator that uses the usual consistent
estimators of variance is asymptotically equivalent to 2 times the normalized
spectral density estimator at the zero frequency, which uses the Bartlett kernel.
Formally, we have
VR
f
= 2
f
y
(0)

2
(1)
(3)
where f
y
(0) represents the estimator of the spectrum evaluated at frequency 0
with
2f
y
(0) = 2
m

j =1
W
k
(
j
)I
y
(
j
)
I
y
= (2T)
1
|d
y
(
j
)|
2
d
y
=
T1

t =1
(y
t
y
t 1
)e
i
j
t
W
k
(
j
) =

| j |k
(1 | j |/k) exp(ij) = k
1
_
sin(k
j
/2)
sin(
j
/2)
_
2
(4)
where
j
= 2j T
1
, j = 1, . . . , T 1, I
y
denotes the periodogram, d
y
is the
discrete Fourier transform, is an estimate of the mean of y
t
and W(
j
) is the
Bartlett window.
If the data-generating process of time series is a random walk, the expected value
of VR(x; k) should be equal to unity for all horizons k. If returns are positively
(negatively) autocorrelated, the VR should be higher (lower) than unity. A time
series (in level) is said to be mean reverting if VR(x; k) is significantly lower than
unity at long horizons k. In contrast, a time series is mean averting, i.e. explosive, if
VR(x; k) is significantly higher than unity at long horizons (Poterba and Summers,
1988).
We describe the most popular individual VR tests developed by Lo and
MacKinlay (1988) as well as some of their improvements.
2.1 Lo and MacKinlay (1988) Tests
Lo and MacKinlay (1988) proposed the asymptotic distribution of VR(x; k) by
assuming that k is fixed when T . They showed that if x
t
is i.i.d., i.e. under
the assumption of homoscedasticity, then under the null hypothesis that V(k) = 1,
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 507
the test statistic M
1
(k) is given by
M
1
(k) =
VR(x; k) 1
(k)
1/2
(5)
which follows the standard normal distribution asymptotically. The asymptotic
variance, (k), is given by
(k) =
2(2k 1)(k 1)
3kT
(6)
To accommodate x
t
s exhibiting conditional heteroscedasticity, Lo and MacKinlay
(1988) proposed the heteroscedasticity
6
robust test statistic M
2
(k)
M
2
(k) =
VR(x; k) 1

(k)
1/2
(7)
which follows the standard normal distribution asymptotically under the null
hypothesis that V(k) = 1, where

(k) =
k1

j =1
_
2(k j )
k
_
2
( j )
( j ) =
_
_
_
T

t =j +1
(x
t
)
2
(x
t j
)
2
_
_
_

_
_
_
_
T

t =1
(x
t
)
2
_
2
_
_
_
The M
2
(k) test is applicable to x
t
s generated from a martingale difference time
series (see the Appendix for a discussion on the assumptions). The usual decision
rule for the standard normal distribution is applied to both tests.
The finite-sample properties of the VR test were studied by Lo and MacKinlay
(1989), who found that the two-sided test has size generally quite close to
the nominal level, as long as the test is robustified against any conditional
heteroscedasticity. The VR statistic has been found by several authors (e.g.
Richardson and Smith, 1991; Faust, 1992) to be particularly powerful when testing
against mean reverting alternatives to the random walk model, particularly when k
is large.
2.2 Chen and Deo (2006) Test
It is critical to note that the conventional VR tests, such as the LoMacKinlay test,
are asymptotic tests in that their sampling distributions are approximated by their
limiting distributions. Indeed, the practical use of the statistic has been impeded
by the fact that the asymptotic theory provides a poor approximation to the small-
sample distribution of the VR statistic. In general, the ability of the asymptotic
distribution to approximate the finite-sample distribution depends crucially on the
value of the horizon k. More specifically, rather than being normally distributed
(when standardized by

T) as the theory states, the statistics are severely biased
and right skewed for large k (relative to T) (Lo and MacKinlay, 1989), which
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
508 CHARLES AND DARN

E
makes application of the statistic problematic. In other words, the finite-sample
null distribution of the test statistic is quite asymmetric and non-normal.
A solution is provided in a series of theoretical papers such as those by
Richardson and Stock (1989), Deo and Richardson (2003), Perron and Vodounou
(2005) and Chen and Deo (2006).
7
For example, to circumvent this problem,
Richardson and Stock (1989) provided an alternative asymptotic distribution of the
VR statistic under the random walk null, assuming that both k and T increase
to infinity but in such a way that k/T converges to a positive constant that is
strictly less than 1.
8
Through Monte Carlo simulations, they demonstrated that this
new distribution provides a far more robust approximation to the small-sample
distribution of the VR statistic. Most current applications of the VR statistic cite
the k/T > 0 result as justification for using Monte Carlo distributions (i.e. set
at k = T) as representative of the VR statistics sampling distribution. Perron and
Vodounou (2005) also studied the VR statistics properties under the Richardson
and Stock (1989) framework, and characterized the maximal possible power by
taking a continuous-time limit given a fixed data span T.
However, Deo and Richardson (2003) argued that the VR statistic is inconsistent
against an important class of mean reverting alternatives under this framework
when the horizon k is increasing proportional to the sample size, i.e. k/T >
0. Chen and Deo (2006) also showed that the k/T > 0 asymptotic distribution
cannot approximate the finite-sample distribution of the VR statistic when k/T is
small and is sensitive to conditional heteroscedasticity.
Chen and Deo (2006) suggested a simple power transformation of the VR statistic
that, when k is not too large,
9
provides a better approximation to the normal
distribution in finite samples and is able to solve the well-known right-skewness
problem. They showed that the transformed VR statistic leads to significant gains
in power against mean reverting alternatives. Furthermore, the distribution of the
transformed VR statistic is shown, both theoretically and through simulations, to
be robust to conditional heteroscedasticity.
10
First, they defined the VR statistic based on the periodogram as
VR
p
(k) =
1
1 k/T
4
T
2
(T1)/2

j =1
W
k
(
j
)I
y
(
j
) (8)
where I
y
(
j
) and W
k
(
j
) are defined as in (4). This expression of the VR statistic
is precisely the normalized discrete periodogram average estimate of the spectral
density of a stationary process at the origin (Brockwell and Davis, 1996). To
obtain their transformed VR statistic, noted VR

p
(k), they applied the following
power transformation
11
to VR
p
(k):
= 1
2
3
_

(T1)/2
j =1
W
k
(
j
)
_ _

(T1)/2
j =1
W
3
k
(
j
)
_
_

(T1)/2
j =1
W
2
k
(
j
)
_
2
(9)
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 509
2.3 Wright (2000) Tests
As already noted, the LoMacKinlay tests, which are asymptotic tests whose
sampling distribution is approximated based on its limiting distribution, are biased
and right-skewed in finite samples. In this respect, Wright (2000) proposed a
non-parametric alternative to conventional asymptotic VR tests using signs and
ranks. Wrights (2000) tests have two advantages over the LoMacKinlay test
when sample size is relatively small: (1) as the rank (R
1
and R
2
) and sign (S
1
and S
2
) tests have an exact sampling distribution, there is no need to resort to
asymptotic distribution approximation, and (2) the tests may be more powerful
than the conventional VR tests against a wide range of models displaying serial
correlation, including fractionally integrated alternatives. The tests based on ranks
are exact under the i.i.d. assumption, whereas the tests based on signs are exact
even under conditional heteroscedasticity. Moreover, Wright (2000) showed that
ranks-based tests display low-size distortions, under conditional heteroscedasticity.
Given a variable in first differences {x
t
}
T
t =1
, let r(x) be the rank of x
t
among
(x
1
, . . . , x
T
). Under the null hypothesis that x
t
is generated from an i.i.d. sequence,
r(x) is a random permutation of the numbers of 1, . . . , T with equal probability.
Wright (2000) suggested the R
1
and R
2
statistics, defined as
R
1
(k) =
_
(Tk)
1

T
t =k
(r
1,t
+ +r
1,t k+1
)
2
T
1

T
t =k
r
2
1,t
1
_
(k)
1/2
(10)
R
2
(k) =
_
(Tk)
1

T
t =k
(r
2,t
+ +r
2,t k+1
)
2
T
1

T
t =k
r
2
2,t
1
_
(k)
1/2
(11)
where the standardized ranks r
1,t
and r
2,t
are given by
r
1,t
=
r(x
t
) (T +1)/2

(T 1)(T +1)/12
r
2,t
=
1
r(x)
T +1
where (k) is defined in (6), and
1
is the inverse of the standard normal
cumulative distribution function. The R
1
and R
2
statistics follow the same exact
sampling distribution. The critical values of these tests can be obtained by
simulating their exact distributions.
The tests based on the signs of first differences are given by
S
1
(k) =
_
(Tk)
1

T
t =k
(s
t
+ +s
t k+1
)
2
T
1

T
t =k
s
2
t
1
_
(k)
1/2
(12)
S
2
(k) =
_
(Tk)
1

T
t =k
(s
t
( ) + +s
t k+1
( ))
2
T
1

T
t =k
s
t
( )
2
1
_
(k)
1/2
(13)
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
510 CHARLES AND DARN

E
where (k) is defined in (6), s
t
= 2u(x
t
, 0), s
t
( ) = 2u(x
t
, ) and
u(x
t
, q) =
_
0.5 if x
t
> q
0.5 otherwise
Similar to R
1
and R
2
tests, the critical values of the S
1
and S
2
tests can be
obtained by simulating their exact sampling distribution.
Note that S
1
assumes a zero drift value. In recent work, Luger (2003) used ranks
and signs to extend the Campbell and Dufour (1997) non-parametric approach to
test for random walk with unknown drift. Belaire-Franch and Contreras (2004) and
Belaire-Franch and Opong (2005) applied the procedure described in Luger (2003)
to compute S
2
.
2.4 Correction of Size Distortions
As pointed out by Wright (2000), using several k values in the Wright tests would
lead to an over rejection of the null hypothesis, as in the Lo and MacKinlays
tests context (see Belaire-Franch and Opong, 2005). To overcome these test-size
distortions, Belaire-Franch and Contreras (2004) and Belaire-Franch and Opong
(2005) proposed different approaches to control the size of Wrights tests. They
considered the application of individual VR tests as the application of different
individual tests. Then, they applied different p-value adjustments for multiplicity,
in line with Psaradakis (2000).
They computed the Sidack-adjusted p-value for each test j as
p
(S)
ij
= 1 (1 p
ij
)
m
i = 1, . . . , m
where p
ij
is the p-value corresponding to the VR test j computed for an
individual k value, and m is the number of k values.
They also employed the Hochberg (1988) adjusted p-values that are obtained
as
p
(H)
ij
= min{[k R( p
ij
) +1] p
ij
, 1}
Given a significance level , the decision rule states that, using the VR test j, the
null is rejected if p
(S)
j
= min
1i m
p
(S)
ij
or p
(H)
j
= min
1i m
p
(H)
ij
.
However, these methods assume that the test statistics computed at different
intervals are uncorrelated. In order to take into account possible correlations among
the statistics, Belaire-Franch and Contreras (2004) and Belaire-Franch and Opong
(2005) suggested to compute bootstrap-adjusted p-values as described in Psaradakis
(2000). The goal of the procedure is to obtain an approximation to the null
sampling distribution of min
1i m
p
ij
by resampling with replacement from the
original returns.
12
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 511
2.5 Choi (1999) Test
When implementing the VR tests, the choice of holding period k is important.
However, this choice is usually rather arbitrary and ad hoc. To overcome this issue,
Choi (1999) proposed a data-dependent procedure to determine the optimal value
of k. Choi (1999) suggested a VR test based on frequency domain since Cochrane
(1988) showed that the estimator of V(k), which uses the usual consistent estimators
of variance, is asymptotically equivalent to 2 times the normalized spectral density
estimator at the zero frequency, which uses the Bartlett kernel. However, Choi
(1999) employed instead the quadratic spectral (QS) kernel because this kernel is
optimal in estimating the spectral density at the zero frequency (Andrews, 1991).
The VR estimator is defined as
VR(k) = 1 +2
T1

i =1
h(i /k) (i ) (14)
where (i ) is the autocorrelation function, and h(x) is the QS window defined as
h(x) =
25
12
2
x
2
_
sin(6x/5)
6x/5
cos(6x/5)
_
The standardized statistic is
VR
f
=
VR(k) 1
(2)
1/2
(T/k)
1/2
(15)
Under the null hypothesis the test statistic VR
f
follows the standard normal
distribution asymptotically.
13
Note that it is assumed that T , k and
T/k .
Various methods for optimally selecting the truncation point for the spectral
density at the zero frequency are available (Andrews, 1991; Andrews and Monahan,
1992; Newey and West, 1994; among others). Choi (1999) employed the Andrews
(1991) methods to select the truncation point optimally and compute the VR test.
Note that the small sample properties of this automatic VR test under
heteroscedasticity are unknown and have not been investigated properly.
3. Multiple Variance Ratio Tests
The LoMacKinlay test is an individual test where the null hypothesis is tested for
an individual value of k. The question as to whether or not a time series is mean
reverting requires that the null hypothesis holds true for all values of k. In view of
this, it is necessary to conduct a joint test where a multiple comparison of VRs over
a set of different time horizons is made. However, conducting separate individual
tests for a number of k values may be misleading as it leads to over rejection of the
null hypothesis of a joint test, above the nominal size. As stressed by Chow and
Denning (1993), this sequential procedure leads to an oversized testing strategy.
Thus, the weakness of LoMacKinlays test is that it ignores the joint nature of
testing for the RWH.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
512 CHARLES AND DARN

E
We present some multiple VR tests based on multiple comparison tests (Chow
and Denning, 1993; Whang and Kim, 2003; Belaire-Franch and Contreras, 2004)
and Wald-type tests (Richardson and Smith, 1991; Cecchetti and Lam, 1993;
Chen and Deo, 2006) that combine the information contained in statistics at several
horizons. These multiple VR tests consider the following joint null hypothesis H
0
:
V(k
i
) = 1 for all i = 1, . . . , m, against the alternative H
1
: V(k
i
) = 1 for some k
i
.
3.1 Chow and Denning (1993) Tests
Chow and Denning (1993) proposed using Hochbergs (1974) procedure for the
multiple comparison of the set of VR estimates with unity, which allows us to
examine a vector of individual VR tests while controlling for overall test size. For
a set of m test statistics, the RWH is rejected if any one of the estimated VRs is
significantly different from one.
To test the joint null hypothesis, Chow-Dennings (1993) test statistic is defined
as
MV
1
=

T max
1i m
|M
1
(k
i
)| (16)
where M
1
(k
i
) is defined in (5). This is based on the idea that the decision regarding
the null hypothesis can be obtained from the maximum absolute value of the
individual VR statistics. In order to control the size of the multiple VR test and
because the limit distribution of these statistics is complex, they applied the Sidak
(1967) probability inequality and give an upper bound to the critical values taken in
the studentized maximum modulus (SMM) distribution. Indeed, the statistic follows
the SMM distribution with m and T degrees of freedom, i.e. SMM(, m, T), where
m is the number of k values.
14
The null hypothesis is rejected at the level of
significance if the MV
1
statistic is greater than the [1 (

/2)]th percentile of the

standard normal distribution where

= 1 (1 )
1/m
.
Similarly, the heteroscedasticity-robust version of the ChowDenning test MV
2
can be written as
MV
2
=

T max
1i m
|M
2
(k
i
)| (17)
where M
2
(k
i
) is defined in (7), and it has the same critical values as MV
1
. However,
with finite-sample sizes it may be preferable to use critical values obtained by
simulations as done by Chow and Denning themselves. Nevertheless, as pointed
out by Fong et al. (1997), Hochbergs approach is valid only if the vector of test
statistics is multivariate normal. This condition is satisfied by VRs if there is little
overlap in the data, i.e. if k/T is small.
3.2 Whang and Kim (2003) Test
Whang and Kim (2003) developed a multiple VR test that uses a subsampling
technique of Politis et al. (1997), which is a data-intensive method of approximating
the sampling distribution. It can show better properties than the conventional VR
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 513
tests when the sample size is relatively small. The Monte Carlo experiment results
reported in Whang and Kim (2003) confirmed that their new VR test shows
excellent power in small samples, coupled with little or no serious size distortions.
To test the joint null hypothesis, Whang and Kim (2003) considered the statistic
MV
T
=

T g
N
(x
1
, . . . , x
T
) (18)
where
g
t
(x
1
, . . . , x
T
) = max
1i m
|M
r
(k
i
)|
with M
r
(k
i
) = VR(x; k
i
) 1, and VR(x; k) is as defined in (2). The sampling
distribution function for the MV
T
statistic is written as
G
T
(x) = P
_

T g
t
(x
1
, . . . , x
T
) x
_
Whang and Kim (2003) showed that the asymptotic null distribution of the
statistic is that of a maximum of a multivariate normal vector with unknown
covariance matrix, which is complicated to estimate. Therefore, they proposed to
approximate the null distribution by means of the subsampling approach.
Consider a subsample (x
t
, . . . , x
t b+1
) of size b for t = 1, . . . , T b + 1.
The statistic MV
T
calculated from the subsample is denoted as g
T,b,t
= g
b
(x
t
, . . . ,
x
t b+1
). Then, G
T
(x) is approximated by the distribution function obtained by the
collection of g
T,b,t
s calculated from all individual subsamples. It can be written as

G
T,b
(x) = (T b +2)
1
Tb+1

t =0
l
_

bg
T,b,t
x
_
where l() is the indicator function that takes 1 if the condition inside the parentheses
is satisfied and 0 otherwise.
The 100(1 )% critical value for the test can be calculated as the (1 )th
percentile of

G
T,b
, while the p-value of the test is estimated as 1

G
T,b
(MV
T
).
The null hypothesis that V(k
i
) = 1 (i = 1, . . . , m) is rejected at the level of
significance if the observed MV
T
is greater than this critical value or if the
p-value is less than . To implement the subsampling technique, a choice of block
length b should be made. Whang and Kim (2003) recommended that a number of
block lengths from an equally spaced grid in the interval of (2.5T
0.3
, 3.5T
0.6
) be
taken. However, they found that the size and power properties of their test are not
sensitive to the choice of block length.
3.3 Belaire-Franch and Contreras (2004) Tests
Recently, Belaire-Franch and Contreras (2004) proposed to substitute the standard
VR tests by Wrights rank- and sign-based tests in the definition of the Chow and
Denning (1993) procedure to create multiple rank and sign VR tests. The statistics
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
514 CHARLES AND DARN

E
are defined as
CD
(R
1
)
= max
1i m
|R
1
(k
i
)|
CD
(R
2
)
= max
1i m
|R
2
(k
i
)|
CD
(S
1
)
= max
1i m
|S
1
(k
i
)|
CD
(S
2
)
= max
1i m
|S
2
(k
i
)|
The ranks-based procedures are exact under the i.i.d. assumption whereas the
signs-based procedures are exact under both the i.i.d. and martingale difference
sequence assumption. Belaire-Franch and Contreras (2004) showed that the
ranks-based tests CD
(R
1
)
and CD
(R
2
)
are more powerful than their signs-based
counterparts, CD
(S
1
)
and CD
(S
2
)
.
15
Moreover, Colletaz (2005) and Kim and Shamsuddin (2008) also proposed an
extension to Wrights VR methodology following ChowDenning, but only for the
rank (R
1
) and sign (S
1
) tests, respectively.
3.4 Wald-type Tests
3.4.1 Richardson and Smith (1991) Test
Richardson and Smith (1991) suggested a joint test based on the following Wald
statistic:
RS(k) = T(VR 1
k
)

1
(VR 1
k
)
where VR is the (k 1) vector of sample k VRs, 1
k
is the (k 1) unit vector and
is the covariance matrix of VR. The joint RS(k) statistic follows a
2
distribution
with k degrees of freedom.
The usefulness of this test relies on the fact that, whenever the VR tests are
computed over long lags with overlapping observations, the distribution of the
VR test is non-normal; then, neither the LoMacKinlay test nor ChowDenning
procedure is valid for drawing inferences.
Moreover, Fong et al. (1997) argued that Richardson and Smiths (1991) joint
VR test can be more powerful than the ChowDenning multiple comparison test for
empirically relevant alternatives, and it displays low-size distortion in the presence
of heteroscedastic increments. However, their simulation results are based on an
autoregressive conditional heteroscedasticity (ARCH) process with slope coefficient
0.1, which is practically an i.i.d. process. Therefore, the conclusion of Fong et al.
(1997) could not hold under heteroscedasticity.
3.4.2 Cecchetti and Lam (1994) Test
Cecchetti and Lam (1994) proposed a multivariate version of the VR statistic to test
the RWH, in order to control the investment horizon. They suggested the following
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 515
Wald statistic that incorporates the correlations between VR statistics at various
horizons and weights them according to their variances:
S(k) = [VR(k) E [VR(k)]]

1
(k) [VR(k) E [VR(k)]]
where E is the expectation operator, VR a column vector sequence of VR statistics
VR(k) = [VR(2), . . . , VR(q)] and (k) is a measure of the covariance matrix of
VR. The joint VR S(k) statistic follows a
2
distribution with k degrees of freedom.
Cecchetti and Lam (1994) studied the empirical distribution of S(k) using Monte
Carlo techniques. For each simulation, they computed a value for the statistic S(k),
using the mean vector and covariance matrix

VR(k) and

(k) taken as the true
population values, and tabulate the distribution.
16
However, as suggested by Cecchetti and Lam (1994), the empirical distributions
of the VR have large positive skewness, suggesting that inference based on the
2
distribution will be misleading.
3.4.3 Chen and Deo (2006) Test
Chen and Deo (2006) also proposed a joint VR test based on their individual power
transformed VR statistic. They define the following Wald statistic:
QP(k) = (V
p,
(k)

(V
p,
(k)

)
where V
p,
is a column vector sequence of VR statistics V
p,
(k) = [VR

p
(2), . . . ,
VR

p
(k)] with VR

p
(k) the power transformed VR as in (9), and

and (k)

are a
measure of the expectation and covariance matrix of V
p,
, respectively. The joint
VR QP(k) statistic follows a
2
distribution with k degrees of freedom. Moreover,
Chen and Deo (2006) showed from Monte Carlo simulations that their joint VR
test displayed much higher power.
Note that the ChenDeo (2006) test is a joint test with one-sided alternative (H
1
:
V(k
i
) < 1, for some k
i
) while the RichardsonSmith (1991) and CecchettiLam
(1994) tests are joint tests with two-sided alternative (H
1
: V(k
i
) = 1). Therefore,
at level of significance, the null hypothesis that V(k
i
) = 1 is rejected if the
test statistic is greater than the upper 2 critical value of a
2
distribution for the
ChenDeo (2006) test and than the upper critical value of a
2
distribution for
the RichardsonSmith (1991) and CecchettiLam (1994) tests.
4. Bootstrapping Variance Ratio Tests
As already noted, Wright (2000), based on ranks and signs, and Whang and Kim
(2003), using the subsampling method, proposed VR tests that do not rely on
asymptotic approximations in order to overcome the difficulties due to using VR
tests based on asymptotic approximations (severe bias and right skewness). As
an alternative, some researchers proposed to employ a bootstrap method, which
is a resampling method that approximates the sampling distribution of a test
statistic (Efron, 1979) to the VR test statistic. The bootstrap is a distribution-free
randomization technique, which can be used to estimate the sampling distribution
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
516 CHARLES AND DARN

E
of the VR statistic when the distribution of the original population is unknown. We
describe the two most used bootstrapping VR tests, i.e. those suggested by Kim
(2006) in a theoretical framework and by Malliaropulos and Priestley (1999) in an
empirical framework.
17
4.1 Kim (2006) Test
Kim (2006) used the wild bootstrap that is a resampling method that approximates
the sampling distribution of the VR statistic, and is applicable to data with unknown
forms of conditional and unconditional heteroscedasticity (see Mammen, 1993;
Davidson and Flachaire, 2001).
Kim (2006) applied the wild bootstrap to LoMacKinlay, M
2
(k), and Chow
Denning, MV
2
(k
i
), VR tests. The wild bootstrap test based on MV
2
(k
i
) can be
conducted in three stages as below.
(1) Form a bootstrap sample of T observations X

t
=
t
X
t
(t = 1, . . . , T) where

t
is a random sequence with E() = 0 and E(
2
) = 1.
(2) Calculate MV

= MV
2
(X

; k
i
), the MV
2
(X

; k
i
) statistic obtained from the
bootstrap sample generated in stage (1).
(3) Repeat (1) and (2) sufficiently many, say m, times to form a bootstrap
distribution of the test statistic {MV
2
(X

; k
j
; j )}
m
j =1
.
The p-value of the test can be obtained as the proportion of {MV
2
(X

; k
j
; j )}
m
j =1
greater than the sample value of MV
2
(k
i
). The wild bootstrap version of the M
2
(k)
test can be implemented in a similar manner as a two-tailed test, where we obtain
M

= M
2
(X

; k) in stage (2) and {M

2
(X

; k
j
; j )}
m
j =1
in stage (3).
Conditionally on X
t
, X

t
is a serially uncorrelated sequence with zero mean
and variance X
2
t
. As such, M

and MV

have the same asymptotic distributions as

M
2
(k) and MV
2
(k
i
), respectively. Since X

t
is a serially uncorrelated sequence, wild
bootstrapping approximates the sampling distributions under the null hypothesis,
which is a desirable property for a bootstrap test.
To implement the wild bootstrap test, a specific form of
t
should be chosen.
Kim (2006) recommends using the standard normal distribution for
t
since he
reports that other choices provided qualitatively similar small sample results.
Note that the wild bootstrap is valid and the test statistics being bootstrapped are
pivotal asymptotically under the condition that X
t
follows a martingale difference
sequence.
18
Kim (2006) showed that the subsampling test of Whang and Kim (2003) displays
small sample properties far inferior to the wild bootstrap test under a small sample
size.
4.2 Malliaropulos and Priestley (1999) Test
Malliaropulos and Priestley (1999) used a weighted bootstrap method proposed by
Wu (1986) which is robust to the presence of heteroscedasticity, which is done by
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 517
resampling normalized returns instead of actual returns. Basically, the returns are
normalized by multiplying each observation of actual returns, for each one of the
time series of returns, by a corresponding random factor and resampling from these
normalized returns.
19
The bootstrap scheme can be summarized with the following
algorithm.
(l) For each t, draw a weighting factor z

t
(t = 1, . . . , T) with replacement from
the empirical distribution of normalized returns z
t
= (r
t
r)/(r), where
r = T
1

T
t =1
r
t
is the mean and (r) =
_
T
1

T
t =1
(r
t
r)
2
is the standard
error of the return.
(2) Form the bootstrap sample of T observations r

t
= z

t
r
t
(t = 1, . . . , T) by
multiplying each observation of actual returns with its corresponding random
weighting factor.
(3) Calculate the VR statistic VR

(k) from the pseudo data r

t
for k = 1, . . . , K.
(4) Repeat steps (1) and (2) M times, obtaining VR

(k; m) (m = 1, . . . , M)
and calculate the relevant quantiles, mean, median and standard deviation
of the sampling distribution of VR

(k) under the null hypothesis of serially

uncorrelated returns.
Using this procedure, resampling from normalized returns instead from actual
returns, the weighted bootstrap method accounts for the possible non-constancy of
the variance of returns. The strongest difficulty with resampling schemes, such
as bootstrap, is that they may generate data that are less dependent than the
original data. The main idea of the weighted bootstrap scheme is to overcome
this difficulty.
20
Malliaropulos and Priestley (1999) and Cajueiro and Tabak (2006) used this
bootstrap method to approximate the sampling distribution of the LoMacKinlay
VR statistics as well as the Wald statistic of the Cecchetti and Lam (1994)
test.
Note that Malliaropulos and Priestley (1999) bootstrapped the VR statistics,
which are not asymptotically pivotal, under heteroscedasticity in an empirical
framework. Their bootstrap tests are not supported by any asymptotic theory or
Monte Carlo evidence to evaluate their properties in contrast to the bootstrap tests
proposed by Kim (2006).
5. Empirical Applications
The VR tests have been widely used and their applications have often covered
emerging markets: Asian markets (Hoque et al., 2007; Kim and Shamsuddin,
2008), Eastern European markets (Smith and Ryoo, 2003), African markets (Smith
et al., 2002; Al-Khazali et al., 2007; Lagoarde-Segot and Lucey, 2008) and Latin
American markets (Chaudhuri and Wu, 2003; Chang et al., 2004). In this section we
propose an illustration by examining the RWH for five emerging markets in Latin
America, including Argentina, Brazil, Chile, Ecuador and Mexico. We use daily
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
518 CHARLES AND DARN

E
Table 1. Summary Statistics of Stock Returns.
Mean SD Skewness Kurtosis LM(10)
Argentina 0.000468 0.022095 0.133875
a
8.158586
a
402.0338
a
Brazil 0.001866 0.025078 0.550350
a
13.58173
a
365.7680
a
Chile 0.000448 0.007172 0.046965 8.218660
a
555.2467
a
Ecuador 0.000025 0.018418 1.125265
a
44.79396
a
225.6864
a
Mexico 0.000792 0.016019 0.001551 9.235503
a
380.2590
a
a
Means significant at 1% level, respectively. LM(10) indicates the Lagrange multiplier test for
conditional heteroscedasticity with 10 lags.
market prices spanning 03 August 1993 to 22 May 2007. All data are collected
from Thomson Financial Datastream.
There have been many studies that tested the efficiency of Latin American stock
markets. However, the results are overall mixed and scattered over studies that
employ different sample periods, methods and data frequencies. Urrutia (1995),
using the LoMacKinlay VR test, rejected the RWH for the Latin American
emerging equity markets of Argentina, Brazil, Chile and Mexico, whereas the runs
test indicated weak-form efficiency over the period 1980:31988:12. In contrast,
Ojah and Karemera (1999) found that the Latin American equity returns follow
a random walk and were generally weak-form efficient. Grieb and Reyes (1999)
re-examined the random walk properties of stocks traded in Brazil and Mexico
over the period 1988:121995:6, using the LoMacKinlay VR tests, and concluded
that the Mexican stock market exhibited mean aversion whereas the Brazilian
stock market showed a tendency toward random walk. Karemera et al. (1999)
also found that Brazil, Chile and Mexico did not follow the random walk under
the LoMacKinlay test, whereas Argentina did, over the period 1987:121997:5.
However, this result changed when they applied ChowDennings multiple VR test,
showing that Argentina and Brazil followed a random walk. Chaudhuri and Wu
(2003) investigated the efficiency for Argentina, Brazil, Chile, Colombia, Mexico
and Venezuela over the period 1985:11997:2. Using the LoMacKinlay VR test,
they rejected the RWH only for Argentina and Brazil. Chang et al. (2004) rejected
the RWH, using a Wald-type test (Cecchetti and Lam, 1994), for Argentina, Brazil,
Chile and Mexico over the period 1991:12004:1.
Table 1 presents summary statistics for the stock returns calculated as the first
differences in the logs of the stock price indices. The data are all leptokurtic as
might be expected from daily stock returns. Three series (Argentina, Brazil and
Ecuador) are skewed. To check for non-linear dependences, we apply the Lagrange
multiplier test for ARCH on the residuals of the ARMA model, where the lag
length is selected based on the Akaike and Schwarz information criterion. This
particular specification of heteroscedasticity was motivated by the observation that,
in many financial time series, the magnitude of residuals appeared to be related
to the magnitude of recent residuals. The LM(10) indicates clearly that all stocks
show strong conditional heteroscedasticity.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 519
Table 2. Results of Individual VR Tests.
VR tests k Argentina Brazil Chile Ecuador Mexico
M
2
(k) 2 3.00
a
2.22
a
8.74
a
1.41 3.60
a
5 1.63 1.61 9.87
a
1.99
a
2.05
a
10 1.13 1.13 10.11
a
1.17 1.56
30 1.45 3.78
a
10.66
a
0.10 1.44
R
1
(k) 2 3.34
a
1.50 16.96
a
1.33 7.31
a
5 2.45
a
1.28 17.29
a
5.11
a
4.64
a
10 1.91
a
2.24
a
16.91
a
8.61
a
3.53
a
30 2.93
a
0.56 17.13
a
14.55
a
4.42
a
R
2
(k) 2 3.10
a
1.61 15.87
a
1.10 7.04
a
5 2.23
a
1.14 16.31
a
4.14
a
4.16
a
10 1.61 2.30
a
15.73
a
7.73
a
3.02
a
30 2.17
a
0.31 15.34
a
13.85
a
3.57
a
S
2
(k) 2 2.09
a
0.90 12.79
a
7.34
a
5.95
a
5 0.96 0.01 12.79
a
14.98
a
3.58
a
10 0.13 0.01 12.92
a
21.45
a
2.37
a
30 1.15 0.41 13.15
a
34.78
a
2.68
a
VR

p
(k) 2 1.63
b
0.71 0.71 0.31 2.83
a
5 1.45
b
0.25 0.54 0.29 1.28
10 0.95 1.27 1.46
b
1.43
b
0.63
30 1.06 0.88 0.12 3.31
a
0.60
We report the VR statistic for each test.
a
Significant at the 5% level.
b
Significant at the 10% level.
Tables 2 and 3 report the results of various individual and multiple VR tests,
21
respectively, for the five Latin American markets. Since these stock returns exhibit
conditional heteroscedasticity, we do not consider the LoMacKinlay M
1
(k) and
ChowDenning MV
1
tests. More precisely, for individual VR tests we apply the
LoMacKinlay M
2
(k) test as well as Wrights R
1
, R
2
and S
2
tests. For multiple
VR tests, we apply the ChowDenning MV
2
test, the RichardsonSmith RS test,
the WhangKim subsampling MV
T
test, the BelaireContreras rank-based CD(R
1
)
and CD(R
2
) tests as well as Kims bootstrap MV

test.
22
The holding periods (k) considered are (2, 5, 10, 30). As advocated by Deo
and Richardson (2003), we use relatively short holding periods when testing for
mean reversion using VR tests. For the wild bootstrap test (MV

), as suggested
by Kim (2006),
23
the number of bootstrap replications m is set to 1000. As
recommended by Whang and Kim (2003), we take a number of block lengths
from an equally spaced grid in the interval (2.5T
0.3
, 3.5T
0.6
) for the subsampling
test (MV
T
).
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
520 CHARLES AND DARN

E
Table 3. Results of Multiple VR Tests.
VR tests Block length Argentina Brazil Chile Ecuador Mexico
MV
2
3.00
a
3.78
a
10.66
a
1.99 3.60
a
RS 29.94
a
83.91
a
380.74
a
46.94
a
43.87
a
QP(k) 3.81 6.17 9.63
a
20.16
a
9.40
a
MV

0.01
a
0.00
a
0.00
a
0.10 0.00
a
CD(R
1
) 3.69
a
4.85
a
18.14
a
13.65
a
7.84
a
CD(R
2
) 4.19
a
6.42
a
18.31
a
11.63
a
7.89
a
MV
T
93 0.00
a
0.00
a
0.00
a
0.01
a
0.00
a
157 0.00
a
0.00
a
0.00
a
0.05 0.00
a
221 0.00
a
0.00
a
0.00
a
0.26 0.00
a
285 0.00
a
0.00
a
0.00
a
0.32 0.00
a
349 0.00
a
0.00
a
0.00
a
0.32 0.00
a
413 0.00
a
0.00
a
0.00
a
0.28 0.00
a
a
Significant at the 5% level. The p-values are given for the MV

(Kim, 2006) and MV

T
(Belaire and
Contreras, 2004) tests whereas the VR statistic is reported for the other tests.
Some convergence among the individual and multiple VR tests is observed for
Chile and Mexico. Indeed, the individual and multiple VR tests reject the RWH
for these two Latin American markets, indicating that these markets have not been
weak-form efficient.
For Argentina, some divergence among the individual tests is observed. The
statistics of individual tests do not provide the same results. Nevertheless, as shown
by Wright (2000), the rank-based VR tests are more powerful than the conventional
LoMacKinlay and the sign-based VR tests. Thus, in this context, it seems that
Argentina does not follow a random walk. Furthermore, the RWH is generally
rejected under the multiple VR tests. Consequently, the exchange rate market of
Argentina seems to be inefficient.
Applying the individual VR tests shows that Brazil follows a random walk.
However, this result changes when the multiple VR tests are employed (except
for the ChenDeo QP(k) test), showing that, as found by Chang et al. (2004),
the Brazilian exchange rate market is not an efficient market. As already noted,
conducting individual tests for a number of k values may be misleading as it leads
to over rejection of the null hypothesis of a joint test, above the nominal size.
Finally, we found mixed results from the various VR tests for Ecuador. Indeed,
the LoMacKinlay M
2
(k) test is not significant, whereas the rank- and sign-
based tests as well as the ChenDeo VR

p
are significant. Furthermore, the Kim
bootstrap and the WhangKim subsampling tests do not reject the RWH while the
other multiple VR tests show that Ecuador follows a random walk. Consequently,
it is impossible to conclude on the weak-form efficiency for the Ecuadorian
market.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 521
6. Conclusion
This paper reviewed the recent developments in the field of the VR tests of the ran-
dom walk and martingale hypothesis. In particular, we presented the conventional
individual and multiple VR tests as well as their improved modifications based
on power-transformed statistics, rank and sign tests, subsampling and bootstrap
methods, among others.
We also re-examined the weak-form efficiency for five emerging equity markets
in Latin America. We found that Argentina, Brazil, Chile and Mexico follow
rejecting the RWH and, consequently, these four Latin American markets are not
weak-form efficient. We do not conclude for Ecuador because the results are mixed.
We did not deal with the possible presence of structural breaks, due to financial
or economic events, which can affect the VR tests. We left this issue to further
research that can be conducted even by applying VR tests using a moving subsample
window (Yilmaz, 2003; Kim and Shamsuddin, 2008) or by modifying VR tests to
take into account structural changes (Lee and Kim, 2006).
Acknowledgements
We would like to thank two anonymous referees for very helpful comments and
suggestions.
Notes
1. When the error terms are not an i.i.d. sequence, the random walk process is
denominated martingale process, whereas the sequence {
t
}
T
t=1
is the so-called
martingale difference sequence. Campbell et al. (1997) refer to the random walk
3.
2. Daniel (2001) explores a wider range of possible test statistics.
3. Recently, Kan (2006) provided analytical formulae for the moments of the sample
VR under both the null and the alternatives. See also Shively (2002) for the case of
the mean.
4. Hoque et al. (2007) proposed a comparison of several VR tests.
5. Lo and MacKinlay (1988) and Campbell et al. (1997) argued that using overlapping
data in estimating the variances allowed a more efficient estimator to be obtained
and hence a more powerful test.
6. It has been argued that misleading conclusions may be obtained with VR statistics
when time-varying volatility is present in the data. See, for example, Kim, M.J.,
et al. (1991), Kim, C.J., et al. (1998) and Kim and Nelson (1998) who also proposed
a solution based on a Bayesian approach and the use of a Gibbs sampler.
7. Kan (2006) presented the exact distributions of the VR test with overlapping data.
Moreover, Tse et al. (2004) suggested a modified VR statistic and proposed to
approximate the small-sample distribution of this statistic using a beta distribution
that matches the exact mean and the asymptotic variance.
8. Richardson and Stock (1989) showed that the VR statistic, without any normaliza-
tion, converges to a functional of Brownian motion.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
522 CHARLES AND DARN

E
9. Deo and Richardson (2003) advocated that large values of k should not be used
when testing for mean reversion using the VR statistics.
10. To adjust for conditional heteroscedasticity, Chen and Deo (2006) proposed a
modified version of the standard deviation of the transformed VR statistic (even
for the standard VR statistic).
11. See Chen and Deo (2004) for a discussion on power transformations.
12. Note that the p-value adjustments can be applied to other VR tests for a joint
hypothesis.
13. The estimator VR
f
based on the Bartlett window as in Cochrane (1988) also has
a limiting normal distribution. However, Cogley (1990) showed that this estimator
seems inappropriate because it is right skewed. Therefore, Cogley (1990) proposed
to approximate this estimator by a multiple of the
2
variate, giving improvement
over the normal.
14. The critical values of the test are tabulated in Hahn and Hendrickson (1971) and
Stoline and Ury (1979). It should be noted that, when T is large, the critical values
of the test can be calculated from the limiting distribution of the statistic.
15. Belaire-Franch and Contreras (2004) also suggested substituting Wrights rank- and
sign-based tests in the definition of the Richardson and Smith (1991) procedure but
found that these tests were inferior to rank- and sign-based CD
()
tests.
16. Note that Cecchetti and Lam (1994) showed that the empirical distribution of the
statistic S(k) is numerically identical to the quadratic sum of the deviations of
the first k 1 autocorrelations from their population values, weighted by their
covariance matrix.
17. Recently, Fleming et al. (2006) developed a bootstrap method for testing multiple
inequality restrictions on VRs.
18. See MacKinnon (2002) for the advantages of bootstrapping asymptotic pivotal
statistics. Note that there are other possible choices of two-point distributions for
the wild bootstrap, which potentially outperform the standard normal distribution
when the sample size is small (Davidson et al., 2007).
19. See Malliaropulos and Priestley (1999) and Cajueiro and Tabak (2006) for a detailed
discussion on the weighted bootstrap method proposed by Wu (1986).
20. Cribari-Neto and Zarkos (1999) evaluated the performance of this bootstrap method-
ology by comparing the weighted with the unweighted bootstrap. Their results
suggested that weighted bootstrap estimators perform very well, outperforming other
estimators, even in the case of heteroscedastic errors and non-normality (fat tails).
21. Some computational resources for VR tests are available from the different authors.
Kim (2006) provides various VR tests written in R (http://www.r-project.org/) avail-
able from http://www-personal.buseco.monash.edu.au/jaekim/vrtest.htm. Moreover,
a Stata module for the LoMacKinlay test written by Baum (2006) is available from
http://fmwww.bc.edu/repec/bocode/l/lomackinlay.ado.
22. We do not apply the S
1
test suggested by Wright (2001) since it assumes a zero
drift that need not be satisfied in practice as well as the multiple sign-based tests
developed by Belaire-Franch and Contreras (2004) since the rank-based tests are
more powerful.
23. Following Kim (2006), we use the standard normal distribution for
t
to implement
the wild bootstrap test. He reports that other choices provided qualitatively similar
sample results.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 523
References
Al-Khazali, O.M., Ding, D.K. and Pyun, C.S. (2007) A new variance ratio test of random
walk in emerging markets: a revisit. Financial Review 42: 303317.
Andrews, D.W.K. (1991) Heteroskedasticity and autocorrelation consistent covariance
matrix estimation. Econometrica 58: 817858.
Andrews, D.W.K. and Monahan, C. (1992) An improved heteroskedasticity and
autocorrelation consistent covariance matrix estimator. Econometrica 60: 953966.
Belaire-Franch, J. and Contreras, D. (2004) Ranks and signs-based multiple variance ratio
tests. Working Paper, Department of Economic Analysis, University of Valencia.
Belaire-Franch, J. and Opong, K.K. (2005) Some evidence of random walk behaviour
of Euro exchange rates using ranks and signs. Journal of Banking and Finance 29:
16311643.
Brockwell, P. and Davis, R. (1996) Time Series: Theory and Methods. New York: Springer.
Cajueiro, D.O. and Tabak, B.M. (2006) Testing for predictability in equity returns for
European transition markets. Economic System 30: 5678.
Campbell, B. and Dufour, J.-M. (1997) Exact nonparametric tests of orthogonality and
random walk in the presence of a drift parameter. International Economic Review 38:
151173.
Campbell, J.Y. and Mankiw, N.G. (1987) Permanent and transitory components in
macroeconomic fluctuations. American Economic Review 77: 111117.
Campbell, J.Y., Lo, A.W. and MacKinlay, A.C. (1997) The Econometrics of Financial
Markets. Princeton, NJ: Princeton University Press.
Cecchetti, S.G. and Lam, P.S. (1994) Variance ratio tests: small-sample properties with an
application to international output data. Journal of Business and Economic Statistics
12: 177186.
Chang, E.J., Lima, E.J.A. and Tabak, B.M. (2004) Testing predictability in emerging
equity markets. Emerging Markets Review 5: 295316.
Chaudhuri, K. and Wu, Y. (2003) Random walk versus breaking trend in stock prices:
evidence from emerging markets. Journal of Banking and Finance 27: 575592.
Chen, W.W. and Deo, R.S. (2004) Power transformations to induce normality and their
applications. Journal of the Royal Statistical Society, Series B 66: 117130.
Chen, W.W. and Deo, R.S. (2006) The variance ratio statistic at large horizons.
Econometric Theory 22: 206234.
Choi, I. (1999) Testing the random walk hypothesis for real exchange rates. Journal of
Applied Econometrics 14: 293308.
Chow, K.V. and Denning, K.C. (1993) A simple multiple variance ratio test. Journal of
Econometrics 58: 385401.
Cochrane, J.H. (1988) How big is the random walk in GNP? Journal of Political Economy
96: 893920.
Cogley, T. (1990) International evidence on the size of the random walk in output. Journal
of Political Economy 98: 501518.
Colletaz, G. (2005) A simple multiple variance-ratio test based on ranks. Working Paper,
LEO, University of Orl eans.
Cribati-Neto, F. and Zarkos, S. (1999) Bootstrap methods for heteroskedastic regression
models: evidence on estimation and testing. Econometric Review 18: 211228.
Daniel, K. (2001) The power and size of mean reversion tests. Journal of Empirical
Finance 8: 493535.
Davidson, R. and Flachaire, E. (2001) The wild bootstrap, tamed at last. Working Paper
99, GREQAM, University of Marseille.
Davidson, R., Monticini, A. and Peel, D. (2007) Implementing the wild bootstrap using
a two-point distribution. Economics Letters 96: 309315.
Deo, R.S. and Richardson, M. (2003) On the asymptotic power of the variance ratio test.
Econometric Theory 19: 231239.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
524 CHARLES AND DARN

E
Efron, B. (1979) Bootstrap methods: another look at the jackknife. Annals of Statistics 7:
126.
Fama, E. (1970) Efficient capital markets: a review of theory and empirical work. Journal
of Finance 25: 383417.
Fama, E. (1991) Efficient capital markets: II. Journal of Finance 46: 15751617.
Faust, J. (1992) When are variance ratio tests for serial dependence optimal? Econometrica
60: 12151226.
Fleming, J., Kirby, C. and Ostdiek, B. (2006) Bootstrap tests of multiple inequality
restrictions on variance ratios. Economics Letters 91: 343348.
Fong, W.M., Koh, S.K. and Ouliaris, S. (1997) Joint variance-ratio tests of the martingale
hypothesis for exchange rates. Journal of Business and Economic Statistics 15: 5159.
Grieb, T. and Reyes, M.G. (1999) Random walk tests for Latin American equity indices
and individual firms. Journal of Financial Research 4: 371383.
Hahn, G.J. and Hendrickson, R.W. (1971) A table of percentage points of the distribution
of the largest value of K Student T variates and its applications. Biometrika 58:
323332.
Hochberg, Y. (1974) Some conservative generalizations of the T-method in simultaneous
inference. Journal of Multivariate Analysis 4: 224234.
Hochberg, Y. (1988) A sharper Bonferroni procedure for multiple tests of significance.
Biometrika 75: 800802.
Hoque, H.A.A.B., Kim, J.H. and 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: 488502.
Kan, R. (2006) Exact variance ratio test with overlapping data. Working Paper, University
of Toronto.
Karemera, D., Ojah, K. and Cole, J.A. (1999) Random walks and market efficiency
tests: evidence from emerging equity markets. Review of Quantitative Finance and
Accounting 13: 171188.
Kim, C.J. and Nelson, C.R. (1998) Testing for mean-reversion in heteroskedastic data II:
Autocorrelation tests based on Gibbs-sampling-augmented randomization. Journal of
Empirical Finance 5: 385396.
Kim, C.J., Nelson, C.R. and Startz, R. (1998) Testing for mean reversion in heteroskedastic
data based on Gibbs-sampling-augmented randomization. Journal of Empirical
Finance 5: 131154.
Kim, J.H. (2006) Wild bootstrapping variance ratio tests. Economics Letters 92: 3843.
Kim, J.H. and Shamsuddin, A. (2008) Are Asian stock markets efficient? Evidence from
new multiple variance ratio tests. Journal of Empirical Finance 15: 518532.
Kim, M.J., Nelson, C.R. and Startz, R. (1991) Mean reversion in stock prices. A
reappraisal of the empirical evidence. Review of Economic Studies 58: 515528.
Lagoarde-Segot, T. and Lucey, B.M. (2007) Efficiency in emerging markets evidence
from the MENA region. International Financial Markets, Institutions and Money 18:
94105.
Lee, Y.-S. and Kim, T.-H. (2006) Spurious rejections by variance ratio tests in the presence
of a break in drift. Working Paper 0603, Durham University.
Lo, A.W. and MacKinlay, A.C. (1988) Stock market prices do not follow random walk:
evidence from a simple specification test. Review of Financial Studies 1: 4166.
Lo, A.W. and MacKinlay, A.C. (1989) The size and power of the variance ratio test in
finite samples: a Monte Carlo investigation. Journal of Econometrics 40: 203238.
Luger, R. (2003) Exact non-parametric tests for a random walk with unknown drift under
conditional heteroscedasticity. Journal of Econometrics 115: 259276.
MacKinnon, J.G. (2002) Bootstrap inference in econometrics. Canadian Journal of
Economics 35: 615645.
Malliaropulos, D. and Priestley, R. (1999) Mean reversion in Southeast Asian stock
markets. Journal of Empirical Finance 6(4): 355384.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 525
Mammen, E. (1993) Bootstrap and wild bootstrap for high dimensional linear models.
Annals of Statistics 21: 255285.
Newey, W. and West, K. (1994) Automatic lag selection in covariance matrix estimation.
Review of Economic Studies 61: 631653.
Ojah, K. and Karemera, D. (1999) Random walks and market efficiency tests of Latin
American emerging equity markets: a revisit. Financial Review 34: 5772.
Perron, P. and Vodounou, C. (2005) The variance ratio test: an analysis of size and power
based on a continuous-time asymptotic framework. Econometric Theory 21: 562592.
Politis, D.N., Romano, J.P. and Wolf, M. (1997) Subsampling for heteroskedastic time
series. Journal of Econometrics 81: 281317.
Poterba, J.M. and Summers, L.H. (1988) Mean reversion in stock prices: evidence and
implications. Journal of Financial Economics 22: 2759.
Psaradakis, Z. (2000) p-value adjustments for multiple tests for nonlinearity. Studies in
Nonlinear Dynamics and Econometrics 4: 95100.
Richardson, M. and Smith, T. (1991) Tests of financial models in the presence of
overlapping observations. Review of Financial Studies 4: 227254.
Richardson, M. and Stock, J. (1989) Drawing inferences from statistics based on multiyear
asset returns. Journal of Financial Economics 25: 323348.
Shively, P.A. (2002) An exact invariant variance ratio test. Economics Letters 75: 347353.
Sidak, Z. (1967) Rectangular confidence regions for the means of multivariate normal
distributions. Journal of the American Statistical Association 62: 626633.
Smith, G. and Ryoo, H.-J. (2003) Variance ratio tests of the random walk hypothesis for
European emerging stock markets. European Journal of Finance 9: 290300.
Smith, G., Jefferis, K. and Ryoo, H. (2002) African stock markets: multiple variance ratio
tests of random walks. Applied Financial Economics 12: 475484.
Stoline, M.R. and Ury, H.K. (1979) Tables of the studentized maximum modulus
distribution and an application to multiple comparisons among means. Technometrics
21: 8793.
Tse, Y.K., Ng, K.W. and Zhang, X.B. (2004) A small-sample overlapping variance ratio
test. Journal of Time Series Analysis 25: 127135.
Urrutia, J.L. (1995) Tests of random walk and market efficiency for Latin American
emerging markets. Journal of Financial Research 18: 299309.
Whang, Y.-J. and Kim, J. (2003) A multiple variance ratio test using subsampling.
Economics Letters 79: 225230.
Wright, J.H. (2000) Alternative variance-ratio tests using ranks and signs. Journal of
Business and Economic Statistics 18: 19.
Wu, C.F.J. (1986) Jackknife, bootstrap and other resampling methods in regression
analysis. Annals of Statistics 14: 12611295.
Yilmaz, K. (2003) Martingale property of exchange rates and central bank intervention.
Journal of Business and Economic Statistics 21: 383395.
Appendix: Assumptions on the VR Tests
We present some of the main assumptions for the underlying time series that drive
the VR tests.
The LoMacKinlay (1988), ChowDenning (1993) and Kim (2006) tests are
driven by Assumption H

of Lo and Mackinlay (1988), which assumes the

following.
H

1: For all t , E(
t
) = 0, and E(
t

t
) = 0 for any = 0
H

2: {
t
} is -mixing with coefficients (m) of size r/(2r 1) or is -mixing
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
526 CHARLES AND DARN

E
with coefficients (m) of size r/(r 1), where r > 1, such that for all t
and for any 0, there exists some > 0 for which
E|
t

t
|
2(r+)
< <
H

3: lim
T
T
1

T
t =1
E(
2
t
) =
2
0
<
H

4: For all t , E(
t

t j

t i
) = 0 for any non-zero j and i, where j = i
Assumption H

1 is the essential property of the random walk. Assumptions H

2
and H

3 are restrictions on the degree of dependence and heterogeneity, which are

allowed and yet still permit some form of law of large numbers and central limit
theorem to obtain. This allows for a variety of forms of heteroscedasticity, including
GARCH-type variances and variances with deterministic changes. Assumption H

4
implies that the sample autocorrelations of
t
are asymptotically uncorrelated.
The Whang and Kim (2003) test is driven by a relaxed version of Assumption
H

. Indeed, they do not impose the restriction that the sample autocorrelations of

t
are asymptotically uncorrelated (Assumption H

4) by assuming
lim
T
T
1
T

t =1
E
_

2
t
_
=
2
0
Their test is then robust to violations of this assumption.
The Wright (2000) tests are driven by Assumption A in which it is considered
that x
t
= + z
t
and z
t
=
t

t
. Letting I
t
= {x
t
, x
t 1
, x
t 2
, . . .}, the assumptions
are
A1: z
t
is i.i.d.
A2:
t
and
t
are independent, conditional on I
t 1
A3: E(
t
|I
t 1
) = 0 and 1(
t
> 0) is an i.i.d. binomial variable that is 1 with
probability
1
2
and 0 otherwise.
Assumption A1 stipulates that the first-differences are i.i.d. while the combination
of Assumptions A2 and A3 is sufficient, but not necessary, for x
t
to be a martingale
difference sequence. Moreover, Assumption A2 is satisfied in a GARCH model
and also by a stochastic volatility model in which the innovations to volatility are
independent of
t
. Assumption A3 allows to
t
to be t-distributed with time-varying
degrees of freedom. The rank-based tests of Wright (2000) and Belaire-Franch and
Contreras (2004) are based on Assumption A1 while their sign-based tests are
based on Assumptions A2 and A3.
Finally, the Chen and Deo (2006) assumptions on the martingale difference
sequence are also different.
B1: {
t
} is ergodic and E(
t
|F
t
) = 0 for all t, where F
t
is a sigma field,
t
is
F
t
measurable, and F
t 1
F
t
for all t.
B2: E(
2
t
) =
2
< .
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd
VARIANCE-RATIO TESTS OF RANDOM WALK 527
B3: For any integer q, 2 q 8, and for q non-negative integers
s
i
, E(

q
i =1

s
i
t
i
) = 0 when at least one s
i
is exactly one and

q
i =1
s
i
8.
B4: For any integer r, 2 r 4, and for r non-negative integers
s
i
, E(

r
i =1

s
i
t
i
|F
t
) = 0 when at least one s
i
is exactly one and

r
i =1
s
i
4,
for all t < t
i
, i = 1, 2, 3, 4.
B5: lim
T
var[E(
2
t +T

2
t +T+j
|F
t
)] = 0 uniformly in j for every j > 0.
B6: lim
T
E(
2
t

2
t T
) =
4
.
Assumptions B1B6 allow the innovations
t
to be a martingale difference
sequence with conditional heteroscedasticity. Chen and Deo (2006) showed that the
stochastic volatility and GARCH models satisfy Assumptions B1B6. Assumptions
B3 and B4 state that the series {
t
} shows product moment behaviour similar to
that of an independent white noise process. Assumptions B5 and B6 state that
t
and
t T
are roughly independent for large lags T.
Journal of Economic Surveys (2009) Vol. 23, No. 3, pp. 503527
C
2009 The Authors. Journal compilation
C
2009 Blackwell Publishing Ltd