Time Series Diagnostic Test
Time Series Diagnostic Test
Time Series Diagnostic Test
test of Lobato, Nankervis, and Savin (2001), the spectral tests of Durlauf (1991), and
the robust spectral test of Deo (2000). There are also tests of the martingale dierence
hypothesis, including Bierens (1982, 1984), Bierens and Ploberger (1997), Hong (1999),
Dominguez and Lobato (2000), Whang (2000, 2001), Kuan and Lee (2004), and Park
and Whang (2005). For the hypothesis of serial independence, two leading tests are the
variance ratio test of Cochrane (1988) and the so-called BDS test of Brock, Dechert, and
Scheinkman (1987); see also Campbell, Lo, and MacKinlay (1997) and Brock, Dechert,
Scheinkman, and LeBaron (1996). Skaug and Tjostheim (1993), Pinkse (1998), and
Hong (1999) also proposed non-parametric tests of serial independence. In addition
to these tests, there are tests of time reversibility, which may also be interpreted as
tests of independence; see e.g., Ramsey and Rothman (1996) and Chen, Chou, and
Kuan (2000). It has been shown that a test of time reversibility is particularly powerful
against asymmetric dependence.
In this note we introduce various diagnostic tests for time series. We will not discuss
non-parametric tests because they are not asymptotically pivotal, in the sense that their
asymptotic distributions are data dependent. This note proceeds as follows. Section 2.1
focuses on the tests of serial uncorrelatedness. In Section 3, we discuss the tests of the
martingale dierence hypothesis. Section 4 presents the variance ratio test and the BDS
test of serial independence. The tests of time reversibility are discussed in Section 5.
c _ Chung-Ming Kuan, 2005
2
2 Tests of Serial Uncorrelatedness
Given a weakly stationary time series y
t
, let denote its mean and () denote its
autocovariance function, where (i) = cov(y
t
y
ti
) for i = 0, 1, 2, . . .. The autocorrelation
function () is such that (i) = (i)/(0). The series y
t
is serially uncorrelated if, and
only if, its autocorrelation function is identically zero.
2.1 Q Tests
As testing all autocorrelations is infeasible in practice, existing tests of serial uncorrelat-
edness focus on a given number of autocorrelations and ignore (i) for large i. The null
hypothesis is
H
0
: (1) = = (m) = 0,
where m is a pre-specied number. Let y
T
denote the sample mean of y
t
,
T
(i) the i
th
sample autocovariance:
T
(i) =
1
T
Ti
t=1
(y
t
y)(y
t+i
y),
and
T
(i) = (i)/ (0) the i
th
sample autocorrelation. For notation convenience, we shall
suppress the subscript T and simply write y, (i) and (i). Writing
m
=
_
(1), . . . , (m)
_
,
the null hypothesis is
m
= o, and the estimator of
m
is
m
=
_
(1), . . . , (m)
_
. Under
quite general conditions, it can be shown that as T tends to innity,
T
_
m
m
_
D
^(o, V ),
where
D
stands for convergence in distribution, and the (i, j)
th
element of V is
v
ij
=
1
(0)
2
_
c
i+1,j+1
(i)c
1,j+1
(j)c
1,i+1
+ (i)(j)c
1,1
,
with
c
i+1,j+1
=
k=
IE
_
(y
t
)(y
t+i
)(y
t+k
)(y
t+k+j
)
IE
_
(y
t
)(y
t+i
)
IE
_
(y
t+k
)(y
t+k+j
)
;
c _ Chung-Ming Kuan, 2005
2.1 Q Tests 3
see e.g., Lobato et al. (2001). Fuller (1976, p. 256) presents for a dierent expression of
V .
1
It follows from the continuous mapping theorem that
T
_
m
m
_
V
1
_
m
m
_
D
2
(m). (2.1)
This distribution result is fundamental for the tests presented in this section.
Under the null hypothesis, V can be simplied such that v
ij
= c
i+1,j+1
/(0)
2
with
c
i+1,j+1
=
k=
IE
_
(y
t
)(y
t+i
)(y
t+k
)(y
t+k+j
)
.
In particular, when y
t
are serially independent,
c
i+1,j+1
=
_
0, i ,= j,
(0)
2
, i = j.
In this case, V is an identity matrix, and the normalized sample autocorrelations
T (i),
i = 1, . . . , m, are independent ^(0, 1) random variables asymptotically. Many computer
programs now raw a condence interval for sample autocorrelations so as to provide a
quick, visual check of the signicance of (i). For example, the 90% (95%) condence
interval of (i) is 1.645/
T (1.96/
m
m
= T
m
i=1
(i)
2
D
2
(m), (2.2)
under the null hypothesis; Q
T
is the well known Q test of Box and Pierce (1970).
When y
t
are independent random variables with mean zero, variance
2
, and nite
6
th
moment, we have from a result of Fuller (1976, p. 242) that
cov
_
T (i),
T (j)
_
=
_
Ti
T
+ O(T
1
), i = j ,= 0,
O(T
1
), i ,= j.
This result provides an approximation up to O(T
1
). Then for suciently large T, the
diagonal elements of V are approximately (T i)/T, whereas the o-diagonal elements
essentially vanish. This leads to the modied Q test of Ljung and Box (1978):
Q
T
= T
2
m
i=1
(i)
2
T i
D
2
(m), (2.3)
1
Fuller (1976, p. 256) shows that the (i, j)
th
element of V is
vij =
k=
(k)(k i + j)+(k + j)(k i) 2(k)(j)(k i)
2(k)(i)(k j) + 2(i)(j)(k)
2
.
c _ Chung-Ming Kuan, 2005
2.1 Q Tests 4
cf. (2.2). Clearly, the Box-Pierce Q test and the Ljung-Box
Q test are asymptotically
equivalent, yet the latter ought to have better nite-sample performance because it em-
ploys a nite-sample correction of V . In practice, the Ljung-Box Q statistic is usually
computed as
T(T + 2)
m
i=1
(i)
2
(T i)
, (2.4)
which is of course asymptotically equivalent to (2.3).
Another modication of the Q test can be obtained by assuming that
IE
_
(y
t
)(y
t+i
)(y
t+k
)(y
t+k+j
)
= 0, (2.5)
for each k when i ,= j and for k ,= 0 when i = j. Given this assumption, c
i+1,j+1
= 0
when i ,= j, but
c
i+1,j+1
= IE
_
(y
t
)
2
(y
t+i
)
2
.
when i = j. Under this assumption, V is diagonal with the diagonal element v
ii
=
c
i+1,i+1
/(0)
2
, which can be consistently estimated by
v
ii
=
1
T
Ti
t=1
(y
t
y)
2
(y
t+i
y)
2
[
1
T
T
t=1
(y
t
y)
2
]
2
.
Then under the null hypothesis,
Q
T
= T
m
i=1
(i)
2
/ v
ii
D
2
(m); (2.6)
Lobato et al. (2001) refer to this test as the Q
test is therefore more suitable for testing processes that are serially uncorrelated
but serially dependent, such as GARCH processes (see Example in Section 3.1.
Remark:
1. The asymptotic distribution of the Box-Pierce and Ljung-Box Q tests is derived
under the assumption that y
t
is serially independent. This distribution result
is also valid when y
t
is a martingale dierence sequence (a precise denition
will be given in Section 3.1) with additional moment conditions. These Q tests
can also be interpreted as independence (or martingale-dierence) tests with a
focus on autocorrelations. Although the Q
j=
(j)e
ij
, [, ],
where i = (1)
1/2
. When the autocorrelations are all zero, the spectral density reduces
to the constant (2)
1
for all . It is then natural to base a test of all autocorrelations
by comparing the sample counterpart of f() and (2)
1
.
Let I
T
() denote the periodogram, the sample spectral density, of the time series y
t
.
The dierence between I
T
() and (2)
1
is
1
2
T1
j=(T1)
(j)e
ij
1
.
Recall that exp(ij) = cos(j) i sin(j), where sin is an odd function such that
sin(j) = sin(j), and cos is an even function such that cos(j) = cos(j). Thus,
1
2
T1
j=(T1)
(j)e
ij
1
=
1
T1
j=1
(j) cos(j).
Integrating this function with respect to on [0, a], 0 a , we obtain the cumulated
dierences:
1
T1
j=1
(j)
sin(ja)
j
.
Consider now the normalized, cumulated dierences:
D
T
(t) =
2T
m(T)
j=1
(j)
sin(jt)
j
,
c _ Chung-Ming Kuan, 2005
2.2 The Spectral Tests 6
where t = a and m(T) is less than T but grows with T at a slower rate.
Recall that a sequence of i.i.d. ^(0, 1) random variables
t
can approximate the
Brownian motion B via
W
T
(t) =
0
t +
j=1
j
sin(jt)
j
B(t), t [0, 1]
where stands for weak convergence. Then,
W
T
(t) tW
T
(1) =
j=1
j
sin(jt)
j
B
0
(t), t [0, 1], (2.7)
where B
0
denotes the Brownian bridge. It can also be shown that when y
t
are condi-
tionally homoeskedastic,
T (j), j = 1, . . . , m, converge in distribution to independent
^(0, 1) random variables under the null hypothesis. In view of the approximation (2.7)
and the asymptotic normality of T
1/2
(j), we obtain
D
T
(t) B
0
(t), t [0, 1]. (2.8)
The spectral tests proposed by Durlauf (1991) are constructed from taking various func-
tionals on D
T
.
Durlauf (1991) considered the following test statistics:
(1) the Anderson-Darling test:
AD
T
=
_
1
0
[D
T
(t)]
2
t(1 t)
dt
_
1
0
[B
0
(t)]
2
t(1 t)
dt;
(2) the Cramer-von Mises test:
CVM
T
=
_
1
0
[D
T
(t)]
2
dt
_
1
0
[B
0
(t)]
2
dt;
(3) the Kolmogorov-Smirnov test:
KS
T
= sup[D
T
(t)[ sup [B
0
(t)[;
(4) the Kuiper test:
Ku
T
= sup
s,t
[D
T
(t) D
T
(s)[ sup[B
0
(t) B
0
(s)[.
The limits of these test statistics are direct consequences of (2.8) and the continuous
mapping theorem. These limits are also the limits of the well-known goodness-of-t tests
in the statistics literature, and their critical values have been tabulated in, e.g., Shorack
and Wellner (1986).
c _ Chung-Ming Kuan, 2005
7
A condition ensuring the result (2.8) is that
T (j) are independent ^(0, 1) asymp-
totically. Similar to the nding of Lobato et al. (2001), Deo (2000) noted that when y
t
are
conditionally heteroskedastic, the asymptotic variance of
2T
m(T)
j=1
(j)
v
jj
sin(jt)
j
,
where
v
jj
=
1
(0)
_
1
T j
Tj
t=1
(y
t
y)
2
(y
t+j
y)
2
_
1/2
,
and (0) is the sample variance of y
t
. With additional regularity conditions, the Cramer-
von Mises test based on D
c
T
has the same limit as the original CVM
T
test, i.e.,
CVM
c
T
=
_
1
0
[D
c
T
(t)]
2
dt
_
1
0
[B
0
(t)]
2
dt.
This modication is analogous to the Q
t2
(
t2
+
t
+ 1),
where
t
are i.i.d. N(0, 1) random variables. It is clear that corr(y
t
, y
tj
) = 0 for all j,
yet y
t
is not a martingale dierence process. Recall that a white noise is a sequence
of uncorrelated random variables that have zero mean and a constant variance. Thus,
a martingale dierence sequence need not be a white noise because the former does not
have any restriction on variance or other high-order moments.
If y
t
is a sequence of serially independent random variables with zero mean, we
have
IE[y
t
h(Y
t1
)] = IE(y
t
) IE[h(Y
t1
)] = 0,
so that y
t
is necessarily a martingale dierence sequence. The converse need not be
true. To see this, consider a simple ARCH (autoregressive conditional heteroskedasticity)
process y
t
such that y
t
= v
1/2
t
t
, where
t
are i.i.d. random variables with mean zero
and variance 1, and
v
t
= a + by
2
t1
,
c _ Chung-Ming Kuan, 2005
9
with a, b > 0. It can be seen that y
t
is a martingale dierence sequence, yet it is
serially dependent due to the correlaions among y
2
t
. We therefore conclude that serial
independence implies the martingale dierence property which in turn implies serial
uncorrelatedness.
4 Tests of Serial Independence
We have seen that serial independence is a much more stringent requirement than the
martingale dierence property. The ARCH example in Section 3.1, for instance, is a mar-
tingale dierence sequence but serially dependent. Given that the ARCH and GARCH
(generalized ARCH) models are popular in nancial applications, it is also of interest to
consider testing a special form of serial dependence, viz., the correlations among squared
returns. McLeod and Li (1983) suggested testing whether the rst m autocorrelations
of y
2
t
are zero using a Q test. That is, one computes (2.3) or (2.4) with (i) the sample
autocorrelations of y
2
t
: where
(i) =
1
T
Ti
t=1
(y
2
t
m
2
)(y
2
t+i
m
2
)
1
T
T
t=1
(y
2
t
m
2
)
2
,
with m
2
the sample mean of y
2
t
. The asymptotic null distributions of the resulting Q
test remains
2
(m). While the McLeod-Li test also checks a necessary condition of
serial independence, its validity requires an even stronger moment condition (nite 8
th
moment). The tests discussed below, on the other hand, focus on the i.i.d. condition which
is sucient for serial independence. Failing to reject the null hypothesis is consistent with
serial independence, yet rejecting the null hypothesis does not imply serial dependence.
4.1 The Variance Ratio Test
The variance-ratio test of Cochrane (1988) is a convenient diagnostic test of the i.i.d.
assumption. Suppose that y
t
are i.i.d. random variables with mean zero and variance
2
. Then for any k, var(y
t
+ + y
tk+1
) is simply k
2
. Let
2
k
denote an estimator
of var(y
t
+ + y
tk+1
) and
2
the sample variance of
2
. Under the null hypothesis,
2
k
/k and
2
should be close to each other. The variance ratio test is simply a normalized
version of
2
k
/(k
2
).
Let
t
be the partial sum of y
i
such that y
t
=
t
t1
. Suppose there are kT + 1
observations
0
,
1
, . . . ,
kT
. Dene the sample average of y
t
as
y =
1
kT
kT
t=1
(
t
t1
) =
1
kT
(
kT
0
).
c _ Chung-Ming Kuan, 2005
4.1 The Variance Ratio Test 10
The variance estimator of
2
= var(
t
t1
) is
2
=
1
kT
kT
t=1
_
t1
y
_
2
.
Under the i.i.d. null hypothesis,
kT(
2
2
)
D
^(0, 2
4
). Consider now the following
estimator of
2
k
= var(
t
tk
):
2
k
=
1
T
T
t=1
_
kt
ktk
k y
_
2
=
1
T
T
t=1
_
k( y
t
y)
2
,
where y
t
=
kt
ktk+1
y
i
/k. Under the i.i.d. hypothesis,
2
k
= k
2
, so that
T(
2
k
k
2
)
D
^(0, 2k
2
4
),
or equivalently,
1
T(
2
k
k
2
) =
kT
_
2
k
k
2
_
=
kT
_
2
k
k
2
_
+
kT
_
2
2
_
D
^(0, 2k
4
).
While
2
is both consistent and asymptotically ecient for
2
under the null hypothe-
sis,
2
k
/k is consistent but not asymptotically ecient. Following Hausman (1978),
2
we
conclude that
2
must be asymptotically uncorrelated with
2
k
/k
2
and hence
kT
_
2
k
k
2
_
D
^
_
0, 2(k 1)
4
_
.
It follows that
kT
_
2
k
k
2
1
_
D
^
_
0, 2(k 1)
_
.
Denote the ratio
2
k
/(k
2
) as VR. The statistic
kT(VR 1)/
_
2(k 1)
D
^(0, 1),
under the null hypothesis.
2
Given the parameter of interest , let
e be a consistent and asymptotically ecient estimator of
and
c is a consistent estimator but not asymptotically ecient. Hausman (1978) showed that
e is
asymptotically uncorrelated with
c
e. For if not, there would exist a linear combination of
e and
2
k
may be estimated by
2
k
=
1
kT
kT
t=k
_
tk
k y
_
2
.
One may also correct the bias of variance estimators and compute
2
=
1
kT 1
kT
t=1
_
t1
y
_
2
,
2
k
=
1
M
kT
t=k
_
tk
k y
_
2
,
where M = (kT k + 1)(1 1/T). For more detailed discussion, we refer to Campbell,
Lo, and MacKinlay (1997).
4.2 The BDS Test
The BDS test of serial independence also checks whether a sequence of random vari-
ables are i.i.d. Let Y
n
t
= (y
t
, y
t+1
, . . . , y
t+n1
). Dene the correlation integral with the
dimension n and distance as:
C(n, ) = lim
T
_
T n
2
_
1
s<t
I
(Y
n
t
, Y
n
s
),
where I
(Y
n
t
, Y
n
s
) = 1 if the maximal norm |Y
n
t
Y
n
s
| < and 0 otherwise. The
correlation integral is a measure of the proportion that any pairs of n-vectors (Y
n
t
and
Y
n
s
) are within a certain distance . If y
t
are indeed i.i.d., Y
n
t
should exhibit no pattern
in the n-dimensional space, so that C(n, ) = C(1, )
n
. The BDS test is then designed
to check whether the sample counterparts of C(n, ) and C(1, )
n
are suciently close.
Specically, the BDS statistic reads
B
T
(n, ) =
T n + 1(C
T
(n, ) C
T
(1, )
n
)/ (n, ),
where
C
T
(n, ) =
_
T n
2
_
1
s<t
I
(Y
n
t
, Y
n
s
),
and
2
(n, ) is a consistent estimator of the asymptotic variance of
T n + 1C
T
(n, );
see Brock et al. (1996) for details. The asymptotic null distribution of the BDS test is
^(0, 1).
c _ Chung-Ming Kuan, 2005
12
The performance of the BDS test depends on the choice of n and . There is, however,
no criterion to determine these two parameters. In practice, one may consider several
values of n and set as a proportion to the sample standard deviation s
T
of the data, i.e.,
= s
T
for some . Common choices of are 0.75, 1, and 1.5. An advantage of the BDS
test is that it is robust to random variables that do not possess high-order moments. The
BDS test usually needs a large sample to ensure proper performance. Moreover, it has
been found that the BDS test has low power against various forms of nonlinearity; see,
e.g., Hsieh (1989, 1991, 1993), Rothman (1992), Brooks and Heravi (1999), and Brooks
and Henry (2000). In particular, the BDS test is not sensitive to certain class of self-
exciting threshold AR processes (Rothman, 1992) and neglected asymmetry in volatility
(Hsieh, 1991; Brooks and Henry, 2000; Chen and Kuan, 2002).
5 Tests of Time Reversibility
A dierent type of diagnostic test focuses on the property of time reversibility. A strictly
stationary process y
t
is said to be time reversible if its nite-dimensional distributions
are all invariant to the reversal of time indices. That is,
F
t
1
,t
2
,...,tn
(c
1
, c
2
, . . . , c
n
) = F
tn,t
n1
,...,t
1
(c
n
, c
n1
, . . . , c
n
).
When this condition does not hold, y
t
is said to be time irreversible. Clearly, indepen-
dent sequences and stationary Gaussian ARMA processes are time reversible. Rejecting
the null hypothesis of time reversibility thus implies that the data can not be serially
independent. As such, the test of time reversibility can also be interpreted as a test of
serial independence.
Time irreversibility indicates some time series characteristics that can not be de-
scribed by the autocorrelation function. When y
t
is time reversible, it can be shown
that for any k, the marginal distribution of y
t
y
tk
must be symmetric about the the
origin; see e.g., Cox (1981) and Chen, Chou, and Kuan (2000). That is, y
t
y
tk
and
y
tk
y
t
should have the same distributions for each k. If this symmetry condition fails
for some k, there is some asymmetric dependence between y
t
and y
tk
, in the sense that
the eect of y
tk
on y
t
is dierent from that of y
t
on y
tk
. In view of this property,
we infer that nonlinear time series are time irreversible in general. Indeed, Tong (1990)
states that time irreversibility is the rule rather than the exception when it comes to
nonlinearity (p. 197). Moreover, linear and stationary processes with non-Gaussian in-
novations are typically time irreversible. Thus, testing time reversibility can complement
existing tests of independence.
c _ Chung-Ming Kuan, 2005
13
Existing tests of time reversibility aim at symmetry of y
t
y
tk
. A necessary condition
of distribution symmetry is its third central moment being zero. One may then test time
reversibility by evaluating whether the sample third moment is suciently close to zero.
Observe that by stationarity,
IE(y
t
y
tk
)
3
= IE(y
3
t
) 3 IE(y
2
t
y
tk
) + 3 IE(y
t
y
2
tk
) IE(y
3
tk
)
= 3 IE(y
2
t
y
tk
) + 3 IE(y
t
y
2
tk
),
where the two terms on the right-hand side are referred to as the bi-covariances. Ramsey
and Rothman (1996) base their test of time reversibility on the sample bi-covariances.
Note that both the third-moment test and bi-covariance test require the data to possess
at least nite 6
th
moment. Unfortunately, most nancial time series do not satisfy this
moment condition. On the other hand, Chen, Chou and Kuan (2000) consider a dierent
testing approach that is robust to the failure of moment conditions.
It is well known that a distribution is symmetric if, and only if, the imaginary part
of its characteristic function is zero. Hence, time reversibility of y
t
implies that
h
k
() := IE
_
sin
_
(y
t
y
tk
)
_
= 0, for all R
+
. (5.1)
Note that (5.1) include innitely moment conditions indexed by . Let g be a positive
function such that
_
g() d < . By changing the orders of integration, (5.1) implies
that
_
R
+
h
k
()g() d =
_
R
__
R
+
sin
_
(y
t
y
tk
)
_
g() d
_
dF = 0,
where F is the distribution function of y
t
. This condition is equivalent to
IE[
g
(y
t
y
tk
)] = 0, (5.2)
where
g
(y
t
y
tk
) =
_
R
+
sin
_
(y
t
y
tk
)
_
g() d.
To test (5.2), Chen, Chou, and Kuan (2000) suggest using the sample average of
g
(y
t
y
tk
):
(
g,k
=
_
T
k
g,k
/
g,k
, (5.3)
where T
k
= T k with T the sample size,
g,k
=
T
t=k+1
g
(y
t
y
tk
)/T
k
, and
2
g,k
is
a consistent estimator of the asymptotic variance of
_
T
k
g,k
. A suitable central limit
theorem then ensures that (
g,k
is asymptotically distributed as ^(0, 1) under the null
hypothesis (5.2).
c _ Chung-Ming Kuan, 2005
14
A novel feature of this test is that, because
g
() is bounded between 1 and 1, no
moment condition is needed when the central limit theorem is invoked. Yet a major
drawback of (
g,k
is that the null hypothesis (5.2) is only a necessary condition of (5.1).
Thus, h
k
may be integrated to zero by some g function even when h
k
is not identically
zero. For such a g function, the resulting (
g,k
test does not have power against asymmetry
of y
t
y
tk
. Choosing a proper g function is therefore crucial for implementing this test.
Chen, Chou, and Kuan (2000) observed that for absolutely continuous distributions,
h
k
() is a damped sine wave and eventually decays to zero as ; see, e.g., Figure ??
for h
k
of various centered exponential distributions. This suggests choosing g as the
density function of a random variable on R
+
. In particular, when g is the density of the
exponential distribution (g = exp) with the parameter > 0, i.e., g() = exp(/)/,
it can be shown that
g
has an analytic expression:
exp
(y
t
y
tk
) =
(y
t
y
tk
)
1 +
2
(y
t
y
tk
)
2
. (5.4)
The closed form (5.4) renders the computation of (
exp,k
test quite easy. One simply plugs
the data into (5.4) and calculates their sample average and sample standard deviation.
The test statistic is now readily computed as (5.3). Chen, Chou, and Kuan (2000)
demonstrate that the (
exp,k
test performs strikingly well in nite samples and is very
robust to the data without higher-order moments. The third-moment-based test and the
bi-covariance test, on the other hand, have little power when the data do not have proper
moments.
Apart from computation simplicity, the (
exp,k
test is very exible. By varying the
value of , (
exp,k
is able to check departures from (5.2) in dierent ways. When is
small, this test concentrates on h
k
() for smaller . By contrast, more h
k
() can be
taken into account as increases. How to choose an optimal that maximizes the test
power remains an unsolved problem, however.
Remarks:
1. One may consider testing a condition equivalent to (5.1). For example, a Cramer-
von Mises type condition is based on
_
R
+
h
k
()
2
g() d,
which is zero if, and only if, (5.1) holds. This condition, however, does not permit
changing the orders of integration. The resulting test is more dicult to implement
and usually has a data-dependent distribution.
c _ Chung-Ming Kuan, 2005
15
2. To apply the test (5.3) to model residuals, Chen and Kuan (2002) noted that one
must compute the sample standard deviation with care. Chen and Kuan (2002)
also demonstrated that the (
exp,k
test is particularly powerful against asymmetric
dependence in data, but other diagnostic tests are not. Specically, they simulated
data from an EGARCH process which exhibits volatility asymmetry but estimated
a GARCH model which is a model of symmetric volatility pattern. It is shown that
the (
exp,k
test on the standardized GARCH residuals can reject the null hypothesis
of time reversibility with high probability, yet the Q-type tests and the BDS test
have low power and fail to reject the null hypothesis of serial uncorrelatedness
(independence). As far as model diagnostic is concerned, the latter tests are unable
to distinguish between GARCH and EGARCH models.
c _ Chung-Ming Kuan, 2005
16
References
Bierens, Herman J. (1982). Consistent model specication tests, Journal of Economet-
rics, 20, 105134.
Bierens, Herman J. (1984). Model specication testing of time series regressions, Journal
of Econometrics, 26, 323353.
Bierens, Herman J. and Werner Ploberger (1997). Asymptotic theory of integrated con-
ditional moment tests, Econometrica, 65, 11291151.
Box, G. E. P. and David A. Pierce (1970). Distribution of residual autocorrelations in
autoregressive-integrated moving average time series models, Journal of the Ameri-
can Statistical Association, 65, 15091526.
Brock, W. A., W. D. Dechert and J. A. Scheinkman (1987). A test for independence based
on the correlation dimension, Working paper, Department of Economics, University
of Wisconsin at Madison.
Brock, W. A., W. D. Dechert, J. A. Scheinkman, and B. LeBaron (1996). A test for
independence based on the correlation dimension, Econometric Reviews, 15, 197
235.
Brooks, Chris and Saeed M. Heravi (1999). The eect of (mis-specied) GARCH lters
on the nite sample distribution of the BDS test, Computational Economics, 13,
147162.
Brooks, Chris and
Oland T. Henry (2000). Can portmanteau nonlinearity tests serve as
general mis-specication tests? Evidence from symmetric and asymmetric GARCH
models, Economics Letters, 67, 245251.
Campbell, John Y., Andrew W. Lo, and A. Craig MacKinlay (1997). The Econometrics
of Financial Markets, Princeton, NJ: Princeton University Press.
Chen, Yi-Ting, Ray Chou, and Chung-Ming Kuan (2000). Testing time reversibility
without moment restrictions, Journal of Econometrics, 95, 199218.
Chen, Yi-Ting and Chung-Ming Kuan (2002). Time irreversibility and EGARCH eects
in U.S. stock index returns, Journal of Applied Econometrics, 17, 565578.
Cochrane, John H. (1988). How big is the random walk in GNP? Journal of Political
Economy, 96, 893920.
Cox, D. R. (1981). Statistical analysis of time series: Some recent developments, Scan-
dinavian Journal of Statistics, 8, 93115.
c _ Chung-Ming Kuan, 2005
17
de Lima, Pedro J. F. (1997). On the robustness of nonlinearity tests to moment condition
failure, Journal of Econometrics, 76, 251280.
Deo, Rohit S. (2000). Spectral tests of the martingale hypothesis under conditional
heteroscedasticity, Journal of Econometrics, 99, 291315.
Dominguez, Manuel A. and Ignacio N. Lobato (2000). A consistent test for the martingale
dierence hypothesis, Working Paper, Instituto Tecnologico Autonomo de Mexico.
Durlauf, Steven N. (1991). Spectral based testing of the martingale hypothesis, Journal
of Econometrics, 50, 355376.
Fuller, Wayne A. (1976). Introduction to Statistical Time Series, New York, NY: Wiley.
Hausman, Jerry A. (1978). Specication tests in econometrics, Econometrica, 46, 1251
1271.
Hong, Yongmiao (1999). Hypothesis testing in time series via the empirical character-
istic function: A generalized spectral density approach, Journal of the American
Statistical Association, 94, 12011220.
Hsieh, David A. (1989). Testing for nonlinear dependence in daily foreign exchange rates,
Journal of Business, 62, 339368.
Hsieh, David A. (1991). Chaos and nonlinear dynamics: Application to nancial markets,
Journal of Finance, 46, 18391877.
Hsieh, David A. (1993). Implications of nonlinear dynamics for nancial risk manage-
ment, Journal of Financial and Quantitative Analysis, 28, 4164.
Ljung, G. M. and G. E. P. Box (1978). On a measure of lack of t in time series models,
Biometrika, 65, 297303.
Lobato, Ignacio, John C. Nankervis, and N. E. Savin (2001). Testing for autocorrelation
using a modied Box-Pierce Q test, International Economic Review, 42, 187205.
McLeod, A. I. and W. K. Li (1983). Diagnostic checking ARMA time series models using
squared-residual autocorrelations, Journal of Time Series Analysis, 4, 269273.
Pinkse, Joris (1998). A consistent nonparametric test for serial independence, Journal
of Econometrics, 84, 205231.
Ramsey, James B. and Philip Rothman (1996). Time irreversibility and business cycle
asymmetry, Journal of Money, Credit, and Banking, 28, 320.
Rothman, Philip (1992). The comparative power of the tr test against simple threshold
c _ Chung-Ming Kuan, 2005
18
models, Journal of Applied Econometrics, 7, S187S195.
Shorack, Galen R. and John A. Wellner (1986). Empirical Processes with Applications
to Statistics, New York, NY: Wiley.
Skaug, Hans J. and Dag Tjostheim (1993). A nonparametric test of serial independence
based on the empirical distribution function, Biometrika, 80, 591602.
Tong, Howell (1990). Non-linear Time Series, A Dynamic System Approach, New York,
NY: Oxford University Press.
Whang, Yoon-Jae (2000). Consistent bootstrap tests of parametric regression functions,
Journal of Econometrics, 98, 2746.
Whang, Yoon-Jae (2001). Consistent specication testing for conditional moment re-
strictions, Economics Letters, 71, 299306.
c _ Chung-Ming Kuan, 2005