Biometrika (2008), 95, 2, pp. 399414 doi: 10.

1093/biomet/asn014
C 2008 Biometrika Trust
Printed in Great Britain
Least absolute deviation estimation for fractionally integrated
autoregressive moving average time series models with
conditional heteroscedasticity
BY GUODONG LI AND WAI KEUNG LI
Department of Statistics and Actuarial Science,
The University of Hong Kong, Pokfulam Road, Hong Kong
ligd@hku.hk hrntlwk@hku.hk
SUMMARY
We consider a unied least absolute deviation estimator for stationary and nonstationary
fractionally integrated autoregressive moving average models with conditional heteroscedasticity.
Its asymptotic normality is established when the second moments of errors and innovations are
nite. Several other alternative estimators are also discussed and are shown to be less efcient
and less robust than the proposed approach. A diagnostic tool, consisting of two portmanteau
tests, is designed to check whether or not the estimated models are adequate. The simulation
experiments give further support to our model and the results for the absolute returns of the Dow
Jones Industrial Average Index daily closing price demonstrate their usefulness in modelling
time series exhibiting the features of long memory, conditional heteroscedasticity and heavy
tails.
Some key words: ARFIMA; Conditional heteroscedasticity; Heavy tail; GARCH; Least absolute deviation; Long
memory.
1. INTRODUCTION
The fractionally integrated autoregressive moving average, ARFIMA or fractional ARIMA, process
was proposed by McLeod & Hipel (1978), Granger & Joyeux (1980) and Hosking (1981), and is
one of the most popular models for explaining the phenomenon of long memory in diverse elds
of statistical application, especially in the eld of nance; see Robinson (2003). On the other
hand, since Engle (1982), it has been accepted that many nancial time series have a time-varying
conditional variance. In fact, some nancial time series may exhibit the features of both long
memory and time-varying conditional variance; these include the consumer price index ination
series in Baillie et al. (1996), the Swiss Euromarket interest rate in Hauser & Hunst (2001) and
the absolute log return sequences in Tsay (2002). The generalized autoregressive conditional
heteroscedasticity, GARCH, models (Bollerslev, 1986) are usually considered for modelling the
phenomenon of time-varying conditional variance and it is natural to consider the ARFIMAGARCH
model dened as follows:
(B)(1 B)
d
Y
t
= (B)
t
, (1)

t
= u
t
h
1/2
t
, h
t
=
0
+
r

i =1

2
t i
+
s

j =1

j
h
t j
, (2)
400 GUODONG LI AND WAI KEUNG LI
where
0
> 0,
i
0 (i = 1, . . . , r)
j
0 ( j = 1, . . . , s) B is the backward shift opera-
tor, (B) = 1

p
i =1

i
B
i
, (B) = 1 +

q
j =1

j
B
j
and (1 B)
d
=

k=0
[(k d)/{(k +
1)(d)}]B
k
. Note that {
t
} is the error sequence and members of the innovation sequence {u
t
}
are independently and identically distributed.
The parameter d describes the extent of long memory generated by the ARFIMAGARCH model
and is called the memory parameter. The process generated by models (1) and (2) is short
memory, long memory, stationary or nonstationary, respectively, when d (05, 0), (0, ),
(05, 05) or (05, ); see Ling & Li (1997). It is important to estimate d as well as the other
parameters. When the errors {
t
} are independent, many procedures, including time domain
and frequency domain methods, have been developed for model (1); see Bhardwaj & Swanson
(2006) and references therein. Under the normality of u
t
, Ling & Li (1997) established the
asymptotic normality of the maximum likelihood estimators. Beran & Feng (2001) considered
local polynomial estimation of semiparameteric models with an ARFIMAGARCH error. However,
both papers required the condition E(
4
t
) < resulting in a more restricted parameter space for
model (2); see Ling (2007). Francq & Zakoian (2004) discussed the asymptotic normality of the
Gaussian quasi-maximum likelihood estimators of the autoregressive moving average models
with GARCH errors, and niteness of E(
4
t
) was needed although niteness of only a smaller-order
moment was required for pure GARCH models. The condition E(
4
t
) < seems unavoidable for
such estimation approaches and it is necessary to develop a different method for ARFIMAGARCH
models with a less restricted parameter space.
Furthermore, recent empirical evidence has increasingly shown that some nancial time series
may be so heavy-tailed that the fourth moment of the innovation u
t
is innite; see Mittnik et al.
(1998) and Mikosch & Starica (2000). Most existing estimation methods for fractional ARIMA and
ARFIMAGARCH models are sensitive to outliers. Haldrup & Nielsen (2007) showed, by simulation
experiments, that some commonly used estimators of the fractional ARIMA models were not robust
to outliers and the estimators of memory parameters may be biased. Granger et al. (1999) applied
the fractional ARIMA models to several price indices, and Ling & Li (1997) tted a ARFIMA
GARCH model to the Hong Kong Heng Seng Index. In order to obtain a good estimator, outliers
were removed before estimation in both papers. However, it is well known that these outliers
may include useful information; see Embrechts et al. (1997). For pure GARCH processes with
E(u
4
t
) = , Hall & Yao (2003) showed that Gaussian quasi-maximumlikelihood estimation may
not be asymptotically normal and the convergence rate is slower than the standard rate of n
1/2
. The
same may happen for models (1) and (2) with innite fourth moment for u
t
. Peng & Yao (2003)
constructed three least absolute deviation estimators for the pure GARCH models and established
their asymptotic normality under only nite second moments of
t
and u
t
. This approach may be
useful for providing robust estimation for ARFIMAGARCH models.
2. THE LEAST ABSOLUTE DEVIATION ESTIMATION
Let l = p +q +r +s +2 and denote the parameter vector of models (1) and (2) by =
(

, where = (d,
1
, . . . ,
p
,
1
, . . . ,
q
)

, = (
0
,
1
, . . . ,
r
,
1
, . . . ,
s
)

and is an
l-dimensional vector. Assume that the parameter space is a compact set of R
l
, the true parameter
vector
0
is an interior point of and each in satises the following two assumptions.
Assumption 1. We assume that
i
> 0, i = 0, 1, . . . , r,
j
> 0, j = 1, . . . , s, E(
2
t
) < , and
the polynomials

r
i =1

i
z
i
and 1

s
j =1

j
z
j
have no common root.
Time series models 401
Assumption 2. We assume that d > 05 and d J 05, where J is the collection of all
positive integers. All roots of the polynomials (z) = 1

p
i =1

i
z
i
and (z) = 1 +

q
j =1

j
z
j
lie outside the unit circle and there is no common factor between (z) and (z).
Denote Eu
2
t
by
2
. Then the condition
2

r
i =1

i
+

s
j =1

j
< 1 is necessary and sufcient
for {
t
} in (2) to exist as a unique strictly stationary sequence with a nite second moment Li & Li
(2005). The cases with d J 05 are very complicated (Beran, 1995; Ling & Li, 1997). Hence
we exclude them from the parameter space and leave them for possible future research.
Under Assumptions 1 and 2, model (1) is invertible and then
t
can be written as

t
= (B)
1
(B)

k=0
(k d)
(k +1)(d)
Y
t k
.
When the true parameter vector in the above equation is replaced by ,
t
can be considered as
a function on and hence can be denoted by
t
( ) or
t
(). Similarly, we can dene the function
h
t
() by the iterative equation
0
+

r
i =1

2
t i
() +

s
j =1

j
h
t j
(). These two functions
depend on the innite past values of the sequences {Y
t
} and {
t
}. They are unobservable in real
applications and some initial values are needed. For simplicity, we set the initial values of {Y
t
}
and {
t
} to zero and replace h
t
and
2
t
for t 0 by (1/n)

n
i =1

2
i
. This will not affect the results
in the following derivation; see Bollerslev (1986) and Ling & Li (1997). Furthermore, to save
space and forestall confusion, we denote
t
(
0
), h
t
(
0
),
t
(
0
)/ and h
t
(
0
)/ respectively
by
t
, h
t
,
t
/ and h
t
/.
There are two different approaches to dening the least absolute deviation estimators: one
is based on the sum of absolute errors and the other on Laplace quasi-maximum likelihood
estimation. These two methods are consistent with each other for linear models, but they
may be totally different for nonlinear models. For the difference between these two methods,
see Peng & Yao (2003) and Berkes & Horvath (2004) for the least absolute deviation esti-
mators of pure GARCH models. We rst consider the sum of absolute errors for models (1)
and (2). Peng & Yao (2003) designed three least absolute deviation estimators by rewriting the
GARCH model in the forms of regression. A natural extension of the best of these three to the
ARFIMAGARCH model is given by

PY
= arg min

t =1
|log
2
t
() log h
t
()|.
However, in order to derive its asymptotic normality, the condition E(1/|u
t
|) < is required
and this excludes many familiar continuous distributions such as the normal distribution and
Students t -distributions. The other two estimators considered require a nite fourth moment
condition. Hence, this paper does not pursue this direction and focuses only on the method of
Laplace quasi-maximum likelihood estimation. By temporarily assuming a Laplace distribution
with density g(x) = 05 e
|x|
for the innovation u
t
, we can dene the least absolute deviation
estimator as

n
= arg min

L
n
(), L
n
() =
n

t =1
_
|
t
()|

{h
t
()}
+
1
2
log h
t
()
_
.
To investigate the asymptotic distribution of

n
, we need another assumption.
Assumption 3. The median of u
t
is equal to zero, E|u
t
| = 1, E(u
2
t
) =
2
< and the prob-
ability density function f (x) of u
t
is continuous at the origin.
402 GUODONG LI AND WAI KEUNG LI
When |d| < 05, under Assumptions 1 and 2, the processes {Y
t
} and {
t
} generated by models
(1) and (2) are strictly stationary and ergodic with a nite second moment; see Ling & Li (1997).
Let = Eu
t
and
2
|u|
= var(|u
t
|). We dene the matrices

= E
_
1
h
t

_
,
1
=
_

0
0 0
_
,
2
= E
_
1
4h
2
t
h
t

h
t

_
,

3
= E
_
1
2h
3/2
t
_

h
t

+
h
t

_
_
,
1
=
1
+
2
|u|

2

3
,

2
= f (0)
1
+05
2
,
where f (0) is the value of the probability density function f (x) of u
t
evaluated at zero, the
matrix

is ( p +q +1) ( p +q +1) and other matrices are l l. Following the method in

Francq & Zakoian (2004), under Assumptions 1 and 2, we can show that the matrices
1
and
2
are positive denite.
THEOREM 1. Suppose that {Y
t
} are generated by models (1) and (2). Under Assumptions 13,
if |d| < 05, then there exists a sequence of local minimizers {

n
} of L
n
() such that

n(

n

0
) N(0, 025
1
2

1

1
2
)
in distribution as n .
Following a suggestion from a referee, we have not assumed that E(u
t
) = 0 in the above
theorem and the quantity = E(u
t
) is included in the covariance matrix as a parameter.
For a complicated ARFIMAGARCH model with many parameters, we may encounter compu-
tational difculty in nding the least absolute deviation estimator

n
. If the innovation u
t
is
further assumed to have a symmetric distribution, then = 0 and
2
is a block-diagonal matrix
since E{h
2
t
(h
t
/ )(h
t
/

)} is equal to zero. Hence, the matrices

1
,
2
and
1
2

1

1
2
are
all block-diagonal. This implies that we can separately minimize the score function L
n
() with
respect to and without incurring an asymptotic loss of efciency.
For the general case, a two-stage estimation approach seems more plausible in which we rst
apply least absolute deviation estimation to the ARFIMA part to nd a minimizer , and then apply
the approach of Peng & Yao (2003) to the residuals {
t
( )} to nd the estimator

. Alternatively,
we may obtain
LS
= arg min

n
t =1

2
t
( ) by least squares. However, to derive the asymptotic
normality of
LS
, we need to show that n
1/2

n
t =1

t
(
t
/ ) converges in distribution to
a normal distribution with a nite variance E{
2
t
(
t
/ )(
t
/

)}. If {
t
} are independent,
then the condition E
2
t
< is enough for the niteness of E{
2
t
(
t
/ )(
t
/

)}, but the

assumption E
4
t
< is unavoidable for the ARFIMAGARCH case. The simulation results in 4
suggest that both two-stage estimators are inferior to the least absolute deviation estimator

n
.
This is not surprising since, by ignoring the full model, two-stage methods should have lower
efciency. In real applications, the two-stage estimator can be used as an initial estimator and
then we can use the local quadratic approximation (Fan & Li, 2001) to minimize
n

t =1
_

2
t
( )
|
t
(
(m)
)|

{h
t
(, )}
+
1
2
log h
t
(, )
_
iteratively, where
(m)
is the minimizer in the mth iteration and
(0)
= . The above score
function is sufciently smooth and algorithms such as NewtonRaphson can be employed for the
optimization.
Time series models 403
When d > 05, the process {
t
} generated by (2) is still stationary. However, the process {Y
t
}
generated by models (1) and (2) is nonstationary.
Let d = m +d
f
, where |d
f
| < 05 and m is a positive integer. If U
t
= (1 B)
m
Y
t
, then U
t
follows the model (B)(1 B)
d
f
U
t
= (B)
t
. This means that, after mth-order differencing,
the nonstationary process {Y
t
} will be transformed to a stationary ARFIMA( p, d
f
, q)GARCH(r, s)
process {U
t
}. We next consider the asymptotic behaviour of the least absolute deviation estimation
for the process {U
t
}.
Let

= (d
f
,
1
, . . . ,
p
,
1
, . . . ,
q
)

and

= (

, where the rst elements d in both

and are replaced by d
f
. Denote by

1
and

2
the corresponding matrices associated with
{U
t
} instead of {Y
t
}. By an argument similar to that in Beran (1995) and Ling & Li (1997), it can
be shown that
1
=

1
and
2
=

2
. Let L

n
() be the corresponding score function. Then, by
Theorem 1, there exists a sequence of local minimizers {

n
} of L

n
() such that

n(

n

0
)
N(0, 025
1
2

1

1
2
) in distribution as n .
Let

n
=

n
+(m, 0, . . . , 0)

. Then

n
is a local minimizer of L
n
() and hence we have the
following results for the nonstationary ARFIMAGARCH models.
THEOREM 2. Suppose that {Y
t
} are generated by models (1) and (2). Under Assumptions 13,
if d > 05, then there exists a sequence of local minimizers {

n
} of L
n
() such that

n(

n

0
) N
_
0, 025
1
2

1

1
2
_
in distribution as n , where the matrices
1
and
2
are given as in Theorem 1.
The fractional ARIMA processes may sometimes include an unknown mean
Y
, leading to the
fractional ARIMA form
(B)(1 B)
d
f
{(1 B)
m
Y
t

Y
},
where |d
f
| < 05 and m 0 is an integer. We followBeran(1995) and Ling & Li (1997) in dealing
with this case. Let U
t
= (1 B)
m
Y
t
and

U = (n m)
1

n
t =m+1
U
t
. Then

U is a consistent
estimator of
Y
. We can centre the sequence {U
t
} on

U and then the methodology introduced
before can be used. The simulation results in 4 show that the estimators obtained by this method
are very similar to those obtained when the mean is known.
3. TWO PORTMANTEAU TESTS
This section constructs two portmanteau tests for checking whether or not the tted ARFIMA
GARCH models in the previous section are adequate. One test is based on the residual autocorre-
lations and the other is based on the absolute residual autocorrelations.
Let
t
and

h
t
be the corresponding values when the parameter vector in functions
t
()
and h
t
() is replaced by

n
, the least absolute deviation estimator from 2. From the proof of
Theorem 1, up to o
p
(1),

n(

n

0
)
1
2
n
1/2

1
2

n
t =1
_
1
2
(|u
t
| 1)h
1
t
h
t

sgn(u
t
)h
1/2
t

_
. (3)
Note that {
t
/

h
1/2
t
} is the residual sequence. Then the lag-k residual autocorrelation is
r
k
=

n
t =k+1
_

t
/

h
1/2
t

n
__

t k
/

h
1/2
t k

n
_

n
t =1
_

t
/

h
1/2
t

n
_
2
,
404 GUODONG LI AND WAI KEUNG LI
where
n
= n
1

n
t =1

t
/

h
1/2
t
, and the lag-k absolute residual autocorrelation is

k
=

n
t =k+1
_
|
t
|/

h
1/2
t
1
_ _
|
t k
|/

h
1/2
t k
1
_

n
t =1
_
|
t
|/

h
1/2
t
1
_
2
.
We next consider the asymptotic distributions of the rst M residual autocorrelations and absolute
residual autocorrelations.
Let

R = ( r
1
, . . . , r
M
)

and = E{u
t
(|u
t
| 1)}. Let X = (X
1
, . . . , X
M
), Z = (Z
1
, . . . , Z
M
)
and
4
= X

1
2
Z + Z

1
2
X, where X
k
= E{(u
t k
)h
1/2
t
(
t
/)} and Z
k
= E{(u
t k

)h
1
t
(h
t
/)} with k = 1, . . . , M. Along the lines of Li (1992), by the approximation in (3),
Taylor expansion, the central limit theorem and the MannWald device, we can show that

n

R N(0, V
1
),
in distribution as n , where
V
1
= I
1

4
X

1
2
(
2
025
1
)
1
2
X +

4
4

4
.
When the innovation u
t
is symmetrically distributed, the quantity is equal to zero and the last
term in the matrix V
1
disappears.
Let = (
1
, . . . ,
M
)

,

C = (

C
1
, . . . ,

C
M
)

and C = (C
1
, . . . , C
M
)

, where

C
k
=
1
n
n

t =k+1
_
|
t
|

h
1/2
t
1
__
|
t k
|

h
1/2
t k
1
_
, k = 1, . . . , M,
and C
k
is the corresponding value when

n
in

C
k
is replaced by the true parameter vector

0
. Let Z

= (Z

1
, . . . , Z

M
) and
5
= H

1
2
Z

+ Z

1
2
H, where Z

k
= E{05(|u
t k
|
1)h
1/2
t
(
t
/)} with k = 1, . . . , M. We can show that n
1

(|
t
|/

h
1/2
t
1)
2
=
2
|u|
+o
p
(1),
and then it is sufcient to derive the asymptotic distributions of

C. However, the vector

C is a
function of

n
and this function is not smooth. The classical Taylor expansion cannot be used
here and this is just the main difculty in deriving the asymptotic distribution of . Fortunately,
by the inequalities in the Appendix, we can show that

C = C H

n

0
) +o
p
_
n
1/2
_
, (4)
where H = (H
1
, . . . , H
M
) and H
k
= E{05(|u
t k
| 1)h
1
t
(h
t
/)} with k = 1, . . . , M. The
above equation plays the same role as Taylor expansion in Li (1992) and Li & Mak (1994).
Hence, as in Li & Li (2005), by the approximations in (3) and (4), the central limit theorem and
the MannWald device, we can obtain that

n N(0, V
2
)
in distribution as n , where
V
2
= I
1

4
|u|
H

1
2
_

2
|u|

2
025
1
_

1
2
H +

4
|u|

5
.
Note that = 0 when the innovation u
t
is symmetrically distributed.
Based on the asymptotic normality of

R and , we can construct two portmanteau tests,
Q
r
(M) = n

R

V
1
1

R, Q
a
(M) = n

V
1
2
.
Time series models 405
It can be shown that, if the ARFIMAGARCH model in the previous section is correctly specied,
the quantities Q
r
(M) and Q
a
(M) will be asymptotically distributed as
2
(M). In practice, we
can obtain the exact values of ,
2
,
2
|u|
and f (0) in the denitions of matrices V
1
and V
2
if
the distribution of u
t
is known. Otherwise, we can use n
1

t
/

h
1/2
t
to replace , n
1

2
t
/

h
t
to replace
2
and n
1

(|
t
|/

h
1/2
t
1)
2
to replace
2
|u|
. For f (0), the sequence {
t
/

h
1/2
t
} is rst
supposed to be independently and identically distributed and then some nonparametric method,
such as kernel estimation, can be applied to t the density function

f (x). Finally, we can use

f (0)
to replace f (0). The entries of X, Z, H, Z

,
2
and
3
can be replaced by the corresponding
sample averages, as in Li & Mak (1994). Tse & Zuo (1997) considered the optimal choice of M
for portmanteau tests proposed in Li & Mak (1994).
The tests Q
r
(M) and Q
a
(M) should be useful in checking whether or not the tted ARFIMA
GARCH models in 2 are adequate, and the simulation results in 4 give further support to this
fact. It can be seen that Q
r
(M) is not sensitive to the misspecication in conditional variances
while Q
a
(M) is not sensitive to the misspecication in conditional means. In fact, Wong & Ling
(2005) observed the same phenomenon for the residual autocorrelations and squared residual
autocorrelations, and proposed a combined portmanteau test based on these two types of auto-
correlation. Hence, it is also of interest to construct a combined portmanteau test, based on the
asymptotic joint distribution of residual autocorrelations and absolute residual autocorrelations.
However, this method is more complicated and we leave it for future research.
4. SIMULATION EXPERIMENTS
41. Performance of the least absolute deviation estimation
When the innovation u
t
is normally distributed, Ling & Li (1997) considered the maximum
likelihood estimator of models (1) and (2),

MLE
= arg min

t =1
_

2
t
()
h
t
()
+log h
t
()
_
,
and its asymptotic normality was also obtained. This estimator can still be used when the normality
of u
t
is violated, and is the so-called Gaussian quasi-maximum likelihood estimator. Note that
the proof in Ling & Li (1997) only needs the conditions E(
4
t
) < and E(u
4
t
) < , and hence

MLE
is still asymptotically normal under these conditions.
The rst experiment compares the least absolute deviation estimator

n
in 2 with the Gaussian
quasi-maximum likelihood estimator

MLE
and the following ARFIMA(0,d,0)GARCH(1,1) process
was involved:
(1 B)
d
Y
t
=
t
,
t
= u
t
h
1/2
t
, h
t
= 05 +02
2
t 1
+07h
t 1
, (5)
where u
t
followed a standard normal distribution or a Students t -distribution with 3 or 5 degrees
of freedom. These three distributions were always employed for the innovation u
t
except in
the next experiment, and were rst standardized to have mean 0 and variance 1. The memory
parameter d was chosen to be 03 or 03 for the stationary case and d = 07 or 13 for the
nonstationary case. For each combination of innovation distributions and memory parameters,
we considered the sample size n = 600 and drew 400 independent replications. The subroutine
DBCPOL in the IMSL library was employed to perform an exhaustive search for

MLE
and

n
at
the same time. We set initially the value of the parameter d to zero and the parameters in the
conditional variance,
0
,
1
and
1
, to 01. The subroutine DBCPOL was also used for optimization
in the following experiments and the real example. Since the values of parameters
0
and
1
406 GUODONG LI AND WAI KEUNG LI
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
__
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
__
_
_
_
_
__
_
_
_
_
_
_
_
_
_
_
_
_
_
_
__
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
__
_
__
_
__
__
_
_
_
_
LAD
t(3)
MLE
t(3)
LAD
t(5)
MLE
t(5)
LAD
nor
MLE
nor
1
.
0
0
.
8
0
.
6
0
.
4
0
.
2
0
.
0
1
.
0
0
.
8
0
.
6
0
.
4
0
.
2
0
.
0
d = 0
.
3 (a)
_
_
_
_
_
_
_
_
_
__
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
__
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
__
_
_
_
_
_
_
_
_
_
_
___
_
_
_
__
_
_
_
_
_
_
_
_
_
__
_
_
_
_
_
LAD
t(3)
MLE
t(3)
LAD
t(5)
MLE
t(5)
LAD
nor
MLE
nor
d = 0
.
7 (b)
Fig. 1. Boxplots for the average absolute errors of

n
and

MLE
; for (a) a stationary case with d = 03, (b) a
nonstationary case with d = 07. Labels t (3), t (5) or nor indicate that the innovation u
t
has the t
3
, t
5
or N(0, 1)
distribution, respectively.
tted by the least absolute deviation approach are different from 05 and 02 by a common factor
as in Peng & Yao (2003), we dene the average absolute error as
1
3
(|
1
/
0
04| +|

1
02| +|

d d|),
which can be used to compare the performance of

n
with

MLE
.
Figure 1 displays the boxplots for the average absolute error for one stationary case, d = 03,
and one nonstationary case, d = 07. The results for d = 03 and d = 13 were very similar to
those in Fig. 1(a) and (b), respectively. There are a few very large values of average absolute
errors for

MLE
when the errors are distributed as t
3
. For convenience of presentation, we have
removed these outliers fromthe gures. The least absolute deviation estimator

n
is much superior
when u
t
t
3
. This may reect the fact that the heavier the tails, the slower the convergence rate
of the Gaussian quasi-maximum likelihood estimator; see Hall & Yao (2003). Note that the t
5
distribution has a nite fourth moment so that

MLE
will enjoy the standard n
1/2
convergence
rate. For this case,

n
also performs better. When the error is normally distributed, of course,

MLE
is the better choice, but the performance of

n
is comparable.
As suggested by a referee, we also conduct an experiment to compare

n
with two two-stage
estimators. For simplicity, an AR(1)ARCH(1) model is employed to generate the realizations:
Y
t
= 03Y
t 1
+
t
,
t
= u
t
_
05 +02
2
t 1
_
1/2
.
The innovation sequence {u
t
} come from a mixed t -distribution, taking the values of |t
f
1
| and
|t
f
2
| respectively with probability 05. Four different distributions of u
t
are considered in this
experiment with ( f
1
, f
2
) respectively equal to (3, 3), (5, 5), (3, 5) and (9, 5). Note that the median
of u
t
is zero, but when f
1
f
2
, u
t
is asymmetrically distributed with E(u
t
) 0. The following
two-stage estimation method, 2LAD, was employed:
(
0
,
1
)

= arg min
n

t =1
|log
2
t
(

) log h
t
(

,
0
,
1
)|,
Time series models 407
____
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
__
_
_
__
_
_
_
__
_
_
_
_
_
_
_
_
_
_
_
_
LAD
MLE
2LS 2LAD LAD
MLE
2LS 2LAD
( f
1
, f
2
) = (3, 3) (a)
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
__
_
_
_
_
_
_
_
_
_
_
_
_
__
_
_
_
_
_
_
_
_
_
_
_
( f
1
, f
2
) = (5, 5) (b)
1
.
0
0
.
8
0
.
6
0
.
4
0
.
2
0
.
0
1
.
0
0
.
8
0
.
6
0
.
4
0
.
2
0
.
0
Fig. 2. Boxplots for average absolute errors of the four types of estimator with symmetric innovation
u
t
for (a) ( f
1
, f
2
) = (3, 3) and (b) ( f
1
, f
2
) = (5, 5), for methods LAD, MLE, 2LS and 2LAD.
where

is the median of the sequence {Y
t
/Y
t 1
, t = 2, . . . , n}. For comparison, another two-
stage procedure, 2LS, was also considered:

= arg min
n

t =1

2
t
(), (
0
,
1
)

= arg min
n

t =1
|log
2
t
(

) log h
t
(

,
0
,
1
)|,
where
t
() = Y
t
Y
t 1
since = Eu
t
may not be equal to zero. The same adjustment is
applied to

MLE
. The boxplots for the average absolute errors of the above four different methods
are presented in Fig. 2 for the cases of symmetric innovation distributions corresponding to
( f
1
, f
2
) = (3, 3) and ( f
1
, f
2
) = (5, 5). For the cases of asymmetric innovation distributions,
the results for ( f
1
, f
2
) = (3, 5) and ( f
1
, f
2
) = (9, 5) were very similar to Fig. 2(a) and (b),
respectively. The results show that the least absolute deviation estimator

n
outperforms the other
three. This conclusion is consistent with the argument in 2.
The nal experiment examines the performance of

n
in nite-sample cases. The generating
process (5) with d = 03 was employed again. Note that the series {Y
t
} has the long-memory
property. The sample size is set to be 300 or 400, and we drew 500 independent replications
for each combination of the sample sizes and the innovation distributions. Table 1 presents the
estimated biases, the empirical root mean squared errors and the root mean asymptotic variance
of the estimators. The biases are all small and the root mean asymptotic variances are very similar
to the empirical root mean squared errors. All biases, empirical root mean squared errors and root
mean asymptotic variances change little when the series is centred by the sample mean and they
decrease as the sample size increases. The empirical root mean squared errors and the root mean
asymptotic variances become very similar when the sample size is larger, n = 400. We also tried
different memory parameters for the generating process (5) and found very similar results.
42. Performance of the portmanteau tests
In this section, we check the empirical sizes and powers of the two portmanteau tests ob-
tained in 3. Three different generating processes were involved: the ARFIMA(0,d,0)GARCH(1,1)
408 GUODONG LI AND WAI KEUNG LI
Table 1. Simulation study. Estimated bias, square root of the mean squared error, RMSE, and
square root of the mean asymptotic variance, RMAV, for the ARFIMA (0, 03, 0)GARCH (1, 1)
model, based on 500 replications; all gures multiplied by 10
Known mean Unknown mean
n

d
0

1

1

d
0

1

1
Innovation distribution t (3)
300 BIAS 0023 0695 0072 0723 0024 0734 0063 0751
RMSE 0457 1938 0565 2068 0499 1964 0568 2099
RMAV 0450 2257 0540 2355 0462 2309 0538 2386
400 BIAS 0009 0516 0046 0478 0045 0542 0045 0490
RMSE 0387 1565 0422 1590 0423 1613 0425 1614
RMAV 0389 1537 0455 1603 0398 1550 0455 1611
Innovation distribution t (5)
300 BIAS 0045 0747 0015 0406 0115 0743 0009 0391
RMSE 0559 2063 0479 1447 0589 2040 0476 1414
RMAV 0520 2114 0486 1448 0532 2147 0485 1467
400 BIAS 0025 0626 0001 0279 0069 0629 0005 0276
RMSE 0469 1765 0402 1163 0479 1759 0394 1154
RMAV 0452 1725 0416 1169 0461 1733 0415 1176
Innovation distribution N(0, 1)
300 BIAS 0035 0919 0006 0309 0129 0918 0000 0305
RMSE 0606 2317 0514 1227 0622 2332 0511 1233
RMAV 0593 2616 0472 1339 0605 3091 0471 1562
400 BIAS 0052 0720 0004 0260 0141 0706 0005 0246
RMSE 0542 1931 0423 1023 0586 1922 0423 1018
RMAV 0513 1984 0409 1055 0522 1977 0407 1052
process,
(1 B)
d
Y
t
=
t
,
t
= u
t
h
1/2
t
, h
t
= 03 +03
2
t 1
+03h
t 1
,
was used to check the empirical sizes; the ARFIMA(0,d,0)GARCH(3,1) process,
(1 B)
d
Y
t
=
t
,
t
= u
t
h
1/2
t
, h
t
= 03 +03
2
t 1
+03
2
t 3
+03h
t 1
,
was used to check the sensitivity for the misspecication of conditional variance, and we call this
Type I power; and the ARFIMA(2,d,0)GARCH(1,1) process,
(1 02B
2
)(1 B)
d
Y
t
=
t
,
t
= u
t
h
1/2
t
, h
t
= 03 +03
2
t 1
+03h
t 1
,
was used to check the sensitivity for the misspecication of conditional mean, and we call this Type
II power. The memory parameter d was taken to be 03, 03 or 07, resulting in series with the
short-memory, long-memory or nonstationary property, respectively. Two different sample sizes,
n = 400 and n = 600, were considered and there were 1000 replications for each combination
of the memory parameters d, sample sizes n, and the innovation distributions. We estimated all
the simulated data with the ARFIMA(0, d, 0)GARCH(1,1) model by the least absolute deviation
approach and calculated the values of Q
a
(M) and Q
r
(M) with M = 6.
Table 2 displays the proportions of rejections based on the upper 5th percentile of the
2
6
distribution. All the sizes of Q
a
(6) and Q
r
(6) in Table 2 are very close to 005 especially for the
cases with n = 600. Type I powers of Q
a
(6) and Type II powers of Q
r
(6) are all greater than 05
when the sample size n is as large as 600. This means that the tests Q
a
(M) and Q
r
(M) can be
used to check respectively whether or not the conditional variance part or the conditional mean
part of the tted model is misspecied. Type I powers of Q
r
(6) and Type II powers of Q
a
(6)
Time series models 409
Table 2. The empirical size and power of Q
a
(6) and Q
r
(6), based on 1000 replications
Size Type I Power Type II Power
Innovation distribution n Q
a
(6) Q
r
(6) Q
a
(6) Q
r
(6) Q
a
(6) Q
r
(6)
Differencing parameter d = 03
t
3
400 0043 0043 0328 0078 0039 0374
600 0046 0044 0500 0088 0037 0620
t
5
400 0055 0063 0733 0099 0052 0620
600 0053 0047 0901 0104 0045 0801
N(0, 1) 400 0058 0047 0919 0105 0048 0691
600 0045 0053 0984 0099 0038 0990
Differencing parameter d = 03
t
3
400 0051 0038 0338 0087 0039 0388
600 0053 0043 0462 0086 0034 0589
t
5
400 0061 0043 0747 0108 0047 0621
600 0051 0052 0860 0094 0039 0824
N(0, 1) 400 0055 0056 0913 0104 0063 0698
600 0048 0053 0981 0092 0049 0886
Differencing parameter d = 07
t
3
400 0052 0048 0319 0082 0038 0346
600 0049 0043 0475 0073 0041 0588
t
5
400 0052 0049 0742 0105 0050 0599
600 0048 0048 0900 0092 0038 0815
N(0, 1) 400 0047 0043 0923 0124 0050 0701
600 0051 0043 0987 0114 0053 0875
are no more than 015. Hence, only the combination of Q
a
(M) and Q
r
(M) forms a complete
diagnostic tool for checking whether or not the tted ARFIMAGARCH model is adequate.
5. AN ILLUSTRATIVE EXAMPLE
The dataset contains the absolute returns, as a percentage, of the Dow Jones Industrial Average
Index daily closing price. There are 2519 observations from January 3, 1995 to December 31,
2004. The mean and standard deviation of the absolute returns are 0352 and 0335, respectively.
Denote the centralized absolute returns by {y
t
}. The conditional heteroscedasticity in the time
series is obvious fromFig. 3(a). Figure 3(b) shows that the sample autocorrelations of the absolute
returns are relatively small in magnitude, but decay very slowly. They appear to be signicant
at the 5% level even after 200 lags, suggesting the presence of long memory (Tsay, 2002).
The ARFIMAGARCH models were considered for the observed series {y
t
} and the least absolute
deviation method was used to nd the memory parameter as well as others.
We considered two different models for the conditional mean, ARFIMA(4, d, 0) and
ARFIMA(2, d, 1), and two different models for the conditional variance, ARCH(4) and GARCH(1, 1),
leading to four models in total: Model 1 combines an ARFIMA(4, d, 0) model with ARCH(4) er-
ror; Model 2 combines an ARFIMA(4, d, 0) model with GARCH(1, 1) error; Model 3 combines an
ARFIMA(2, d, 1) model with ARCH(4) error; and Model 4 combines an ARFIMA(2, d, 1) model with
GARCH(1, 1) error. The methodology in 2 and 3 was applied to these four models. We set M to
be 15 and the values of ,
2
,
2
|u|
and f (0) were estimated with the methods mentioned in 3.
The bandwidth was set to be 005. The results are summarized in Table 3.
Only Model 4 is not rejected by either of the test statistics Q
a
(15) and Q
r
(15) at signicance
level 005 and hence is adequate; note that
2
15,095
= 2500. Consistent with the simulation results
410 GUODONG LI AND WAI KEUNG LI
0 500 1000 1500 2000 2500
Time in days
(a)
0 50 100 150 200
Lag
(b)
0
.
0
1
.
0
2
.
0
3
.
0
1
.
0
0
.
8
0
.
6
0
.
4
0
.
2
0
.
0
Fig. 3. Daily closing Dow Jones Industrial Average Index, 19952004. (a) Time plot of absolute returns,
(b) Empirical autocorrelation function of absolute returns.
Table 3. Modelling results for the absolute returns of Dow Jones Industrial Average Index daily
closing price, 19952004
Model 1 Model 2 Model 3 Model 4
Parameter

n
SD

n
SD

n
SD

n
SD
d 04349 5672 04651 5541 06594 1240 07117 1350

1
04573 6818 04786 6570 00784 7118 00618 7815

2
02611 6526 02827 6393 00170 4334 00193 3912

3
01794 5543 02052 5591

4
00707 4280 01004 4499

1
07699 7925 08004 7673

0
00242 1982 00009 0242 00229 1925 00008 0223

1
00973 2150 00465 7264 01011 2160 00469 7201

2
01045 2200 01086 2230

3
00746 1950 00815 2010

4
00876 2060 00923 2090

1
09084 1270 09100 1220
Q
r
(15) 3043 3312 1769 1596
Q
a
(15) 10940 1079 9783 1088
SD, estimated asymptotic standard deviation multiplied by 10
3
.
Time series models 411
2 4 6 8 10 12 14
2 4 6 8 10 12 14
2 4 6 8 10 12 14
2 4 6 8 10 12 14
2 4 6 8 10 12 14
2 4 6 8 10 12 14
2 4 6 8 10 12 14
2 4 6 8 10 12 14
0
.
10
0
.
00
0
.
10
0
.
10
0
.
00
0
.
10
0
.
10
0
.
00
0
.
10
0
.
10
0
.
00
0
.
10
0
.
10
0
.
00
0
.
10
0
.
10
0
.
00
0
.
10
0
.
10
0
.
00
0
.
10
0
.
10
0
.
00
0
.
10
Model 1
R
e
s
i
d
u
a
l
s
A
b
s
o
l
u
t
e

R
e
s
i
d
u
a
l
s
Model 2 Model 3 Model 4
Fig. 4. DowJones example. Sample autocorrelation functions of residuals and absolute residuals obtained fromtting
Models 14. The dotted lines show 196A
i
, the 95% asymptotic condence intervals, where A
i
, i = 1, . . . , 15, is
the asymptotic standard error for lag i .
in 5, at signicance level 005, the portmanteau test Q
a
(15) rejects Models 1 and 3, suggesting
that the conditional variances are misspecied, while Q
r
(15) rejects Models 1 and 2, suggesting
that the conditional means are misspecied. Figure 4 presents the sample autocorrelation functions
of residuals and absolute residuals coming from the above four tted models, along with 95%
condence bands. These plots are consistent with the above ndings. We tried several other values
of M and similar results were obtained.
The value

d = 07117 can be considered as a reliable estimate of the true memory parameter
since Model 4 is the selected model, indicating that this sequence is probably nonstationary.
Note that all the estimated memory parameters are signicantly different, with 95% condence
intervals that do not even overlap. Models 1 and 2 even fail to identify that the sequence is
nonstationary. Hence it is important to specify the conditional mean and variance correctly when
the least absolute deviation method is employed to estimate ARFIMAGARCH models.
ACKNOWLEDGEMENT
W. K. Li thanks the Croucher Foundation for awarding a Senior Research Fellowship and the
Hong Kong Research Grant Council for a grant for partial support. The authors thank Dr S. Ling
for useful comments and discussion. We thank Professor D.M. Titterington and two referees for
comments that led to improvement of the paper.
APPENDIX
Technical details
Some properties of the derivative functions of
t
() and h
t
(). Under Assumptions 1 and 2, the functions,

t
() and h
t
(), dened in 2 are both meaningful and their rst-order derivatives are as follows:

t
()

j
=
1
(B)
t j
(),

t
()

j
=
1
(B)
t j
(),

t
()
d
=

k=1
1
k

t k
(),
h
t
()

=
t
+
s

i =1

i
h
t i
()

,
412 GUODONG LI AND WAI KEUNG LI
h
t
()

= 2
r

i =1

t i
()

t i
()

+
s

i =1

i
h
t i
()

,
where
t
= (1,
2
t 1
(), . . . ,
2
t r
(), h
t 1
(), . . . , h
t s
())

; see Ling & Li (1997).

If the condition |d| < 05 is also assumed, the processes {Y
t
} and {
t
} generated by models (1) and (2)
are strictly stationary and have a nite second moment. Let

= {|d| < 05}, which is still compact.

Then we can show that
E
_
sup

_
_
_
_

t
()

_
_
_
_
2
_
< , E
_
sup

_
_
_
_
1

{h
t
()}

t
()

_
_
_
_
2
_
< ,
E
_
sup

_
_
_
_

t
()

_
_
_
_
2
_
< , E
_
sup

_
_
_
_
1

{h
t
()}
h
t
()

_
_
_
_
2
_
< ,
E
_
sup

_
_
_
_
1
h
t
()
h
t
()

_
_
_
_
2
_
< , E
_
sup

_
_
_
_
1
h
t
()

2
h
t
()

_
_
_
_
2
_
< ,
where is the Euclidean norm. The above inequalities are necessary for the derivations in 3 and the
proof of Theorem 1.
Proof of Theorem 1. For any v = (v

1
, v

2
)

R
l
, where v
1
R
p+q+1
, v
2
R
r+s+1
and v 0, let
S
n
(v) = L
n
_

0
+n
1/2
v
_
L
n
(
0
)
=
n

t =1
_
1

{h
t
(
0
)}
_

t
_

0
+n
1/2
v
_

|
t
(
0
)|
_
_
+
n

t =1
__
1

_
h
t
_

0
+n
1/2
v
__
1

{h
t
(
0
)}
_
_
|
t
_

0
+n
1/2
v
_
| |
t
(
0
)|
_
_
+
n

t =1
_
|
t
(
0
)|

_
h
t
_

0
+n
1/2
v
__ +
1
2
log h
t
_

0
+n
1/2
v
_

|
t
(
0
)|

{h
t
(
0
)}

1
2
log h
t
(
0
)
_
= S
1n
(v) + S
2n
(v) + S
3n
(v),
S

1n
(v) =
n

t =1
_
1

{h
t
(
0
)}
_

t
(
0
) +n
1/2
v
1

t
(
0
)

|
t
(
0
)|
__
.
It holds that, for x 0,
|x + y| |x| = y sgn(x) +2
_
y
0
{I (x s) I (x 0)}ds,
where the function sgn(x) is equal to 1 for x > 0, 0 for x = 0 and 1 for x < 0; see Knight (1998). By a
method similar to that in Peng & Yao (2003) together with the above equation, we can show that
S

1n
(v) = n
1/2
v

1
n

t =1
sgn(u
t
)h
1/2
t

+ f (0)v

v
1
+o
p
(1).
Time series models 413
As in Davis & Dunsmuir (1997), by the inequalities at the beginning of this Appendix, we can show that
S
1n
(v) n
1/2
v

1
n

t =1
sgn(u
t
)h
1/2
t

+ f (0)v

v
1
in distribution as n . Hence, S
2n
(v) = o
p
(1).
Note that S
3n
(v) is a smooth function so that Taylor expansion can be used. After some algebra, the
inequalities at the beginning of this Appendix ensure that
S
3n
(v) = 05 n
1/2
v

t =1
(|u
t
| 1)h
1
t
h
t

+05 v

2
v +o
p
(1).
Let
s
n
= v

t =1
_
sgn(u
t
)h
1/2
t

05(|u
t
| 1)h
1
t
h
t

_
.
Note that, under Assumption 3, s
n
is a martingale and (1/n)Es
2
n
= v

1
v > 0, where the matrix
1
=

1
+
2
|u|

2

3
is dened in 2. By the strict stationarity and ergodicity of {Y
t
} and {
t
}, it holds that
_
(1/n)Es
2
n
_
1
_
(1/n)E
_
s
2
n

F
n1
__
1,
almost surely. Using the central limit theorem of Stout (1974), we can show that n
1/2
s
n
v

W in
distribution, where W is a multivariate normal vector with mean 0 and covariance matrix
1
. Hence,
S
n
(v) v

W +v

2
v
in distribution, where
2
= f (0)
1
+05
2
is dened in 2. Following Lemma 22 and Remark 1 of
Davis et al. (1992) we complete the proof of Theorem 1.
REFERENCES
BAILLIE, R. T., CHUNG, C. F. & TILES, M. A. (1996). Analyzing ination by the fractionally integrated ARFIMA-GARCH
model. J. Appl. Economet. 11, 2340.
BERAN, J. (1995). Maximum likelihood estimation of the differencing parameter for invertible short and long memory
autoregressive integrated moving average models. J. R. Statist. Soc. B 57, 65972.
BERAN, J. & FENG, Y. H. (2001). Local polynomial estimation with a ARFIMA-GARCH error process. Bernoulli 7,
73350.
BERKES, I. & HORVATH, L. (2004). The efciency of the estimators of the parameters in GARCH processes. Ann. Statist.
32, 63355.
BHARDWAJ, G. & SWANSON, N. R. (2006). An empirical investigation of the usefulness of ARFIMA models for predicting
macroeconomic and nancial time series. J. Economet. 131, 53978.
BOLLERSLEV, T. (1986). Generalized autoregression conditional heteroskedasticity. J. Economet. 31, 30727.
DAVIS, R. A. & DUNSMUIR, W. T. M. (1997). Least absolute deviation estimation for regression with ARMA errors. J.
Theor. Prob. 10, 48197.
DAVIS, R. A., KNIGHT, K. & LIU, J. (1992). M-estimation for autoregressions with innite variances. Stoch. Proces.
Applic. 40, 14580.
EMBRECHTS, P., KLUPPELBERG, C. & MIKOSCH, T. (1997). Modelling Extremal Events. Berlin: Springer.
ENGLE, R. F. (1982). Autoregression conditional heteroscedasticity with estimates of the variance of U.K. ination.
Econometrica 50, 9871008.
FAN, J. & LI, R. (2001). Variable selection via nonconcave penalized likelihood and its oracle properties. J. Am. Statist.
Assoc. 96, 134860.
FRANCQ, C. & ZAKOIAN, J. M. (2004). Maximum likelihood estimation of pure GARCH and ARMA-GARCH processes.
Bernoulli 10, 60537.
GRANGER, C. W. J. &JOYEUX, R. (1980). An introduction to long-memory time series models and fractional difference.
J. Time Ser. Anal. 1, 1539.
GRANGER, C. W. J., SPEAR, S. & DING, Z. X. (1999). Stylized facts on the temporal and distributional properties
of absolute returns: an update. In Statistics and Finance: An Interface, Ed. W. S. Chan, W. K. Li and H. Tong,
pp. 97120. London: Imperial College Press.
414 GUODONG LI AND WAI KEUNG LI
HALDRUP, N. & NIELSEN, M. O. (2007). Estimation of fractional integration in the presence of data noise. Comp.
Statist. Data Anal. 51, 310014.
HALL, P. & YAO, Q. (2003). Inference in ARCH and GARCH models with heavy-tailed errors. Econometrica 71, 285317.
HAUSER, M. A. & HURST, R. M. (2001). Forecasting high-frequency nancial data with the ARFIMA-ARCH model. J.
Forecasting 20, 50118.
HOSKING, J. R. M. (1981). Fractional differencing. Biometrika 68, 16576.
KNIGHT, K. (1998). Limiting distributions for l
1
regression estimators under general conditions. Ann. Statist. 26,
75570.
LI, G. & LI, W. K. (2005). Diagnostic checking for time series models with conditional heteroscedasticity estimated
by the least absolute deviation approach. Biometrika 92, 691701.
LI, W. K. (1992). On the asymptotic distribution of residual autocorrelations in nonlinear time series modelling.
Biometrika 79, 4357.
LI, W. K. & MAK, T. K. (1994). On the squared residual autocorrelations in nonlinear time series with conditional
heteroskedasticity. J. Time Ser. Anal. 15, 62736.
LING, S. (2007). A double AR(p) model: structure and estimation. Statist. Sinica 17, 16175.
LING, S. & LI, W. K. (1997). On fractionally integrated autoregressive moving average time series models with
conditional heteroscedasticity. J. Am. Statist. Assoc. 92, 118493.
MCLEOD, A. I. & HIPEL, K. W. (1978). Preservation of the rescaled adjusted range 1. A reassessment of the hurst
phenomenon. Water Resour. Res. 14, 491508.
MIKOSCH, T. &STARICA, C. (2000). Limit theory for the sample autocorrelations and extremes of a GARCH(1,1) process.
Ann. Statist. 28, 142751.
MITTNIK, S., RACHEV, S. T. &PAOLELLA, M. S. (1998). Stable Paretian models in nance: some empirical and theoretical
aspects. In A Practical Guide to Heavy Tails, Ed. R. J. Adler, R. E. Feldman and M. S. Taqqu, pp. 79110. Boston:
Birkhauser.
PENG, L. & YAO, Q. (2003). Least absolute deviations estimation for ARCH and GARCH models. Biometrika 90, 96775.
ROBINSON, P. M. (2003). Long-memory time series. In Time Series with Long Memory, Ed. P. M. Robinson, pp. 432.
Oxford: Oxford University Press.
STOUT, W. F. (1974). Almost Sure Convergence. New York: Academic Press.
TSAY, R. S. (2002). Analysis of Financial Time Series. New York: John Wiley & Sons.
TSE, Y. K. & ZUO, X. L. (1997). Testing for conditional heteroscedasticity: some Monte Carlo results. J. Statist. Comp.
Simul. 58, 23753.
WONG, H. & LING, S. (2005). Mixed portmanteau tests for time-series models. J. Time Ser. Anal. 26, 56979.
[Received January 2006. Revised November 2007]