Metodolo ski zvezki, Vol. 2, No.

2, 2005, 243-257
Properties and Estimation of GARCH(1,1) Model
Petra Posedel
1
Abstract
We study in depth the properties of the GARCH(1,1) model and the assump-
tions on the parameter space under which the process is stationary. In particular, we
prove ergodicity and strong stationarity for the conditional variance (squared volatil-
ity) of the process. We show under which conditions higher order moments of the
GARCH(1,1) process exist and conclude that GARCH processes are heavy-tailed.
We investigate the sampling behavior of the quasi-maximum likelihood estimator
of the Gaussian GARCH(1,1) model. A bounded conditional fourth moment of the
rescaled variable (the ratio of the disturbance to the conditional standard deviation) is
sufcient for the result. Consistent estimation and asymptotic normality are demon-
strated, as well as consistent estimation of the asymptotic covariance matrix.
1 Introduction
Financial markets react nervously to political disorders, economic crises, wars or natural
disasters. In such stress periods prices of nancial assets tend to uctuate very much.
Statistically speaking, it means that the conditional variance for the given past
Var(X
t
|X
t1
, X
t2
, . . .)
is not constant over time and the process X
t
is conditionally heteroskedastic. Econome-
tricians usually say that volatility

t
=
_
Var(X
t
|X
t1
, X
t2
, . . .)
changes over time. Understanding the nature of such time dependence is very important
for many macroeconomic and nancial applications, e.g. irreversible investments, option
pricing, asset pricing etc. Models of conditional heteroskedasticity for time series have a
very important role in todays nancial risk management and its attempts to make nancial
decisions on the basis of the observed price asset data P
t
in discrete time. Prices P
t
are
believed to be nonstationary so they are usually transformed in the so-called log returns
X
t
= log P
t
log P
t1
.
Log returns are supposed to be stationary, at least in periods of time that are not too
long. Very often in the past it was suggested that (X
t
) represents a sequence of inde-
pendent identically distributed random variable, in other words, that log returns evolve
1
Faculty of Economics, University of Zagreb, Zagreb, Croatia
244 Petra Posedel
like a random walk. Samuelson suggested modelling speculative prices in the continuous
time with the geometric Brownian motion. Discretization of that model leads to a random
walk with independent identically distributed Gaussian increments of log return prices in
discrete time. This hypothesis was rejected in the early sixties. Empirical studies based
on the log return time series data of some US stocks showed the following observations,
the so-called stylized facts of nancial data:
serial dependence are present in the data
volatility changes over time
distribution of the data is heavy-tailed, asymmetric and therefore not Gaussian.
These observations clearly showthat a random walk with Gaussian increments is not a
very realistic model for nancial data. It took some time before R. Engle found a discrete
model that described very well the previously mentioned stylized facts of nancial data,
but it was also relatively simple and stationary so the inference was possible. Engle called
his model autoregressive conditionally heteroskedastic- ARCH, because the conditional
variance (squared volatility) is not constant over time and shows autoregressive structure.
Some years later, T. Bollerslev generalized the model by introducing generalized au-
toregressive conditionally heteroskedastic - GARCH model. The properties of GARCH
models are not easy to determine.
2 GARCH(1,1) process
Denition 2.1 Let (Z
n
) be a sequence of i.i.d. random variables such that Z
t
N(0, 1).
(X
t
) is called the generalized autoregressive conditionally heteroskedastic or GARCH(q,p)
process if
X
t
=
t
Z
t
, t Z (2.1)
where (
t
) is a nonnegative process such that

2
t
=
0
+
1
X
2
t1
+. . . +
q
X
2
tq
+
1

2
t1
+. . . +
p

2
tp
, t Z (2.2)
and

0
> 0,
i
0 i = 1, . . . , q
i
0 i = 1, . . . , p. (2.3)
The conditions on parameters ensure strong positivity of the conditional variance (2.2).
If we write the equation (2.2) in terms of the lag-operator B we get

2
t
=
0
+(B)X
2
t
+(B)
2
t
, (2.4)
where
(B) =
1
B +
2
B
2
+. . . +
q
B
q
and
(B) =
1
B +
2
B
2
+. . . +
p
B
p
. (2.5)
Properties and Estimation of GARCH(1,1) Model 245
If the roots of the characteristic equation, i.e.
1
1
x
2
x
2
. . .
p
x
p
= 0
lie outside the unit circle and the process (X
t
) is stationary, then we can write (2.2) as

2
t
=

0
1 (1)
+
(B)
1 (B)
X
2
t
=

0
+

i=1

i
X
2
ti
(2.6)
where

0
=

0
1 (1)
, and
i
are coefcients of B
i
in the expansion of (B)[1(B)]
1
.
Note that the expression (2.6) tells us that the GARCH(q,p) process is an ARCH process
of innite order with a fractional structure of the coefcients.
From (2.1) it is obvious that the GARCH(1,1) process is stationary if the process
(
2
t
) is stationary. So if we want to study the properties and higher order moments of
GARCH(1,1) process it is enough to do so for the process (
2
t
).
The following theorem gives us the main result for stochastic difference equations that
we are going to use in order to establish the stationarity of the process (
2
t
).
Theorem 2.2 Let (Y
t
) be the stochastic process dened by
Y
t
= A
t
+B
t
Y
t1
, t N, (2.7)
or explicitly
Y
t
= Y
0
t

j=1
B
j
+
t

m=1
A
m
t

j=m+1
B
j
, t N. (2.8)
Suppose that Y
0
is independent of the i.i.d. sequence
_
(A
t
, B
t
)
_
t1
. Assume that
E ln
+
|A| < and E ln |B| < 0. (2.9)
Then
(a) Y
t
D
Y for some random variable Y such that it satises the identity in law
Y = A+BY, (2.10)
where Y and (A, B) are independent.
(b) Equation (2.10) has a solution, unique in distribution, which is given by
Y
D
=

m=1
A
m
m1

j=1
B
j
. (2.11)
The right hand side of (2.11) converges absolutely with probability 1.
246 Petra Posedel
(c) If we choose Y
0
D
= Y as in (2.11), then the process (Y
t
)
t0
is strictly stationary.
Now assume the moment conditions
E|A|
p
< and E|B|
p
< 1 for some p [1, ).
(d) Then E|Y |
p
< , and the series in (2.11) converges in pth mean.
(e) If E|Y
0
|
p
< , then (Y
t
) converges to Y in pth mean, and in particular
E|Y
t
|
p
E|Y |
p
as t .
(f) The moments EY
m
are uniquely determined by the equations
EY
m
=
m

k=0
_
m
k
_
E
_
B
k
A
mk
_
EY
k
, m = 1, . . . , p (2.12)
where p denotes the oor function.
In the next theorem we present the stationarity of the conditional variance process (
2
t
).
Theorem 2.3 Let (
2
t
) be the conditional variance of GARCH(1,1) process dened with
(2.1) and (2.2). Additionally, assume that
E
_
ln
_

1
Z
2
0
+
1
_
< 0 (2.13)
and that
2
0
is independent from (Z
t
). Then it holds
(a) the process (
2
t
) is strictly stationary if

2
0
D
=
0

m=1
m1

j=1
_

1
+
1
Z
2
j1
_
(2.14)
and the series (2.14) converges absolutely with probability 1.
(b) Assume that (
2
t
) is strictly stationary and let =
2
0
, Z = Z
1
. Let E
_

1
+

1
Z
2
_
p
< 1 for some p [1, ). Then E
_

2
_
m
< for some 1 m p. For
such integer m it holds
E
_

2m

=
_
1 E
_

1
+
1
Z
2
_
m

1
m1

k=0
_
m
k
_
E
_

1
Z
2
+
1
_
k

mk
0

E
_

2k
] < . (2.15)
Proof: From (2.2) we have

2
t
=
0
+
1
X
2
t1
+
1

2
t1
,
or

2
t
=
0
+
_

1
Z
2
t1
+
1
_

2
t1
Properties and Estimation of GARCH(1,1) Model 247
that represents a stochastic difference equation
Y
t
= A
t
+B
t
Y
t1
,
where Y
t
=
2
t
, A
t
=
0
and B
t
=
1
Z
2
t1
+
1
. From the assumptions of the theorem we
have that E ln
+
|A| < and E ln |B| = E
_
ln
_

1
+
1
Z
2
t1
_
< 0. So, fromTheorem2.2
we have that
_

2
t
_
is strictly stationary with unique marginal distribution given by (2.14)
and this shows the rst statement of the theorem. Additionally, suppose that E
_

1
+

1
Z
2
_
p
< 1. In that case we have E|B|
p
= E
_

1
+
1
Z
2
_
p
< 1 for some p [1, ) so
from part (f) of Theorem 2.2 it follows (2.15).

Example 2.4 Let (X

t
) be GARCH(1,1) process.Let
(
1
,
1
, p) = E
_

1
Z
2
+
1
_
p
, p [1, ).
In that case, it follows from Theorem 2.3 that a necessary condition for the existence of
the stationary moment of order 2m, 1 m p, of a GARCH(1,1) process is given by
(
1
,
1
, p) < 1.
In the special case of m = 2 it follows that the stationary fourth moment of the GARCH(1,1)
process exists if
(
1
,
1
, 2) =
2

j=0
_
2
j
_
a
j

j
1

mj
1
< 1,
that is equivalent to

2
1
+ 2
1

1
+ 3
2
1
< 1.
From the recursive formula given in the Theorem 2.2 in the case of m = 1 and m = 2 we
obtain
E
_
X
2
t
_
= E
_
Z
2
t
_
E
_

2
t
_
=

0
1
1

1
and
E
_
X
4
t
_
= E
_
Z
4
t
_
E
_

4
t
_
= 3
_

2
0
+ 2E
_
X
2
t
_

0
(
1
+
1
)

_
1
2
1
2
1

1
3
2
1

1
= 3
_

2
0
+ 2

2
0
1
1

1
(
1
+
1
)
_

_
1
2
1
2
1

1
3
2
1

1
= 3
2
0
_
1 + 2

1
+
1
1
1

1
_

_
1
2
1
2
1

1
3
2
1

1
= 3
2
0
(1 +
1
+
1
)
_
(1
1

1
)(1
2
1
2
1

1
3
2
1
)

1
.
Since the marginal kurtosis is given by
k =
E(X
4
t
)
_
E(X
2
t
)

2
,
248 Petra Posedel
from the previous calculus it immediately follows that
k =
3(1 +
1
+
1
)(1
1

1
)
1
2
1
2
1

1
3
2
1
.
A little calculus shows
3Var(
2
t
) = E
_
X
4
t
_
3
_
E
_
X
2
t
_
2
=
3
2
0
(1 +
1
+
1
)
(1
1

1
)(1
2
1
2
1

1
3
2
1
)
3
_

0
1
1

1
_
2
=
3
2
0
(1
1

1
)
2

2
2
1
(1
2
1
2
1

1
3
2
1
)
. (2.16)
Since from the assumptions we have that
0
> 0, 1
1

1
> 0 and 1
2
1
2
1

3
2
1
< 1, it follows that all the factors in (2.16) are positive so we conclude that the
GARCH(1,1) process has the so-called leptokurtic distribution.
3 Estimation of the GARCH(1,1) model
Although in this section we assume that (Z
t
) are i.i.d. sequence of random variables,
the results we shall present can also be shown for the (Z
t
) strictly stationary and ergodic
sequence of random variables. In that case, the assumptions for the process (Z
t
) are little
modied but the main part of the calculus we present here also holds for not such strong
assumptions.
3.1 Description of the model and the quasi-likelihood function
Suppose we observe the sequence (Y
t
) such that
Y
t
= C
0
+
0t
, t = 1, . . . , n,
where we assume that (
0t
) is GARCH(1,1) process, exactly

0t
= Z
t

0t
, F
t
=
_
{
0s
, s t}
_
,
where (Z
t
) is a sequence of i.i.d. random variables and

2
0t
=
0
(1
0
) +
0

2
0t1
+
0

2
0t1
a.s. (3.1)
From Theorem 2.2 we have that the strict stationary solution of (3.1) is given by

2
0t
=
0
+
0

k=0

k
0

2
0t1k
a.s.
if it holds E
_
ln
_

0
+
0
Z
2
_
< 0. The process is described with the vector of parameters

0
=
_
C
0
,
0
,
0
,
0
_
.
Properties and Estimation of GARCH(1,1) Model 249
The model for the unknown parameters =
_
C, , , )

is given by
Y
t
= C +
t
, t = 1, . . . , n,
and

2
t
() = (1 ) +
2
t1
+
2
t1
(), t = 2, . . . , n
and with the initial condition
2
1
() = . With that kind of notation we have the following
expression for the process of conditional variance:

2
t
= +
t2

k=0

2
t1k
.
Let us dene the compact space
=
_
: C
l
C C
d
, 0 <
l

d
, 0 <
l

d
,
0 <
l

d
< 1
_

_
: E
_
ln
_
+Z
2
_
< 0
_
.
Additionally, assume that
0
so it immediately follows that
0
> 0 and
0
> 0.
Inference for GARCH(1,1) process usually assumes that (Z
t
) are i.i.d. random variables
such that Z
t
N(0, 1) so the likelihood function is easy to determine. Assuming that
the likelihood function is Gaussian, the log-likelihood function is of the form (ignoring
constants)
L
T
() =
1
2T
T

t=1
l
t
(), where l
t
() =
_
ln
2
t
() +

2
t

2
t
()
_
.
Since the likelihood function does not need to be Gaussian, in other words, the process
(Z
t
) does not need to be the Gaussian white noise, L
T
is called the quasi-likelihood
function.
3.2 Consistency of the quasi-maximum likelihood estimator
Although a nite data set is available in practice, this is not enough to determine good
properties of an estimator. We shall see in this section how useful results can be obtained
taking into consideration the strictly stationary model for the conditional variance that we
have previously dened. We shall note it in the following way

2
ut
() = +

k=0

2
t1k
,
t
= Y
t
C,
to avoid confusion with the original conditional variance process (
2
t
). In that case the
quasi-likelihood function is given by
L
uT
() =
1
2T
T

t=1
l
ut
(), where l
ut
() =
_
ln
2
ut
() +

2
t

2
ut
()
_
.
250 Petra Posedel
Additionally, we are going to show that the stationary and the non-stationary model
are not far away in some sense. So, all the calculus is done using the stationary model
and then connecting the two models.
Let us dene

2
t
() = +

k=0

2
0t1k
.
The process (
2
ut
) is a strictly stationary model of the conditional variance which assumes
an innite history of the observed data. The process (
2
t
) is in fact identical to the process
(
2
ut
) except that it is expressed as a function of the true innovations (
0t
) instead of the
residuals (
t
).
We suppose that the following conditions on the process (Z
t
) hold:
(1) (Z
t
) is a sequence of i.i.d. random variables such that EZ
t
= 0;
(2) Z
2
t
is nondegenerate;
(3) for some > 0 exists S

< such that E

_
Z
2+
t

< ;
(4) E
_
ln
_

0
+
0
Z
2
t
_
< 0;
(5)
0
is in the interior of ;
(6) if for some t holds

2
0t
= c
0
+

k=1
c
k

2
tk
i
2
0t
= c

0
+

k=1
c

2
tk
then c
i
= c

i
for every 1 i < .
We call the conditions (1) (6) elementary conditions.
The proof for the following result for the case of the general GARCH(q, p) process can
be found in [5].
Proposition 3.3 If the elementary conditions hold, there are not two different vectors
_
, , , C
_
and
_

, C

_
such that

2
0t
=

_
Y
t1
C

_
2
+

2
0t1
and

2
0t
= +
_
Y
t1
C
_
2
+
2
0t1
.
The following lemma would be very helpful for the results we shall provide. The
proof can be found in [10].
Lemma 3.4 Uniformly on
B
1

2
t
()
2
ut
() B
2
t
() a.s.
where
B = 1 + 2(1
d
)

1
2
(C
d
C
l
) max
_

l
, 1
_
+

d

l
(1
d
)
_
C
d
C
l
_
2
.
Properties and Estimation of GARCH(1,1) Model 251
Although we are not going to discuss the rational moments of the process
_

2
0t
_
, we will
still mention that, under the elementary conditions, there exists 0 < p < 1 such that
E
_

2
0t
_
p
< . (3.2)
The proof for such a result can be found in [13], Theorem 4.
The following lemma gives us the basic properties of the process
_

2
ut
_
and the likelihood
function (l
ut
).
Lemma 3.5 If the elementary conditions hold
(i) The process
_

2
ut
()
_
is strictly stationary and ergodic;
(ii) The process
_
l
ut
()
_
and the processes of its rst and second derivatives with re-
spect to are strictly stationary and ergodic for every in ;
(iii) For some 0 < p < 1 and for every it holds
E

2
ut
()

p
H
p
< .
Proof: The statement (1) follows from Theorem 2.3.
Since l
ut
() =
_
ln
2
ut
() +

2
t

2
ut
()
_
, and
l
ut

=
_

2
t

2
ut
1
_

2
ut
()

2
ut
()
, (3.3)
l
ut

=
_

2
t

2
ut
1
_

2
ut
()

2
ut
()
, (3.4)
l
ut
C
=
_

2
t

2
ut
1
_

2
ut
()
C
1

2
ut
()
2

t

2
ut
()
(3.5)
and
l
ut

=
_

2
t

2
ut
1
_

2
ut
()

2
ut
()
, (3.6)
where

2
ut

= 1 +

2
ut1

, (3.7)

2
ut

=
2
t1
+

2
ut1

, (3.8)

2
ut
C
= 2
t1
+

2
ut1
C
(3.9)
and

2
ut

=
2
ut1
+

2
ut1

, (3.10)
it follows that the process l
ut
() and processes of its rst and second derivatives are mea-
surable functions of strictly stationary and ergodic process (
t
) and so they are also strictly
252 Petra Posedel
stationary and ergodic. Finally, let 0 < p < 1 from (3.2). Then it follows from Lemma
3.4
E
_

2
ut
()
_
p
B
p
E
_

2
t
()
_
p
= B
p
E
_
+

k=0

2
0t1k
_
p
B
p
_

p
+
p

k=0

kp
E
_

2p
0tk1
_
_
.
Since
2
0t1k

1
0

2
0t
for every k, using (3.2) it follows
E
_

2
ut
_
p
B
p
_

p
+
p

k=0

kp
1

p
0
E
_

2p
0t
_
_
= B
p
_

p
+

p

p
0
E
_

2p
0t
_
1
1
p
_
B
p
_

p
d
+

p
d

p
0
E
_

2p
0t
_
1
1
p
d
_
H
p
< .

Some nontrivial calculus give us the following result.

Lemma 3.6 Under the elementary conditions it holds
sup

L
uT
() L
T
()

0 a.s. when T .
Finally, we want to nd additional constraints for the expression

2
0t

2
ut
and its in-
verse uniformly on . We will do so by splitting the parameter space. Let R
l
=
R(K
1
l

l
) < 1 where R() =
2 +P
2 +
< 1, for > 0, P = 1
_
1
2
2+

S
2

_
(0, 1)
and S

dene in the elementary conditions and K

l
=

d

0
+

d

0
< . Let
l
and
d
be
positive constants such that

l
<
0
_
1 R
1
12
l
_
and
d
<
0
_
1 R
1
12
0
_
,
where R
0
= R(
0
) < 1. For 1 r 12
2
dene constants

rl
=
0
R
1
r
l
+
l
<
0
and
rd
=

0

d
R
1
r
0
>
0
,
subspaces

r
l
= { :
rd

0
} and
r
d
= { :
0

rd
}
2
We will need r to be 12 in Lemma 4.2. Our aim is to nd the minimal r so that all the statements
presented bellow hold for every
r
.
Properties and Estimation of GARCH(1,1) Model 253
and
r
=
r
l

r
d
. The values
l
and
d
will depend on constants R
l
and R
0
which are
functions of the parameter space .
Observe that we can choose =
rmax

r
, for all 1 r 12. Now we are able
to present the result about the convergence in probability of the unconditional likelihood
process.
Lemma 3.7 Under the elementary conditions for every
1
it holds:
(1) E
_

2
t

2
ut
()
_
H
1

(C
d
C
l
)
2

l
+BH
c
where H
c
=

0

l
+

0

l
< .
In this case it holds
(2) L
uT
()
P
L() when T , where L() = E
_
l
ut
()
2
_
.
Proof: It is straightforward to show that
_
_

2
0t

2
t
()
_
_
r
H
c
. Hence, using Lemma 3.4 and
g = C
0
C we have the following
E
_

2
t

2
ut
_
= E
_
_

0t
+g
_
2

2
ut
_
= BE
_

2
0t

2
t
_
+ 2gE
_
1

2
ut
E
_

0t
|F
t1
_
_
+ E
_
g
2

2
ut
_
BE
_

2
0t

2
t
_
+
g
2

l
= BE
_

2
0t

2
t
_
+
g
2

l
B

2
0t

2
t

1
+
g
2

l
so
E
_

2
t

2
ut
_
BH
c
+
g
2

l
BH
c
+
_
C
d
C
l
_
2

l
H
1
that proves the rst statement. Additionally, we have
E

l
ut
()

= E

ln
_

2
ut
()
_
+

2
t

2
ut
()

ln
_

2
ut
()
_

+E
_

2
t

2
ut
()
_
.
254 Petra Posedel
But, for x 1 and 0 < p < 1 it holds the inequality ln x <
1
p
x
p
, so we have
E

ln
2
ut
()

ln
l

+E

ln

2
ut
()

ln
l

+
1
p
E
__

2
ut
()

l
_
p
_
=

ln
l

+
1
p
l
E
_

2p
ut
()

since

2
ut
()

l
1. Finally, using Lemma 3.5 we have
E

l
ut
()

< .
Since (l
ut
()) is strictly stationary and ergodic, it follows
L
uT
() =
1
2T
T

t=1
l
ut
()
P

1
2
E
_
l
ut
()

= L(),
1
.

The convergence in probability that we have presented in Lemma 3.7 is not a sufcient
condition for the consistency of the quasi-maximum likelihood estimator. It is necessary
that the convergence we have previously obtained holds uniformly. In order to obtain that,
it is sufcient to nd an upper bound for the score vector of the log-likelihood function
l
ut
() uniformly on . The details regarding the explicit forms of the upper bounds can
be found in [10].
Let
|A| =
_
tr(AA

)
_1
2
and A
r
=
_
E|A|
r
_1
r
be the Euclidean norm of a matrix or a vector and the L
r
norm of a random matrix or a
vector respectively.
Now we are going to present the local consistency of the quasi-maximum likelihood
estimator. Let us dene

T
= arg max

3
L
T
().

T
is the parameter value that maximizes the likelihood function on the set
3
.
Theorem 3.8 Under the elementary conditions

T
P

0
when T .
Properties and Estimation of GARCH(1,1) Model 255
4 Asymptotic normality of the quasi-maximumlikelihood
estimator
In this section we present the asymptotic distribution of the quasi-maximum likelihood
estimator (QMLE). In order to do so, we need stronger conditions on the process (Z
t
)
than the elementary conditions we have given in the previous section. In fact, we pretend
that the fourth moment of the random variable Z
t
is nite. We are going to call the
following condition additional condition.
E
_
Z
4
0
_
K < .
We do not present the proof for the following results as this would require long and non-
trivial calculus.
Lemma 4.1 Under the elementary conditions and under additional condition it holds
(i) E|l
ut
()l
ut
()

| < , for every

12
;
(ii)
1

T
T

t=1
l
t
(
0
)
D
N(0, A
0
), where A
0
= E
_
l
ut
(
0
)l
ut
(
0
)

_
.
Let
B
T
() =
1
T
T

t=1

2
l
t
() and B() = E
2
l
ut
().
Lemma 4.2 Suppose the elementary conditions and the additional condition to hold.
Then
(i) E sup

12

2
l
ut
()

< ;
(ii) For i = 1, 2, 3, 4, E sup

12

2
l
ut
()

< , where
i
is the i-th element of ;
(iii) sup

12
|B
T
() B()|
P
0 and B() is a continuous function on
12
.
The following result presents one of the classical results in asymptotic analysis and it
will be the basic tool for our further considerations. The details regarding the proof can
be found in [9, p. 185].
Theorem 4.3 Let (X
T
) be a sequence of random (m n) matrices and let (Y
T
) be a
sequence of random (n1) vectors such that X
T
P
C and Y
T
D
Y N(, ) when
T . Then the limiting distribution of (X
T
Y
T
) is the same as that of CY ; that is
X
T
Y
T
D
N(C, CC

) when T .
The following result assures that B
0
is a regular matrix.
256 Petra Posedel
Lemma 4.4 Suppose that the joint distribution of (
t
,
2
t
,
2
ut
) is nondegenerate. Then for
every the matrix
E
_

2
ut

2
ut

4
ut
_
is positive denite.
Finally, we have all the necessary results for studying the asymptotic behavior of the
parameter estimator. In fact, using the results presented above, the following theorem can
be proved.
Theorem 4.5 Suppose the elementary conditions and the additional condition to hold.
Then

T
_

T

0
_
D
N(0, V
0
),
where V
0
= B
1
0
A
0
B
1
0
, B
0
= B(
0
) = E
_

2
l
ut
(
0
)
_
and A
0
is dened in Lemma 4.1.
Notice that A
0
=
1
2
_
EZ
4
0
1
_
B
0
. So, in the case in which (Z
t
) is a sequence of random
variables such that Z
t
N(0, 1) we would have EZ
4
0
1 = 2 and A
0
= B
0
.
Let

B
T
= B
T
(

T
). In the case of maximum likelihood estimator,

B
T
would be the stan-
dard estimator of the covariance matrix. But in a more general case of quasi-maximum
likelihood estimator, the asymptotic covariance matrix is B
1
0
A
0
B
1
0
according to Theo-
rem 4.5. Since this is not equal to B
1
0
,

B
T
would not be a consistent estimator of that
value.
Let us dene
A
T
() =
1
T
T

t=1
l
t
()l
t
()

and

A
T
= A
T
(

T
) and A() = El
ut
()l
ut
()

.
The following result presents the consistency of the covariance matrix estimator.
Lemma 4.6 Suppose the elementary conditions and the additional condition to hold.
Then
(i) sup

12

A
T
() A()

P
0 and A() is continuous on
12
;
(ii)

V
T
=

B
1
T

A
T

B
1
T
P
B
1
0
A
0
B
1
0
.
Lemma 4.6 completes our characterization of classical properties of the QMLE for
GARCH(1, 1) model. We show that the covariance matrix estimator is consistent for the
asymptotic variance of the parameter estimator.
Properties and Estimation of GARCH(1,1) Model 257
References
[1] Amemiya, T. (1985): Advanced Econometrics. Cambridge: Harvard University
Press.
[2] Anderson, T.W. (1971): The Statistical Analysis of Time Series. New York: Wiley.
[3] Basrak, B., Davis R.A., and Mikosch, T. (2002): Regular variation of GARCH pro-
cesses. Stochastic. Process. Appl., 99, 95-115.
[4] Basrak, B., Davis R.A., and Mikosch, T. (2002): A characterization of multivariate
regular variation. Ann. Appl. Probab., 12, 908-920.
[5] Berkes, I., Horv ath, L., and Kokoszka, P. (2003): GARCH process: structure and
estimation. Bernoulli, 9, 201-227.
[6] Billingsley, P. (1968): Convergence of Probability Measures. New York: Wiley.
[7] Billinglsley, P. (1979): Probability and Measure. New York: Wiley.
[8] Embrechts, P., Kl uppelberg, C., and Mikosch, T. (1997): Modelling Extremal
Events. Berlin: Springer-Verlag.
[9] Hamilton, J.D. (1994): Time Series Analysis. Princeton: Princeton University Press.
[10] Lee, S.W. and Hansen, B.E. (1994): Asymptotic theory for the GARCH(1,1) quasi-
maximum likelihood estimator. Econometric Theory, 10, 29-52.
[11] Mikosch, T. amd St aric a, C. (1999): Change of structure in nancial time series,
long range dependence and the GARCH model. IWI Preprint, 5.
[12] Mikosch, T. and Straumann, D. (2003): Stable limits of martingale transforms with
application to the estimation of GARCH parameters. Working Paper, no. 189.
[13] Nelson, D.B. (1990): Stationarity and persistance in the GARCH(1,1) model.
Econometric Theory, 6, 318-334.
[14] Straumann, D. (2003): Estimation in Conditionally Heteroscedastic Time Series
Models, Ph.D. Thesis, University of Copenhagen.
[15] Williams, D. (1991): Probability with Martingales. Cambridge: Cambridge Univer-
sity Press.