Sei sulla pagina 1di 6

Statistica Sinica 8 (1998), 505-510

A NOTE ON THE STATIONARITY AND THE EXISTENCE OF MOMENTS OF THE GARCH MODEL

Min Chen and Hong Zhi An

Abstract: In the present paper we examine the strict stationarity and the existence of higher-order moments for the GARCH(p,q) model under general and tractable assumptions.

Key words and phrases: GARCH model, higher-order moments, nonlinear time series, strict stationarity,

1. Introduction

Consider the following non-linear time series model

x t = h t =

1

t h t α 0 + α 1 x t 1 + ··· + α p x tp + φ 1 h t 1 + ··· + φ q h tq ,

2

2

2

(1.1)

, q, { t } is a sequence of

independent identically distributed(i.i.d.) random variables with zero mean and unit variance, and t is independent of x ts ,s> 0. The model (1.1) is called the GARCH(p,q) model, which is proposed by Bollerslev (1986) and is one of many generalizations of the so-called ARCH (au- toregressive conditional heteroskedasticity) model proposed by Engle (1982) in the literature. The GARCH models have been widely applied in modelling mon- etary and ﬁnancial data such as inﬂation rate, interest rate and stock prices. The recent review by Bollerslev et al. (1992) contains an extensive literature on this subject. The strict stationarity and the existence of moments for a time series model are fundamental for statistical inference. Therefore, it is signiﬁcant to ﬁnd nec- essary and suﬃcient conditions for the strict stationarity and the existence of moments for a time series model. Bougeral and Picard (1992) gave a necessary and suﬃcient condition for the strict stationarity of the GARCH model. How- ever, as they pointed out, the conditions proposed are diﬃcult to verify and can only be checked by Monte Carlo methods. Bollerslev (1986) discussed conditions for the existence of higher-order moments for GARCH(1,1) model. So far it ap-

where α 0 > 0, α i 0, i = 1,

,p,

φ j

0, j = 1,

pears that there is not any paper about the existence of higher-order moments

506

MIN CHEN AND HONG ZHI AN

for the GARCH(p, q) model in the literature. The purpose of this paper is to give some suﬃcient conditions for the strict stationarity and the existence of moments for the GARCH(p, q) model.

2. The Strict Stationarity of the GARCH Model

First, we introduce some notation. Let {x t } conform to model (1.1). De-

,

ﬁne y t = x t 2 , η t = t 2 , Y t = (y t ,

0, α 0 , 0,

,y

tp +1 , h t ,

,h tq +1 ) τ , B t = (α 0 η t , 0,

,

0) τ , where “τ ” denotes the transposition of a matrix, and

A t =

α 1 η t ··· α p 1 η t α p η t φ 1 η t ··· φ q 1 η t φ q η t

00

1 ···

. .

. .

. .

0 ···

···

0 ···

α 1

. .

. .

. .

0 ···

0

.

1

α p 1

0

.

0

···

.

.

···

···

···

.

.

···

0

.

0

φ q 1

0

.

1

0

.

0

φ q

0

.

0

.

. .

00

α p

φ 1

01

.

. .

00

The expectation of the random matrix A t is deﬁned element-wise, hence it is obvious that E A t is a constant matrix, and

α 1 ··· α p 1 α p φ 1 ··· φ q 1 φ q

1

.

.

.

0

···

.

.

.

···

0

.

1

00 ···

.

00 ···

.

.

.

.

0

.

0

0

.

0

α 1 ··· α p 1 α p φ 1 ··· φ q 1 φ q

0

.

.

.

0

···

.

.

.

···

,0,

0

.

0

α 0 , 0,

01 ···

.

00 ···

.

.

.

.

0

.

1

0

.

0

E A t =

A.

Similarly, EB t = B (α 0 ,0,

,

0) τ . Then from (1.1) we have

x t 2 = t 2 α 0 +

p q

α i x ti 2 +

i=1

=1 φ j h tj

j

=

p q

α i t 2 x ti 2 +

i=1

j =1

φ j t 2 h tj + α 0 t 2 ,

which implies

y t =

p

q

α i η t y ti +

i=1 j =1

φ j η t h tj + α 0 η t .

(2.1)

(2.2)

STATIONARITY AND MOMENTS OF GARCH MODELS

507

Thus, {y t } is a solution of (2.2) if and only if {Y t } is a solution of the following stochastic diﬀerence equation

(2.3)

Y t = A t Y t 1 + B t .

Lemma 2.1. If p

q

i=1 α i + j

=1 φ j < 1, then the series of random vectors

=1 k1

k

j

=0 A tj B tk

converges almost surely. Furthermore if

Y t = B t +

=1 k1

k

j

=0 A tj B tk ,

(2.4)

then {Y t } is a strictly stationary, vector-valued process satisfying (2.3).

Proof. By the deﬁnition of A t and B t , it is easy to see that both {A t } and {B t } are sequence of independent, non-negative random vectors and A tj is independent of B tk for k = j. Therefore, we have

E k1

j =0

A tj B tk = k1

j

=0 E A tj EB tk = A k B.

It is easy to veriﬁed that the characteristic polynomial of A is given by

i=1 (α i + φ i )λ i , where m = max{p, q}, and α i = 0, for i>p,

φ i = 0, for i > q. Let ρ(A) be the spectral radius of the matrix A , hence ρ(A ) < 1

if and only if p

=1 φ j < 1, then (see

Horn and Johnson (1985))

det(λA I)=1 m

q

i=1 α i + j

=1 φ j < 1. Thus, if p

q

i=1 α i + j

A k < ,

k =1

which implies that

=1 E k1

k

j

=0 A tj B tk < ,

and hence

=1 k1

k

j

=0 A tj B tk < ,

a.s.

It is obvious that the vector-valued stochastic process {Y t } deﬁned by (2.4) is strictly stationary. Furthermore, we have

Y t = B t + A t B t1 +

k1

k =2 j

(

=0 A tj )B tk

508

MIN CHEN AND HONG ZHI AN

= B t + A t B t 1 +

= B t + A t Y t 1 .

l1

l =1

(

=0 A t1l )B t1l

j

Lemma 2.2. If (2.3) admits a strictly stationary solution with ﬁnite ﬁrst mo-

ment, then p

q

i=1 α i + j

=1 φ j < 1. Moreover, the strictly stationary solution of

(2.3) is unique.

Proof. By (2.3), we have

Y 0 = A 0 Y 1 + B 0

= B 0 +

A 0 B 1 + A 0 A 1 Y 2

.

.

.

= B 0 +

n1

(

k1

k =1 j =0

A

j )B k + (

n1

j =0

A

j )Y n .

(2.5)

Noting that all A n , B n and Y n are non-negative, {A t } is a sequence of

independent random matrices, A nj and B nk are independent for k = j, and

E Y 0 < . By taking expectation of each side of (2.5), it follows that

This shows that

Therefore,

Let {δ i , i = 1, δ i,p + q ) τ , where δ ij

E Y 0

n1 E k1

k

=1

j

=0 A j B k =

n

k =1

A k B < .

n1

k

=1

A k B.

,p

lim A n B = 0.

n

+ q} be the canonical basis of R p + q , i.e.

(2.6)

,

δ i = (δ i, 1 ,

= 0, for i

= j, δ ii = 1. If we can prove that

for 1 i p + q,

lim

n

A n δ i = 0,

(2.7)

then (2.7) implies that lim n A n = 0, which again implies that ρ(A ) < 1. As

we showed before the later is equivalent to p

to the ﬁrst part of this lemma. In fact, since B = α 0 (δ 1 +δ p +1 ) and 0 < α 0 < , by (2.6) and the deﬁnition of matrix A , (2.7) holds for i = 1 and i = p + 1. Again by the deﬁnition of A and B, Aδ p + q = φ q (δ 1 + δ p +1 ). If φ q = 0, then A δ p + q = 0, hence (2.7) holds. If φ q > 0, by the above equalities,

q

i=1 α i + j

=1 φ j < 1, which leads

lim A n δ p + q = lim A n 1 φ q (δ 1 + δ p +1 )=0.

n

n

STATIONARITY AND MOMENTS OF GARCH MODELS

509

It is easy to see that for 2 i < p,

A δ i = α i δ 1 + δ i+1 + α i δ p +1 .

Since (2.7) holds for i = 1 and i = p + 1, by a backward recursion (2.7) holds for

i = p, p 1,

A δ i = φ i (δ 1 + δ p +1 ) + δ i+1 .

Noting that (2.7) holds for i = p + q, by a backward recursion, (2.7) holds for

i = p + q 1, p + q 2,

For the proof of the uniqueness, let {U t } be another strictly stationary solu- tion satisfying (2.3). Then {U t } also satisﬁes an equation similar to (2.5). Note that the third term on the right hand side of the last line in (2.5) goes to zero in

probability. Then the uniqueness follows immediately.

Theorem 2.1. The GARCH (p, q) model (1.1) admits a strictly stationary solu-

=1 φ j < 1. Moreover, this

tion with ﬁnite variance if and only if p strictly stationary solution is also unique.

,

1, respectively. Similarly, for p + 1 <i<p + q,

,p

+ 2. Finally, (2.7) holds for any i = 1, 2,

,p

+ q.

q

i=1 α i + j

 Proof. The desired resultis obtained by combining Lemma 2.1 and Lemma 2.2.

3. The Existence of Higher-Order Moments

Let

L s = {x : x s = E 1 s |x| s < ,s> 0},

where x is a random variable.We need the Kronecker product (), the direct operations (“vec” operations ), the notation A m = A A ⊗···⊗ A , and the basic identity vec(ABC)=(C τ A)vec(B ). We denote the (i, j )th element of a matrix D by (D) ij and deﬁne Σ m = E(A t m ).

Theorem 3.1. Let {x t } be speciﬁed to be the strictly stationary solution of model

(1.1) and p

q

i=1 α i + j

=1 φ j < 1.

(i)

If E| t | 4

< and ρ(Σ ) < 1, then |x | 2 L 2 .

2

t

(ii) If E| t | 4(s 1) < , for some integer s > 2, and ρs ) < 1, then |x t | 2 L s .

˜

˜

Proof. Let Y t be generated according to (2.3) with the starting value Y 0 = 0.

Let Y be a random variable having the same distribution as that of (2.4) with

t = 0, which is the (marginal) distribution of the unique stationary solution of

˜

(2.3). It is clear that Y t Y in distribution. Let ψ(Y ) be a random variable.

From weak convergence theory (Billingsley (1968)), it is known that to show

˜

˜

(Y ) < , it suﬃces to show lim inf ( Y t ) < . Let V (t) = E( Y t ). Then,

taking expectation on both sides of (2.3), we get

V (t) = AV (t 1) + B.

510

MIN CHEN AND HONG ZHI AN

It is well know that lim V (t) exists and is ﬁnite if the spectral radius of A is less than 1(cf. Subba Rao (1981)). Hence if the spectral radius of A is less than 1,

{V (t)} is

bounded. Next, let V 1 (t) = E(vec(

˜

Y t

Y

t

˜

τ )). Then

V 1 (t) = E(A t A t )V 1 (t 1) + E(B t A t )V 1 (t 1)

+E(A t B t )V 1 (t 1) + vec(E(B t B

τ

t

)).

))

are constant and ﬁnite matrices. As {V (t)} is bounded, it is clear that lim V 1 (t) exists and is ﬁnite if the spectral radius of Σ 2 = E(A t A t ) is less than 1. Note

that the ﬁrst element of V 1 (t) is Ex

converges in distribution to

the proof of (i). The proof

x 4 and lim V 1 (t) is ﬁnite, E(x 4 ) < . This completes of (ii) is similar and hence omitted.

Note that the matices E(A t A t ), E(B t A t ), E(A t B t ) and vec(E(B t B

0

4

t

). Because x˜

4

t

τ

t

0

Acknowledgement

We are very grateful to referees for their careful reading and valuable sug- gestions which have enabled us to improve the earlier version of this paper. The work was supported by NNCF of China and Prob. Lab. Inst. Applied Math. Academia Sinica.

References

Billingsley, P. (1968). Convergence of Probability Measure. Wiley, New York. Bollerslev, T. (1986). Generalized autoregressive conditional heteroscedasticity. J. Economet- rics 31, 307-327. Bollerslev, T., Chou, R. Y. and Kroner, K. F. (1992). ARCH modeling in ﬁnance. J. Econo- metrics 52, 5-59. Bougeral, P. and Picard, N. (1992). Stationarity of GARCH processes and of some nonnegative time series. J. Econometrics 52, 115-127. Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the variance of U. K. inﬂation. Econometrica 50, 987-1007. Gu´egan, D. and Diebolt, J. (1994). Probabilistic properties of the β -ARCH model. Statist. Sinica 4, 71-87. Horn, R. A. and Johnson, C. R. (1985). Matrix Analysis. Cambridge Univ. Press. Subba Rao, T. ( 1981). On the theory of bilin ear time series models. J. Roy. Statist. Soc. Ser. B 43, 244-255.

Institute of Applied Mathematics, Academia Sinica, Beijing, 100080.

E-mail: mchen@iss01.iss.ac.cn

E-mail: hzan@sun.ihep.ac.cn

(Received October 1995; accetped March 1997)