Econometric Theory and Methods

Answers to Starred Exercises

140

Solution to Exercise 13.21

13.21 Consider the regression model y = X + u, where X is an n k matrix, in which the errors follow a GARCH(1, 1) process with conditional variance given 2 by equation (13.78). Show that the skedastic function t ( , ) used in the loglikelihood contribution t ( , ) given in (13.86) can be written explicitly as
2 t ( , ) t1 0 (1 1 ) = + 1 1 1 t1 s1 2 uts + 1 s=1 t1 0 1 , 1 1 1

(13.96)

where ut stands for the residual yt Xt , and all unavailable instances of both 2 u2 t and t are replaced by the unconditional expectation 0 /(1 1 1 ). Then show that the rst-order partial derivatives of as follows:
2 2 (u2 t ) t t ut t t Xt ut = + = 1 t 2 2 4 ut t t t t ( , ) t1 s 1 1 Xts uts s=1

can be written

t = 0

2 t t 2 0 t

u2 t

2 t 4 2t

t 1 1 1

1 1
t1

t 1 1

1 1 1

(13.97)

2 2 t t t u2 t = = t 2 1 4 1 t 2t

s 1 2 1 uts + s=1 t2

t 1 0 1 , (1 1 1 )2 t 1 0 (1 1 ) (1 1 )2

2 2 0 (t 1)1 t t u2 t t = = t 2 4 1 1 1 1 t 2t t1

+ 1
s=1

s2 2 (s 1)1 ut s +

t1 t 2 0 1 0 (t 1)1 + 1 1 1 (1 1 1 )2

The conditional variance of a GARCH(1,1) process is a recursive equation that was given in equation (13.78):
2 2 t = 0 + 1 u2 t1 + 1 t1 .

(13.78)

If this recursion is written as

2 2 2 t 1 t 1 = 0 + 1 ut1 ,

then it has the same algebraic form as the recursion (7.29) that denes an AR(1) process. It can thus be solved in the same way. This gives
t2 2 t = s=0 t1 2 s 1 (0 + 1 u2 ts1 ) + 1 1 .

Copyright c 2003, Russell Davidson and James G. MacKinnon

Econometric Theory and Methods

Answers to Starred Exercises

141

Similar results may be found in equations (7.30) and (7.95). By performing the summation of the rst term explicitly and changing the summation index of the second, we see that this equation can be rewritten as
2 t = t 1 0 (1 1 ) s1 2 t1 2 + 1 1 uts + 1 1 . 1 1 s=1 t1

(S13.28)

2 It is straightforward to check that the series t given by this equation does 2 indeed satisfy the recursion (13.78). If we now replace the unavailable 1 in equation (S13.28) by 0 /(1 1 1 ), we obtain equation (13.96), which is what we set out to do.

It follows directly from expression (13.86) for the contribution loglikelihood function that t ut = 2 ut t and
2 t 1 u2 u2 t t t = + = . 2 2 4 4 t 2t 2t 2t

t ( , )

to the

(S13.29)

In order to establish equations (13.97), we must calculate the derivatives of 2 ( , ) with respect to the parameters , 0 , 1 , ut ( ) yt Xt and t and 1 . It is easy to see that ut / = Xt and that the partial derivatives 2 , we obtain from of ut with respect to the other parameters are zero. For t equation (13.96) that
2 t s1 = 21 1 Xts uts , s=1 t 1 t1 2 t 1 1 1 = + , 0 1 1 1 1 1 2 t = 1 t 1 s 1 2 1 uts + s=1 t 1 0 1 , (1 1 1 )2 t1

and

(S13.30)

t 2 t1 2 t 0 (t 1)1 0 (1 1 ) = + 2 1 1 1 (1 1 ) t1

+ 1
s=1

(s 1) s2 u2 ts +

t1 t2 0 (t 1)1 0 1 + . 1 1 1 (1 1 1 )2

Equations (13.97) follow directly from equations (S13.29) and (S13.30).

Copyright c 2003, Russell Davidson and James G. MacKinnon