ANALYSIS OF FINANCIAL TIME SERIES: Exercise Sheet 2

1. Considering T observations of a time series y1 , . . . , yT as a random

sample of a random variable Y , yields for the likelihood function the
joint distribution of the sample L() = f (y1 , . . . , yT |), where the yet
unknown parameter vector is to be determined by maximazation of
L. This approach is known as estimation by maximum likelihood.
Using the chain rule for conditional probabilities (which applies to prob-
ability densities as well), the joint density may be written as

f (y1 , . . . , yT |) = f (yT |FT 1 ; )f (yT 1 |FT 2 ; ) f (y2 |F1 ; )f (y1 ; ),

where the information set Ft collects all available information up to

time t, e.g. Ft = {y1 , . . . , yt } for a univariate time series. The log-
likelihood function is obtained as the natural logarithm of the likelihood
function:
3T
() = log L() = log f (yt |Ft1 ; )
t=1

Applying this approach to the observations yt = xt + ut with ARCH

residuals ut |Ft1 N (0, ht ) such that yt |Ft1 N (xt , ht ), yields the
log-likelihood function given in the lecture notes. More generally, the
residuals of an ARCH process may be defined to follow some other sym-
metric distributions, which are not necessarily normal. As an example,
consider the scaled Student t-distribution with degrees of freedom,

mean , variance 2 = H(2) , and probability density function

(( + 1)/2) /2 J o +1
g(y) = + H(y )2 2
H,
(/2)

where () denotes the gamma function.

a) Show that the log-likelihood function of an ARCH specification

yt = xt + ut with ut |Ft1 scaled Student t( = 0, t2 = Ht (2) )
may be written as:
3 T F w W } ]k
1 ( 2)(/2)2 1 2 +1 (yt xt )2
() = log log t log 1 + 2
t=1
2 (( + 1)/2)2 2 2 t ( 2)

Hint: yt |Ft1 scaled Student t( = xt , t2 = Ht (2)
).
 b) Show that the likelihood function of an ARCH model with con-
ditionally Student t distributed residuals as above converges to
the likelihood function of an ARCH model with conditionally nor-
mally distributed residuals for .
Hint: D i
z + 14 z p x Qr r x
D i z, 1 + e
z 14 r

2. Consider the GARCH(1,1) model ht = + u2t1 + ht1 with > 0,

> 0, 0, and + < 1.
a) Show that

u2 := var(ut ) = .
1
Hint: Show first that n + 1 applications of the law of iterated
expectations yield an expression of the form
var(ut ) = (1 + ( + ) + . . . + ( + )n ) + ( + )n+1 E(utn1 2 )
and take the limit n .
Alternatively you may wish to write down the ARMA representa-
tion of the model and use the stationarity of u2t .
b) Show that
1 ( + )2
E(u4t ) = 3u4
1 ( + )2 22
provided that ( + )2 + 22 < 1.
Hint: Show first that n + 1 applications of the law of iterated
expectations yield an expression of the form

1 + ( + ) D i
E(u4t ) =3 2 1 + (32 + 2 + 2 ) + . . . + (32 + 2 + 2 )n
1 ( + )
+ (32 + 2 + 2 )n+1 E(utn1 4 )
and take the limit n .
c) Show that the k-step ahead forecast of the variance is
Et (u2t+k ) = u2 + ( + )k1 (ht+1 u2 ),

which implies that long horizon forecasts converge to the

unconditional variance u2 = /(1 ).
Hint: Use the law of iterated expectations.