Arch Uni Slides

GARCH Models
Eduardo Rossi
University of Pavia
Empirical regularities
c
Eduardo Rossi
Econometria mercati finanziari (avanzato)
c
Eduardo Rossi
GARCH models have been developed to account for empirical

regularities in financial data. Many financial time series have a
number of characteristics in common.
Asset prices are generally non stationary. Returns are usually
stationary. Some financial time series are fractionally integrated.
Return series usually show no or little autocorrelation.
Serial independence between the squared values of the series is
often rejected pointing towards the existence of non-linear
relationships between subsequent observations.
c
Eduardo Rossi
Volatility of the return series appears to be clustered.

Normality has to be rejected in favor of some thick-tailed
distribution.
Some series exhibit so-called leverage effect, that is changes in
stock prices tend to be negatively correlated with changes in
volatility. A firm with debt and equity outstanding typically
becomes more highly leveraged when the value of the firm
falls.This raises equity returns volatility if returns are constant.
Black, however, argued that the response of stock volatility to the
direction of returns is too large to be explained by leverage alone.
Volatilities of different securities very often move together.
c
Eduardo Rossi
Why do we need ARCH models?
Wolds decomposition theorem establishes that any covariance

stationary {yt } may be written as the sum of a linearly deterministic
component and a linearly stochastic with a square-summable,
one-sided moving average representation. We can write,
yt = dt + ut
dt is linearly deterministic and ut is a linearly regular covariance
stationary stochastic process, given by
ut = B (L) t
B (L) =
X
i=0
c
Eduardo Rossi
bi Li
X
b2i <
b0 = 1
i=0
E [t ] = 0
2 < , if t =
E [t ] =
0, otherwise
The uncorrelated innovation sequence need not to be Gaussian and

therefore need not be independent. Non-independent innovations are
characteristic of non-linear time series in general and conditionally
heteroskedastic time series in particular.
c
Eduardo Rossi

Now suppose that yt is a linear covariance stationary process with
i.i.d. innovations as opposed to merely white noise. The
unconditional mean and variance are
E [yt ] = 0
X
2
E yt = 2 b2i
i=0
which are both invariant in time. The conditional mean is time

varying and is given by
X
E [yt |t1 ] =
bi ti
i=1
where the information set is t1 = {t1 , t2 , . . .}.
c
Eduardo Rossi

This model is unable to capture the conditional variance dynamics.
In fact, the conditional variance of yt is constant at
h
i
2
E (yt E [yt |t1 ]) |t1 = 2 .
This restriction manifests itself in the properties of the k-step-ahead

conditional prediction error variance. The k -step-ahead conditional
prediction is
E [yt+k |t ] =
bk+i ti
i=0
and the associated prediction error is

yt+k E [yt+k |t ] =
k1
X
bi t+ki
i=0
c
Eduardo Rossi
which has a conditional prediction error variance

k1
h
i
X
2
2
E (yt+k E [yt+k |t ]) |t =
b2i
i=0
k1
h
i
X
X
2
E (yt+k E [yt+k |t ]) |t = 2
b2i 2
b2i , as k
i=0
i=0
For any k, the conditional prediction error variance depends only on

k and not on t .
In conclusion, the simple i.i.d. innovations model is
unable to take into account the relevant information which
is available at time t.
c
Eduardo Rossi
10
The ARCH(q) Model
(Bollerslev, Engle and Nelson, 1994) The process {t (0 )} follows an

ARCH (AutoRegressive Conditional Heteroskedasticity) model if
Et1 [t (0 )] = 0
t = 1, 2, . . .
(1)
and the conditional variance

t2
(0 ) V art1 [t (0 )] =
Et1 2t
(0 )
t = 1, 2, . . .
(2)
depends non trivially on the -field generated by the past

observations: {t1 (0 ) , t2 (0 ) , . . .} .
c
Eduardo Rossi
11
The ARCH(q) Model
Let {yt (0 )} denote the stochastic process of interest with

conditional mean
t (0 ) Et1 (yt )
t = 1, 2, . . .
(3)
t (0 ) and t2 (0 ) are measurable with respect to the time t 1

information set. Define the {t (0 )} process by
t (0 ) yt t (0 ) .
c
Eduardo Rossi
(4)
12
The ARCH(q) Model
It follows from eq.(1) and (2), that the standardized process

zt (0 ) t (0 ) t2 (0 )1/2
t = 1, 2, . . .
(5)
with
Et1 [zt (0 )] = 0
V art1 [zt (0 )] = 1
will have conditional mean zero (Et1 [zt (0 )] = 0) and a time
invariant conditional variance of unity.
c
Eduardo Rossi
13
The ARCH(q) Model
We can think of t (0 ) as generated by

t (0 ) = zt (0 ) t2 (0 )
1/2
where 2t (0 ) is unbiased estimator of t2 (0 ).

Lets suppose zt (0 ) N ID (0, 1) and independent of t2 (0 )
2
2
2
2
Et1 t = Et1 t Et1 zt = Et1 t = t2
because zt2 |t1 2(1) .
c
Eduardo Rossi
14
The ARCH(q) Model
If the conditional distribution of zt is time invariant with a finite

fourth moment, the fourth moment of t is
4
4 4
4 2 2
4 2 2
E t = E zt E t E zt E t = E zt E t
where the last equality follows from
E[t2 ] = E[Et1 (2t )] = E[2t ]
4
4 2 2
E t E zt E t
by Jensens inequality. The equality holds true for a constant

conditional variance only.
c
Eduardo Rossi
15
The ARCH(q) Model
4
If zt N ID (0, 1), then E zt = 3, the unconditional distribution for
t is therefore leptokurtic
4
2 2
E t
3E t
4
2 2
3
E t /E t
The kurtosis can be expressed as a function of the variability of the

conditional variance.
c
Eduardo Rossi
16
The ARCH(q) Model

0, t2
In fact, if t |t1 N
4
2 2
Et1 t
= 3Et1 t
h
4
i

2

2 2
2
2 2
3 E Et1 t
= 3 E t
E t
= 3E Et1 t
4

2 2
E t 3 E t
4
E t
c
Eduardo Rossi
2 2 o

2
2
3E Et1 t
3 E Et1 t
n
o

2
2
2
2
2
2
3 E t
+ 3E Et1 t
3 E Et1 t
n
2 2 o
2
4
2
E Et1 t
E Et1 t
E t
2 =3+3
2
2
[E (t )]
[E (2t )]
2

2
V ar Et1 t
V ar t
3+3
=3+3
2
2 .
2
2
[E (t )]
[E (t )]
17
The ARCH(q) Model

Another important property of the ARCH process is that the process
is conditionally serially uncorrelated. Given that
Et1 [t ] = 0
we have that with the Law of Iterated Expectations:
Eth [t ] = Eth [Et1 (t )] = Eth [0] = 0.
This orthogonality property implies that the {t } process is
conditionally uncorrelated:
Covth [t , t+k ] = Eth [t t+k ] Eth [t ] Eth [t+k ] =
= Eth [t t+k ] = Eth [Et+k1 (t t+k )] =
= Eth [t Et+k1 [t+k ]] = 0
The ARCH model has showed to be particularly useful in modeling
the temporal dependencies in asset returns.
c
Eduardo Rossi
18
The ARCH(q) Model

The ARCH (q) model introduced by Engle (Engle (1982)) is a linear
function of past squared disturbances:
t2 = +
q
X
i 2ti
(6)
i=1
In this model to assure a positive conditional variance the parameters

have to satisfy the following constraints: > 0 e
1 0, 2 0, . . . , q 0. Defining
t2 2t vt
where Et1 (vt ) = 0 we can write (6) as an AR(q) in 2t :
2t = + (L) 2t + vt
where (L) = 1 L + 2 L2 + . . . + q Lq .
c
Eduardo Rossi
19
The ARCH(q) Model
(1 (L))2t = + vt
The process is weakly stationary if and only if
q
P
i < 1; in this case
i=1
the unconditional variance is given by

2
E t = / (1 1 . . . q ) .
c
Eduardo Rossi
(7)
20
The ARCH(q) Model
The process is characterised by leptokurtosis in excess with respect

to the normal distribution. In the case, for example, of ARCH(1)

2
with t |t1 N 0, t , the kurtosis is equal to:
E
4t
/E

2 2
t
=3 1
12
/ 1
with 312 < 1, when 312 = 1 we have

4
2 2
E t /E t = .
312
(8)
In both cases we obtain a kurtosis coefficient greater than 3,

characteristic of the normal distribution.
c
Eduardo Rossi
21
The ARCH(q) Model
The result is readily obtained:

4
4
E t = 3E t

2

4
2
4
2
E t = 3 + 1 E t1 + 21 E t1
E
4t
=
=
c
Eduardo Rossi
2t1
3 + 21 E
(1 312 )
2

2
3 + 21
(1 312 )
22
The ARCH(q) Model
substituting 2 = / (1 1 ):
2

2
2

3
(1
)
+
2
3
(1 + 1 )
1
1
4
E t =
=
(1 312 ) (1 1 )
(1 312 ) (1 1 )
finally
4t
c
Eduardo Rossi
/E

2 2
t
3 (1 + 1 ) (1 1 )
=
2
2
(1 31 ) (1 1 )
3 1 12
1 312
(9)
23
The ARCH Regression Model
We have an ARCH regression model when the disturbances in a

linear regression model follow an ARCH process:
yt = xt b + t
Et1 (t ) = 0
0, t2
t |t1 N

2
Et1 t t2 = + (L) 2t
c
Eduardo Rossi
24
The GARCH(p,q) Model
In order to model in a parsimonious way the conditional

heteroskedasticity, Bollerslev (1986) proposed the Generalised ARCH
model, i.e GARCH(p,q):
t2 = + (L) 2t + (L) t2 .
(10)
where (L) = 1 L + . . . + q Lq , (L) = 1 L + . . . + p Lp .

The GARCH(1,1) is the most popular model in the empirical
literature:
2
.
t2 = + 1 2t1 + 1 t1
c
Eduardo Rossi
(11)
25
To ensure that the conditional variance is well defined in a

GARCH (p,q) model all the coefficients in the corresponding linear
ARCH () should be positive:
t2
!
1
q
p
X
X
+
i Li
j 2tj
1
i=1
= +
j=1
k 2tk1
(12)
k=0
t2 0 if 0 and all k 0.
The non-negativity of and k is also a necessary condition for the
non negativity of t2 .
c
Eduardo Rossi
26
In order to make e {k }k=0 well defined, assume that :

i. the roots of the polynomial (x) = 1 lie outside the unit circle,
and that 0, this is a condition for to be finite and positive.
ii. (x) e 1 (x) have no common roots.
c
Eduardo Rossi
27
These conditions are establishing nor that t2 neither that

2
t t= is strictly stationary. For the simple GARCH(1,1) almost
sure positivity of t2 requires, with the conditions (i) and (ii), that
(Nelson and Cao, 1992),
c
Eduardo Rossi
(13)
28
For the GARCH(1,q) and GARCH(2,q) models these constraints can

be relaxed, e.g. in the GARCH(1,2) model the necessary and
sufficient conditions become:
0 1 < 1
1 1 + 2
c
Eduardo Rossi
(14)
29

For the GARCH(2,1) model the conditions are:
1 + 2
< 1
12 + 42
(15)
These constraints are less stringent than those proposed by Bollerslev

(1986):
i = 1, . . . , p
j = 1, . . . , q
c
Eduardo Rossi
(16)
30
These results cannot be adopted in the multivariate case, where the

2
requirement of positivity for t means the positive definiteness for
the conditional variance-covariance matrix.
c
Eduardo Rossi
31
The Yule-Walker equations for the squared process
A GARCH(p,q) can be represented as an ARMA process, given that

2
2
2
t = t + t , where Et1 [t ] = 0, t t , :
max(p,q)
2t
=+
X
j=1
(j + j ) 2tj +
p
X
i=1
i ti
2t ARMA(m,p) with m = max(p, q). We can apply the classical

results of ARMA model.
c
Eduardo Rossi
32
Example: GARCH(1,1)
2
t2 = + 1 2t1 + 1 t1
replacing t2 with 2t t , we obtain

2t t = + 1 2t1 + 1 (2t1 t1 )
hence
2t = + (1 + 1 )2t1 + t 1 t1
c
Eduardo Rossi
33
We can study the autocovariance function, that is:

2
2 2
(k) = cov t , tk
!
p
m
X
X
2 (k) = cov +
(j + j ) 2tj + t
i ti , 2tk
j=1
2 (k) =
m
X
j=1
i=1
"
#
p
X

2
2
(j + j ) cov tj , tk +cov t
i ti , 2tk
i=1
(17)
When k is big enough, the last term on the right of expression (17) is
null.
c
Eduardo Rossi
34
The sequence of autocovariances satisfy a linear difference equation

of order max (p, q), for k p + 1
m
X
2 (k) =
(j + j ) 2 (k j)
j=1
This system can be used to identify the lag order m and p, that is the
p and q order if q p, the order p if q < p.
c
Eduardo Rossi
35
The GARCH Regression Model
1, 2t1 ,
2
, 2tq , t1
,

2
, tp ,
Let wt =
= (, 1 , , q , 1 , , p ) and , where = (b , ) and is
a compact subspace of a Euclidean space such that t possesses finite
second moments. We may write the GARCH regression model as:
t = yt xt b
t |t1 N
t2 = wt
c
Eduardo Rossi
0, t2
36
Stationarity
The process {t } which follows a GARCH(p,q) model is a martingale

difference sequence. In order to study second-order stationarity its
sufficient to consider that:
2
V ar [t ] = V ar [Et1 (t )] + E [V art1 (t )] = E t
and show that is asymptotically constant in time (it does not depend
upon time).
A process {t } which satisfies a GARCH(p,q) model with positive
coefficient 0, i 0 i = 1, . . . , q, i 0 i = 1, . . . , p is covariance
stationary if and only if:
(1) + (1) < 1
c
Eduardo Rossi
37
Stationarity
This is a sufficient but non necessary conditions for strict

stationarity. Because ARCH processes are thick tailed, the conditions
for covariance stationarity are often more stringent than the
conditions for strict stationarity.
A GARCH(1,1) model can be written as
"
#
k
XY

2
2
t = 1 +
1 + 1 zti
k=1i=1
c
Eduardo Rossi
38
Stationarity
In fact,
t2
=+
2
t1
2
t2
c
Eduardo Rossi
2
t1
2
1 zt1
+ 1
=+
2
t2
2
1 zt2
=+
2
t3
2
1 zt3
+ 1
+ 1

39
Stationarity
t2
= + +
2
1 zt2
"
= 1+
c
Eduardo Rossi
Y
k
X
k=1 i=1
2
1 zti
+ 1
2
1 zt1
+ 1

2
2
2
2
= + 1 zt1 + 1 + t2 1 zt2 + 1 1 zt1 + 1

2
2
2
= + 1 zt1 + 1 + 1 zt1 + 1 1 zt2 + 1

2
2
2
2
+t3 1 zt3 + 1 1 zt2 + 1 1 zt1 + 1
finally,
t2
2
t2
+ 1
40
Stationarity
t2

2
Nelson shows that when > 0,
< a.s. and t , t is strictly

2
stationary if and only if E ln 1 + 1 zt < 0

2
2
E ln 1 + 1 zt ln E 1 + 1 zt = ln (1 + 1 )
when 1 + 1 = 1 the model is strictly stationary.

2
E ln 1 + 1 zt < 0 is a weaker requirement than 1 + 1 < 1.
c
Eduardo Rossi
41
Stationarity
Example
ARCH(1), with 1 = 1, 1 = 0, zt N (0, 1)

2
2
E ln zt ln E zt = ln (1)
Its strictly but not covariance stationary. The ARCH(q) is

covariance stationary if and only if the sum of the positive
parameters is less than one.
c
Eduardo Rossi
42
Forecasting volatility
Forecasting with a GARCH(p,q) (Engle and Bollerslev 1986):

2
t+k
=+
q
X
i 2t+ki +
i=1
p
X
2
i t+ki
i=1
we can write the process in two parts, before and after time t:
2
t+k
= +
n
X

i=1
i 2t+ki
2
i t+ki
m
X
i 2t+ki
2
i t+ki
i=k
where n = min {m, k 1} and by definition summation from 1 to 0

and from k > m to m both are equal to zero.
c
Eduardo Rossi
43
Thus
2
Et t+k
= +
c
Eduardo Rossi
n
X

i=1
m
X

2

2
2
(i + i ) Et t+ki +
i t+ki + i t+ki .
i=k
44
In particular for a GARCH(1,1) and k > 2:

2
Et t+k
k2
X
2
(1 + 1 )i + (1 + 1 )k1 t+1
i=0
1 (1 + 1 )
k1
2
+ (1 + 1 )k1 t+1
[1 (1 + 1 )]
h
i
k1
k1 2
= 2 1 (1 + 1 )
+ (1 + 1 )
t+1

k1 2
2
2
= + (1 + 1 )
t+1
2
When the process is covariance stationary, it follows that Et t+k
converges to 2 as k .
c
Eduardo Rossi
45
The IGARCH(p,q) model
The GARCH(p,q) process characterized by the first two conditional

moments:
Et1 [t ] = 0
q
p
X
X

2
t2 Et1 2t = +
i 2ti +
i ti
i=1
i=1
where 0, i 0 and i 0 for all i and the polynomial

1 (x) (x) = 0
has d > 0 unit root(s) and max {p, q} d root(s) outside the unit
circle is said to be:
c
Eduardo Rossi
46
Integrated in variance of order d if = 0

Integrated in variance of order d with trend if > 0.
The Integrated GARCH(p,q) models, both with or without trend, are
therefore part of a wider class of models with a property called
persistent variance in which the current information remains
important for the forecasts of the conditional variances for all horizon.
c
Eduardo Rossi
47

So we have the Integrated GARCH(p,q) model when (necessary
condition)
(1) + (1) = 1
To illustrate consider the IGARCH(1,1) which is characterised by
1 + 1 = 1
t2
t2
2
= + 1 2t1 + (1 1 ) t1
= +
2
t1
1 2t1
2
t1
0 < 1 1
For this particular model the conditional variance k steps in the

future is:
2
2
Et t+k = (k 1) + t+1
c
Eduardo Rossi
48
Persistence
In many studies of the time series behavior of asset volatility the

question has been how long shocks to conditional variance persist.
If volatility shocks persist indefinitely, they may move the whole
term structure of risk premia.
There are many notions of convergence in the probability theory
(almost sure, in probability, in Lp ), so whether a shock is
transitory or persistent may depend on the definition of
convergence.
c
Eduardo Rossi
49
Persistence
In linear models it typically makes no difference which of the

standard definitions we use, since the definitions usually agree.
In GARCH models the situation is more complicated.
In the IGARCH(1,1):
2
t2 = + 1 2t1 + 1 t1
where 1 + 1 = 1. Given that 2t = zt2 t2 , we can rewrite the

IGARCH(1,1) process as

2
2
2
t = + t1 (1 1 ) + 1 zt1
0 < 1 1.
When = 0, t2 is a martingale.
c
Eduardo Rossi
50
Persistence
Based on the nature of persistence in linear models, it seems that

IGARCH(1,1) with > 0 and = 0 are analogous to random walks
with and without drift, respectively, and are therefore natural models
of persistent shocks.
This turns out to be misleading, however:
in IGARCH(1,1) with = 0, t2 collapses to zero almost
surely
in IGARCH(1,1) with > 0, t2 is strictly stationary and ergodic
and therefore does not behave like a random walk, since random
walks diverge almost surely.
c
Eduardo Rossi
51
Persistence
Two notions of persistence.

t2
t2
1. Suppose
is strictly stationary and ergodic. Let F
be the

unconditional cdf for t2 , and Fs t2 the conditional cdf for t2 ,

2
2
given information at time s < t. For any s F t Fs t 0
at all continuity points as t . There is no persistence when
2
t is stationary and ergodic.
2. Persistence is defined in terms of forecast moments. For some
2
>0, the
shocks
to
fail to persist if and only if for every s,

t

Es t2
converges, as t , to a finite limit independent of
time s information set.
c
Eduardo Rossi
52
Persistence
2
Whether or not shocks to t persist depends very much on
which definition is adopted. The conditional moment may diverge to
infinity for some , but converge to a well-behaved limit independent
2
of initial conditions for other , even when the t is stationary and
ergodic.
c
Eduardo Rossi
53
Persistence
GARCH(1,1):
2
t+1
t2
t2
=+
"
1 2t
= 1+
=+
2
t1
t2
c
Eduardo Rossi
Y
k
X
=+
2
1 zti
2
1 zt1
2
t2
+ 1
2
1 zt2
2
t2
t2
+ 1
k=1 i=1
2
t1
=+
1 t2
+ 1
2
1 zt2
1 zt2
+ 1
2
1 zt1
+ 1

2
2
2
2
= + 1 zt1 + 1 + t2 1 zt2 + 1 1 zt1 + 1

2
2
2
= + 1 zt1 + 1 + 1 zt1 + 1 1 zt2 + 1

2
2
2
2
+t3 1 zt3 + 1 1 zt2 + 1 1 zt1 + 1
= + +
+ 1
54
Persistence
Et3 t2
Es
t(t3)1
k=0
(1 + 1 )k
2
+t3
(1 + 1 ) (1 + 1 ) (1 + 1 )
"ts1
#
X

k
ts
2
2
t =
(1 + 1 ) + ts
(1 + 1 )
k=0

2
Es t converges to the unconditional variance of / (1 1 1 )

as t if and only if 1 + 1 < 1.
In the IGARCH(1,1) model with > 0 and 1 + 1 = 1

2
Es t a.s. as t . Nevertheless, IGARCH models are
strictly stationary and ergodic.
c
Eduardo Rossi
55
The EGARCH(p,q) Model
GARCH models assume that only the magnitude and not the
positivity or negativity of unanticipated excess returns
determines feature t2 .
There exists a negative correlation between stock returns and
changes in returns volatility, i.e. volatility tends to rise in
response to bad news, (excess returns lower than expected)
and to fall in response to good news (excess returns higher
than expected).
c
Eduardo Rossi
56
If we write t2 as a function of lagged t2 and lagged zt2 , where

2t = zt2 t2
t2 = +
q
X
j=1
2
2
j ztj
tj
+
p
X
2
i ti
i=1
it is evident that the conditional variance is invariant to changes in

2
2
sign of the zt s. Moreover, the innovations ztj
tj
are not i.i.d.
The nonnegativity constraints on and k , which are imposed
to ensure that t2 remains nonnegative for all t with probability
one. These constraints imply that increasing zt2 in any period
2
increases t+m
for all m 1, ruling out random oscillatory
behavior in the t2 process.
c
Eduardo Rossi
57
The GARCH models are not able to explain the observed

covariance between 2t and tj . This is possible only if the
conditional variance is expressed as an asymmetric function of
tj .
In GARCH(1,1) models, shocks may persist in one norm and die
out in another, so the conditional moments of GARCH(1,1) may
explode even when the process is strictly stationary and ergodic.
GARCH models essentially specify the behavior of the square of
the data. In this case a few large observations can dominate the
sample.
c
Eduardo Rossi
58

In the EGARCH(p,q) model (Exponential GARCH(p,q)) put forward
by Nelson the t2 depends on both size and the sign of lagged
residuals. This is the first example of asymmetric model:

2
ln t
= +
p
X
i ln
2
ti
i=1
q
X
i [zti + (|zti | E |zti |)]
i=1
1 1, E |zt | = (2/)1/2 given that zt N ID(0, 1), where the

parameters , i , i are not restricted to be nonnegative. Let define
g (zt ) zt + [|zt | E |zt |]
by construction {g (zt )}
t= is a zero-mean, i.i.d. random sequence.
c
Eduardo Rossi
59
The components of g (zt ) are zt and [|zt | E |zt |], each with
mean zero.
If the distribution of zt is symmetric, the components are
orthogonal, but not independent.
Over the range 0 < zt < , g (zt ) is linear in zt with slope + ,
and over the range < zt 0, g (zt ) is linear with slope .
The term [|zt | E |zt |] represents a magnitude effect.
c
Eduardo Rossi
60
2
t+1
If > 0 and = 0, the innovation in ln

is positive
(negative) when the magnitude of zt is larger (smaller) than its
expected value.
If = 0 and < 0, the innovation in conditional variance is now
positive (negative) when returns innovations are negative
(positive).
A negative shock to the returns which would increase the debt to
equity ratio and therefore increase uncertainty of future returns
could be accounted for when i > 0 and < 0.
c
Eduardo Rossi
61
2
t+1
In the EGARCH model ln

is homoskedastic conditional on t2 ,

2
and the partial correlation between zt and ln t+1 is constant
conditional on t2 .
An alternative possible specification of the news impact curve is the

following (Bollerslev, Engle, Nelson (1994))

1 |zt |
1 zt
1 |zt |
20
20
2
g(zt , t ) = t
+t
Et
1 + 2 |zt |
1 + 2 |zt |
1 + 2 |zt |
The parameters 0 and 0 parameters allow both the conditional

2
variance of ln t+1 and its conditional correlation with zt to vary
with the level of t2 .
c
Eduardo Rossi
62
If 1 < 0 then Corrt (ln
2
t+1
, zt ) < 0: leverage effect.
The EGARCH model constraints 0 = 0 = 0, so that the conditional

2
correlation is constant, as is the conditional variance of ln t .
The , 2 , and 2 parameters give the model flexibility in how much
weight to assign to the tail observations: e.g., 2 > 0, 2 > 0, the
model downweights large |zt |s.
Nelson assumes that zt has a GED distribution (exponential power
family). The density of a GED random variable normalized is:

1
exp 2 |z/|
f (z; ) =
< z < , 0 <
(1+1/)
2
(1/)
c
Eduardo Rossi
63

where () is the gamma function, and
h
i1/2
(2/)
2
(1/) / (3/)
is a tail thickness parameter.
zs distribution
=2
standard normal distribution
<2
thicker tails than the normal
=1
double exponential distribution
>2
thinner tails than the normal

1/2 1/2
uniformly distributed on 3 , 3
21/ (2/)
With this density, E |zt | =
.
(1/)
c
Eduardo Rossi
64
The Generalized t Distribution takes the form:

1
f t t ; , =
+1/
2t b 1/ B (1/, ) [1 + |t | / (b t )]
where B (1/, ) (1/) () (1/ + ) denotes the beta function,
b [ () (1/) / (3/) ( 2/)]
1/2
t t1
and > 2, > 0 and > 0. The factor b makes V ar

= 1.
The Generalized t nests both the Students t distribution and the
GED.
c
Eduardo Rossi
65
The Students t-distribution sets = 2 and =

freedom
1
2
times the degree of
The GED is obtained for = . The GED has only one shape
parameter , which is apparently insufficient to fit both the central
part and the tails of the conditional distribution.
c
Eduardo Rossi
66
Stationarity of EGARCH(p,q)
In order to simply state the stationarity conditions, we write the

EGARCH(p,q) model as:
"
ln
p
X
i Li ln t
i=1
t2

2
ln t
c
Eduardo Rossi
"
= 1

2
p
X
i=1
q
X
=+
i Li [zt + (|zt | E |zt |)]
#1
i=1
"
+ 1
X
= +
i g (zti )
p
X
i Li
i=1
#1 " q
X
i=1
i Li g (zt )
i=1
67
Given 6= 0 or 6= 0, then

ln t2 <
a.s. when
2i <
i=1
follows from the independence and finite variance of the g (zt ) and
from Billingsley (1986, Theorem 22.6). From this we have that

2

t
ln
< a.s.

exp( )

t2

exp( ) < a.s.

2
exp ( ) t < , {exp ( /2) t } < , a.s., where t = zt t , zt
is i.i.d.. We can also show that they are ergodic and strictly
stationarity.
c
Eduardo Rossi
68
t2
The first two moments of ln

are finite and time invariant:

2
E ln t = 0 for all t

2
V ar ln t
= V ar (g (zt ))
2i
i=1
t2
since V ar (g (zt )) is finite and the distribution of ln

is

independent of t, the first two moments of ln t2 are finite

2
and time invariant, so ln t is covariance stationary if

P
P
2
2
2

i < . If
i = , then ln t = almost surely.
i=1
c
Eduardo Rossi
i=1
69

2
p
P
Since ln t is written in ARMA(p,q) form, when 1

i xi and
i=1
q

P
i xi have no common roots, conditions for strict stationarity of
i=1

2
exp ( ) t and {exp ( /2) t } are equivalent to all the roots

p
P
of 1
i xi lying outside the unit circle.
i=1
2
The strict stationarity of exp ( ) t , {exp ( /2) t } need

2
not imply covariance stationarity, since exp ( ) t ,
{exp ( /2) t } may fail to have finite unconditional means and
variances.
c
Eduardo Rossi
70
For some distribution of {zt } (e.g., the Student t with finite degrees

2
of freedom), exp ( ) t and {exp ( /2) t } typically have no
finite unconditional moments.
If the distribution of zt is GED and is thinner-tailed than the double

P
2
2
exponential ( > 1), and if
i < , then exp ( ) t and
i=1
{exp ( /2) t } are not only strictly stationary and ergodic,

but have arbitrary finite moments, which in turn implies
that they are covariance stationary.
c
Eduardo Rossi
71
Other Asymmetric Models
The Non linear ARCH(p,q) model (Engle - Bollerslev 1986):

q
p
X
X
t = +
i |ti | +
i ti
i=1
i=1
q
p
X
X
t = +
i |ti k| +
i ti
i=1
i=1
for k 6= 0, the innovations in t will depend on the size as well as the

sign of lagged residuals, thereby allowing for the leverage effect in
stock return volatility.
c
Eduardo Rossi
72
The Glosten - Jagannathan - Runkle model (1993):

t2
=+
p
X
2
i ti
i=1
q
X
i 2t1
2
ti
i Sti
i=1
where
St
c
Eduardo Rossi
1
=
0
if
t < 0
if
t 0
73
The Asymmetric GARCH(p,q) model (Engle, 1990):

t2 = +
q
X
i (ti + ) +
i=1
p
X
2
i ti
i=1
The QGARCH by Sentana (1995):

t2 = 2 + xtq + xtq Axtq +
p
X
2
i ti
i=1
when xtq = (t1 , . . . , tq ) . The linear term ( xtq ) allows for

asymmetry. The off-diagonal elements of A accounts for interaction
effects of lagged values of xt on the conditional variance. The
QGARCH nests several asymmetric models.
c
Eduardo Rossi
74
The GARCH-in-mean Model
The GARCH-in-mean (GARCH-M) proposed by Engle, Lilien and

Robins (1987) consists of the system:

2
yt = 0 + 1 xt + 2 g t + t
t2 = 0 +
q
X
i 2t1 +
i=1
t | t1 N (0, t2 )
p
X
2
i t1
i=1
When yt (rt rf ), where (rt rf ) is the risk premium on holding

the asset, then the GARCH-M represents a simple way to model the
relation between risk premium and its conditional variance.
c
Eduardo Rossi
75
This model characterizes the evolution of the mean and the variance
of a time series simultaneously.
The GARCH-M model therefore allows to analize the possibility of
time-varying risk premium.
It turns out that:
t2
yt | t1 N (0 + 1 xt + 2 g
, t2 )
p 2

2
2
2
2
In applications, g t = t , g t = ln t and g t = t2 have
been used.
c
Eduardo Rossi
76
The News Impact Curve
The news have asymmetric effects on volatility.

In the asymmetric volatility models good news and bad news
have different predictability for future volatility.
The news impact curve characterizes the impact of past return
shocks on the return volatility which is implicit in a volatility
model.
c
Eduardo Rossi
77
Holding constant the information dated t 2 and earlier, we can

2
examine the implied relation between t1 and t2 , with ti
= 2
i = 1, . . . , p.
This impact curve relates past return shocks (news) to current
volatility.
This curve measures how new information is incorporated into
volatility estimates.
c
Eduardo Rossi
78
For the GARCH model the News Impact Curve (NIC) is centered on
t1 = 0.
GARCH(1,1):
2
t2 = + 2t1 + t1
The news impact curve has the following expression:

t2 = A + 2t1
A + 2
c
Eduardo Rossi
79

In the case of EGARCH model the curve has its minimum at
t1 = 0 and is exponentially increasing in both directions but with
different paramters.
EGARCH(1,1):
t2
= + ln
2
t1
+ zt1 + (|zt1 | E |zt1 |)
where zt = t /t . The news impact curve is

+
A
exp
t1
f or t1 > 0

t2 =
t1
f or t1 < 0
A exp
< 0
c
Eduardo Rossi
i
p
exp 2/
+>0
80
The EGARCH allows good news and bad news to have different
impact on volatility, while the standard GARCH does not.
The EGARCH model allows big news to have a greater impact
on volatility than GARCH model. EGARCH would have higher
variances in both directions because the exponential curve
eventually dominates the quadrature.
c
Eduardo Rossi
81
The Asymmetric GARCH(1,1) (Engle, 1990)

2
t2 = + (t1 + )2 + t1
the NIC is
t2 = A + (t1 + )
A + 2
> 0, 0 < 1, > 0, 0 < 1.
is asymmetric and centered at t1 = .
c
Eduardo Rossi
82

The Glosten-Jagannathan-Runkle model
2
t2 = + 2t + t1
+ St1
2t1
1 if
t1 < 0
St1 =
0 otherwise
The NIC is
A + 2t1 if t1 > 0
2
t =
A + ( + ) 2
t1 < 0
t1 if
A + 2
> 0, 0 < 1, > 0, 0 < 1, + < 1

is centered at t1 = .
c
Eduardo Rossi
83
The GARCH-in-mean (GARCH-M) proposed by Engle, Lilien and

Robins (1987) consists of the system:

2
yt = 0 + 1 xt + 2 g t + t
t2 = 0 +
q
X
i 2t1 +
i=1
t | t1 N (0, t2 )
p
X
2
i t1
i=1
When yt (rt rf ), where (rt rf ) is the risk premium on holding

the asset, then the GARCH-M represents a simple way to model the
relation between risk premium and its conditional variance.
c
Eduardo Rossi
84
This model characterizes the evolution of the mean and the variance
of a time series simultaneously.
The GARCH-M model therefore allows to analyze the possibility of
time-varying risk premium.
It turns out that:
t2
yt | t1 N (0 + 1 xt + 2 g
, t2 )
p 2

2
2
2
2
In applications, g t = t , g t = ln t and g t = t2 have
been used.
c
Eduardo Rossi
85
The news have asymmetric effects on volatility.

In the asymmetric volatility models good news and bad news
have different predictability for future volatility.
The news impact curve characterizes the impact of past return
shocks on the return volatility which is implicit in a volatility
model.
c
Eduardo Rossi
86
Holding constant the information dated t 2 and earlier, we can

2
examine the implied relation between t1 and t2 , with ti
= 2
i = 1, . . . , p.
This impact curve relates past return shocks (news) to current
volatility.
This curve measures how new information is incorporated into
volatility estimates.
c
Eduardo Rossi
87
For the GARCH model the News Impact Curve (NIC) is centered on
t1 = 0.
GARCH(1,1):
2
t2 = + 2t1 + t1
The news impact curve has the following expression:

t2 = A + 2t1
A + 2
c
Eduardo Rossi
88

In the case of EGARCH model the curve has its minimum at
t1 = 0 and is exponentially increasing in both directions but with
different paramters.
EGARCH(1,1):
t2
= + ln
2
t1
+ zt1 + (|zt1 | E |zt1 |)
where zt = t /t . The news impact curve is

+
A
exp
t1
f or t1 > 0

t2 =
t1
f or t1 < 0
A exp
< 0
c
Eduardo Rossi
i
p
exp 2/
+>0
89
The EGARCH allows good news and bad news to have different
impact on volatility, while the standard GARCH does not.
The EGARCH model allows big news to have a greater impact
on volatility than GARCH model. EGARCH would have higher
variances in both directions because the exponential curve
eventually dominates the quadrature.
c
Eduardo Rossi
90
The Asymmetric GARCH(1,1) (Engle, 1990)

2
t2 = + (t1 + )2 + t1
the NIC is
t2 = A + (t1 + )
A + 2
> 0, 0 < 1, > 0, 0 < 1.
is asymmetric and centered at t1 = .
c
Eduardo Rossi
91

The Glosten-Jagannathan-Runkle model
2
t2 = + 2t + t1
+ St1
2t1
1 if
t1 < 0
St1 =
0 otherwise
The NIC is
A + 2t1 if t1 > 0
2
t =
A + ( + ) 2
t1 < 0
t1 if
A + 2
> 0, 0 < 1, > 0, 0 < 1, + < 1

is centered at t1 = .
c
Eduardo Rossi
92

Arch Uni Slides

Transféré par

Informations du document

Description originale:

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Arch Uni Slides

Transféré par

Droits d'auteur :

Formats disponibles

GARCH Models

Econometria mercati finanziari (avanzato)

Econometria mercati finanziari (avanzato)

GARCH models have been developed to account for empirical

Econometria mercati finanziari (avanzato)

Volatility of the return series appears to be clustered.

Econometria mercati finanziari (avanzato)

Why do we need ARCH models?

Wolds decomposition theorem establishes that any covariance

Econometria mercati finanziari (avanzato)

Why do we need ARCH models?

The uncorrelated innovation sequence need not to be Gaussian and

Econometria mercati finanziari (avanzato)

Why do we need ARCH models?

which are both invariant in time. The conditional mean is time

where the information set is t1 = {t1 , t2 , . . .}.

Econometria mercati finanziari (avanzato)

Why do we need ARCH models?

This restriction manifests itself in the properties of the k-step-ahead

and the associated prediction error is

Econometria mercati finanziari (avanzato)

Why do we need ARCH models?

which has a conditional prediction error variance

For any k, the conditional prediction error variance depends only on

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

(Bollerslev, Engle and Nelson, 1994) The process {t (0 )} follows an

and the conditional variance

depends non trivially on the -field generated by the past

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

Let {yt (0 )} denote the stochastic process of interest with

t (0 ) and t2 (0 ) are measurable with respect to the time t 1

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

It follows from eq.(1) and (2), that the standardized process

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

We can think of t (0 ) as generated by

where 2t (0 ) is unbiased estimator of t2 (0 ).

because zt2 |t1 2(1) .

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

If the conditional distribution of zt is time invariant with a finite

by Jensens inequality. The equality holds true for a constant

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

The kurtosis can be expressed as a function of the variability of the

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

In this model to assure a positive conditional variance the parameters

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

i < 1; in this case

the unconditional variance is given by

Econometria mercati finanziari (avanzato)

The ARCH(q) Model

The process is characterised by leptokurtosis in excess with respect

with 312 < 1, when 312 = 1 we have