Modelling Volatility -

GARCH

Heteroscedasticity Revisited

An example of a structural model is


with u
t
~ N(0, ).

The assumption that the variance of the errors is constant is known as
homoskedasticity, i.e. Var (u
t
) = .

What if the variance of the errors is not constant?
- Heteroskedasticity
- Implies that standard error estimates could be wrong.
- Then inference may be wrong (t-tests, F-tests etc.)
Is the variance of the errors likely to be constant over time? Not for
financial data.
o
u
2
o
u
2
y
t
= |
1
+ |
2
x
2t
+ |
3
x
3t
+ |
4
x
4t
+ u
t

A Sample Financial Asset Returns Time Series
Daily S&P 500 Returns for January 1990 December
1999


-0.08
-0.06
-0.04
-0.02
0.00
0.02
0.04
0.06
1/01/90 11/01/93 9/01/97
Return
Date
Financial Data
Financial data shows certain periods to have higher
volatility than others
Consider an investor who buys an asset at time t and
wishes to sell it at time t+1. The investor will want a
forecast on the rate of return but will also be interested in
the variance of the return over the period.
[We usually think of the variance of returns as being a
measure of risk!!!]
An asset that always yields a return of 5% is different from one
that sometimes loses 20% and sometimes gains 30%, the
average return is the same but the variance is different!
Autoregressive Conditionally Heteroskedastic
(ARCH) Models
Periods of high volatility are often concentrated and
followed by periods of lower volatility that are often
concentrated.
The assumption of homoskedasticity is very limiting in such
circumstances
So we use a model which does not assume that the variance is
constant.
Unconditional Variance: the long run forecast of the
variance and can still be considered constant
Conditional variance: based on our best model of the
variable under consideration

Autoregressive Conditionally Heteroskedastic (ARCH) Models
What could the current value of the variance of the errors plausibly depend
upon?
The conditional variance of a zero-mean normally distributed
random variable u
t
is equal to the conditional expected value of the
square of u
t
.
Assume our error term is a normally distributed random variable:
=>Variance of error terms depend on Previous squared error terms.
This leads to autocorrelation in volatility, which is modelled by
allowing the conditional variance of the error term to depend on the
immediately previous value of the squared error.
This leads to the autoregressive conditionally heteroskedastic model for
the variance of the errors:
= o
0
+ o
1

This is known as an ARCH(1) model.
[Basically what we are doing is modelling the volatility in addition to the
usual model!]
o
t
2
u
t 1
2
Autoregressive Conditionally Hetroscedastic
(ARCH) Models (contd)
The full model would be
y
t
= |
1
+ |
2
x
2t
+ ... + |
k
x
kt
+ u
t
, with u
t
~ N(0, )

where = o
0
+ o
1

We can easily extend this to the general case where the error
variance depends on q lags of squared errors:
= o
0
+ o
1
+o
2
+...+o
q


This is an ARCH(q) model.
Instead of calling the variance , in the literature it is usually
called h
t
so the model is
y
t
= |
1
+ |
2
x
2t
+ ... + |
k
x
kt
+ u
t
, u
t
~ N(0,h
t
)
where h
t
= o
0
+ o
1
+o
2
+...+o
q
o
t
2
o
t
2
o
t
2
u
t 1
2
u
t q
2
u
t q
2
o
t
2
2
1 t
u
2
2 t
u
2
1 t
u
2
2 t
u
Another Way of Writing ARCH Models
For illustration, consider an ARCH(1).
y
t
= |
1
+ |
2
x
2t
+ ... + |
k
x
kt
+ u
t
, with u
t
~ N(0, )

where = o
0
+ o
1


Instead of the above, we can write

y
t
= |
1
+ |
2
x
2t
+ ... + |
k
x
kt
+ u
t
, with u
t
= v
t
o
t

, v
t
~ N(0,1)

The two are different ways of expressing exactly the same model. The first
form is easier to understand while the second form is required for
simulating from an ARCH model, for example.

o o o
t t
u = +
0 1 1
2
o
t
2
u
t 1
2
Testing for ARCH Effects

1. First, run any postulated linear regression of the form given in the equation
above, e.g. y
t
= |
1
+ |
2
x
2t
+ ... + |
k
x
kt
+ u
t

saving the residuals, .

2. Then square the residuals, and regress them on q own lags to test for ARCH
of order q, i.e. run the regression


where v
t
is an error term.
Obtain R
2
from this regression

t
u
t q t q t t t
v u u u u + + + + + =

2 2
2 2
2
1 1 0
2
...
Testing for ARCH Effects (contd)

3. The test statistic is defined as T*R
2
(the number of observations
multiplied by the coefficient of multiple correlation) from the last
regression, and is distributed as a _
2
(q).

4. The null and alternative hypotheses are
H
0
:
1
= 0 and
2
= 0 and
3
= 0 and ... and
q
= 0
[i.e. No autocorrelation in residuals -> ARCH not appropriate]
H
1
:
1
= 0 or
2
= 0 or
3
= 0 or ... or
q
= 0.
[i.e. ARCH is appropriate]

If the value of the test statistic is greater than the critical value from the
_
2
distribution, then reject the null hypothesis.



Problems with ARCH(q) Models

How do we decide on q?
The required value of q might be very large
Non-negativity constraints might be violated.
When we estimate an ARCH model, we require o
i
>0
for all i=1,2,...,q (since variance cannot be negative)

A natural extension of an ARCH(q) model which gets
around some of these problems is a GARCH model.

Generalised ARCH (GARCH) Models
Due to Bollerslev (1986). Allow the conditional variance to be dependent upon previous
own lags
The variance equation is now
(1)
This is a GARCH(1,1) model, which is like an ARMA(1,1) model for the variance
equation.
We could also write


Substituting into (1) for o
t-1
2
:



An infinite number of successive substitutions would yield


o
t
2
= o
0
+ o
1
2
1 t
u +|o
t-1
2

o
t-1
2
= o
0
+ o
1
2
2 t
u +|o
t-2
2

o
t-2
2
= o
0
+ o
1
2
3 t
u +|o
t-3
2

o
t
2
= o
0
+o
1
2
1 t
u +|(o
0
+o
1
2
2 t
u +|o
t-2
2
)
= o
0
+o
1
2
1 t
u +o
0
| +o
1
|
2
2 t
u +|o
t-2
2

2
3
2
3
2 2
2
2
1 1
2
0
2
...) ( ...) 1 (

+ + + + + + + + =
t t t t t
u u u o | | | o | | o o
Generalised ARCH (GARCH) Models (contd)



The first expression on the RHS in the above equation is a
constant as the number of substitutions tends to infinity while

tends to zero. So the GARCH(1,1) model can be written as
an infinite order ARCH model.

=
1
2 1
1
0
2
1
j
j t
j
t
u | o
|
o
o
2
3
2
3
2 2
2
2
1 1
2
0
2
...) ( ...) 1 (

+ + + + + + + + =
t t t t t
u u u o | | | o | | o o
Generalised ARCH (GARCH) Models (contd)
We can again extend the GARCH(1,1) model to a GARCH(p,q):



But in general a GARCH(1,1) model will be sufficient to capture the
volatility clustering in the data.

Why is GARCH Better than ARCH?
- more parsimonious - avoids overfitting
- less likely to breech non-negativity constraints



o
t
2

= o
0
+o
1
2
1 t
u +o
2
2
2 t
u +...+o
q
2
q t
u

+|
1
o
t-1
2
+|
2
o
t-2
2
+...+|
p
o
t-p
2

o
t
2
=

= =

+ +
q
i
p
j
j t j i t i
u
1 1
2
2
0
o | o o
Estimation of ARCH / GARCH Models
Since the model is no longer of the usual linear form, we cannot use OLS.
Remember OLS seeks to minimise the RSS. This is no longer an appropriate
objective when modelling conditional variance

We use another technique known as maximum likelihood. An intuitive account
of that method is provided here. See Brooks for a more formal explanation.

The method works by finding the most likely values of the parameters given
the actual data. We select a set of values for the parameters to be estimated, in
principle we then calculate the probability that the set of endogenous variables,
which we have observed in our data set, would actually have occurred. We then
select the set of parameters which maximise this probability.

More specifically, we form a log-likelihood function and maximise it.




Extensions to the Basic GARCH Model

Since the GARCH model was developed, a huge number of extensions and
variants have been proposed. Three of the most important examples are
EGARCH, GJR, and GARCH-M models.

Problems with GARCH(p,q) Models:
- Non-negativity constraints may still be violated
- GARCH models cannot account for leverage effects
It is argued that a negative shock to a financial time series is likely to
cause volatility to rise by more than a positive shock. Such
asymmetries are generally referred to as leverage effects

Possible solutions: The threshold GARCH (TGARCH) model which is an
asymmetric GARCH model.


Extensions to the Basic GARCH Model
TGARCH (or GJR model)
Allows for negative shocks (bad news) to have a larger impact on
volatility than positive shocks (good news)

TGARCH was put forward by Zakoian(1990) and Glosten,
Jaganathan and Runkle(1993)

To capture the asymmetries of the negative and positive shocks
a multiplicative dummy variable is added to the variance equation

h
t
= o
0
+ o
1
u
t-1
2

+

u
t-1
2
I
t-1
+ h
t-1
where I
t
takes a value of 1 for u
t
<0, and 0 otherwise.
Hence good news has an impact o
1
while bad news has an
impact
1
+.
If the coefficient =0 the news impact is symmetric

An Example of the use of a TGARCH (GJR) Model

Using monthly S&P 500 returns, December 1979- June 1998

Estimating a GJR model, we obtain the following results for
conditional variance. T-statistics in parenthesis



There is asymmetry as the coefficient has the correct sign and
is significant
) 772 . 5 ( ) 999 . 14 ( ) 437 . 0 ( ) 372 . 16 (
604 . 0 498 . 0 015 . 0 243 . 1
1
2
1
2
1
2
1
2

+ + + =
t t t t t
I u u o o
News Impact Curves
The news impact curve plots the next period volatility (h
t
) that would arise from
various positive and negative values of u
t-1
, given an estimated model.

News Impact Curves for S&P 500 Returns using Coefficients from GARCH and GJR
Model Estimates:

0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
-1 -0.9 -0.8 -0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Value of Lagged Shock
V
a
l
u
e

o
f

C
o
n
d
i
t
i
o
n
a
l

V
a
r
i
a
n
c
e
GARCH
GJR
GARCH-in Mean

We expect a risk to be compensated by a higher return. So why not let the
return of a security be partly determined by its risk?

Engle, Lilien and Robins (1987) suggested the ARCH-M specification. A
GARCH-M model would be



o can be interpreted as a sort of risk premium.
If o is positive and statistically significant then increased risk given by an
increase in the conditional variance


y
t
= + oo
t-1
+ u
t
, u
t
~ N(0,o
t
2
)
o
t
2
= o
0
+ o
1
2
1 t
u +|o
t-1
2