The Persistence of volatility
High frequency financial time series data exhibit volatility clustering, which
shows that some periods are highly volatile with some degree of
autocorrelation.
0.02
0
Forecasting Mean and Variance
Nifty return (2/01/1996 - 20/12/2002)
0.1
0.05
0
-0.05
-0.1
1
87
173
259
345
431
517
603
689
775
861
947
1033
1119
1205
1291
1377
1463
1549
1635
1721
0.01
0.002
0.004
0.006
0.008
0.012
0
02-Jan-1996
29-Feb-1996
30-Apr-1996
25-Jun-1996
22-Aug-1996
18-Oct-1996
16-Dec-1996
14-Feb-1997
17-Apr-1997
16-Jun-1997
12-Aug-1997
15-Oct-1997
12-Dec-1997
10-Feb-1998
10-Apr-1998
10-Jun-1998
05-Aug-1998
05-Oct-1998
30-Nov-1998
28-Jan-1999
24-Mar-1999
24-May-1999
19-Jul-1999
14-Sep-1999
09-Nov-1999
06-Jan-2000
03-Mar-2000
SQUARE OF NIFTY RETURNS
05-May-2000
30-Jun-2000
28-Aug-2000
25-Oct-2000
20-Dec-2000
16-Feb-2001
18-Apr-2001
13-Jun-2001
08-Aug-2001
08-Oct-2001
06-Dec-2001
04-Feb-2002
03-Apr-2002
30-May-2002
25-Jul-2002
23-Sep-2002
22-Nov-2002
variance of the return. In this context, ARCH models are highly useful.
useful. For instance, asset holders want to forecast both mean and
It is easy to imagine a situation in which prediction of volatility is highly
Why to choose an ARCH model?
The Problem of Heteroscedasticity
Define a linear regression model:
Yt = 0 + 1 X t + t
One of the basic assumptions of OLS method is:
E ( t2 ) = 2
Implies that the conditional variance of Yi is a constant or the errors (i) are
homocedastic. On the contrary, if
E ( t2 ) = t2
then the variance is not a constant or the errors are Heteroscedastic. This is a
common problem with cross section data.
If so, estimates (0 and 1) are unbiased, but conventional estimates of
standard errors and confidence intervals are very narrow, giving false sense of
precision. For instance, the usual variance formula is:
var(1 ) = 2 xt2
whereas the true variance with heteroskedastic errors is:
*
var(1 ) = x 2
t t
2
( x )
2 2
t
Obviously,
*
var(1 ) var(1 )
Thus, application of OLS method provides biased variance estimate; hence,
inference based on OLS estimates will be misleading.
Conditional Forecast Variance
Estimate a stationary AR (1) model:
Yt = a0 + a1Yt −1 + t
The conditional forecast of Yt+1 is:
EtYt +1 = a0 + a1Yt
The forecast error variance is:
Et [(Yt+1 − a0 − a1Yt )]2 = Et t2+1 = 2
The unconditional forecast of {Yt} is always its long run mean:
a0 (1 − a1 )
The unconditional forecast error variance is:
E{[Yt +1 − a0 /(1 − a1 )]2 } = E[( t +1 + a1 t + a12 t −1 + ...) 2 ]
= 2 /(1 − a12 )
Since
1 /(1 − a12 ) 1
The unconditional forecast variance is greater than conditional forecast variance.
ARCH Process
Thus far, we have assumed that the variance is a constant. However, time series
data with volatility clustering indicate that the variance is not a constant; hence,
it is essential to model the variance as an AR process. Suppose we estimate the
model:
Yt = X t + t
2 2 2
t = 0 + 1 t −1 + 2 t −2 + t
where t is a white noise process. The conditional variance for t is:
2
E t = 0 + 1 t2−1 + 2 t2−2
The above equation is called an ARCH (2) process.
The most convenient way is to model both mean and variance simultaneously
using maximum likelihood techniques. Engle (1982) provides a multiplicative
conditionally heteroscedastic model:
t = t 0 + 1 t2−1
where t is a white noise process with unit variance; 0 > 0 and 0 < 1 < 1.
An Illustration
Suppose, we want to construct and forecast the volatility of stock return (rt)
using an ARCH (1) model:
rt = + t t2
t2 = 0 + 1 t2−1
where t −1 = t −1 t2−1
They are mean and variance equations respectively. There are three
parameters to be estimated (, 0, and 1) using maximum likelihood
method.
The one step ahead forecast of conditional variance of rt is:
t2+1 = 0 + 1 t2
The GARCH Model
Bollerslev (1986) defines conditional variance as an ARMA process, which is
called gerneralised ARCH model or GARCH model. The GARCH (1,1) model
is:
rt = + t t2
t2 = 0 + 1 t2−1 + 2 t2−1
where t-1 and t-1 are ARCH and GARCH terms respectively.
Eviews uses the following maximum likelihood method to estimate these
parameters simultaneously:
1 1 1
lt = − log( 2 ) − log( t ) − ( yt − xt )2 / t2
2 '
2 2 2
The ARCH-M Model
In some of the applications, the dependent variable in the mean equation
may be determined by its conditional variance. For instance, the expected
return on an asset is related to expected asset risk. If so, the mean equation
can be defined as:
rt = + t2 + t t2
t2 = 0 + 1 t2−1 + 2 t2−1
Such models are called ARCH-in-Mean or ARCH-M Models due to
(Engle, Lillien, Robins, 1987).
The Asymmetric ARCH Models
Often, we notice that the downward movements in the market are highly
volatile than upward movement of the same magnitude. In such case,
symmetric ARCH model undermines the true variance process. Engle and
Ng (1993) provide a news impact curve with asymmetric response to good
and bad news.
Volatility
Bad news Good news
Eviews supports estimating two types of asymmetric ARCH models –
TARCH and EGARCH.
The TARCH Model
The T ARCH or Threshold ARCH due to Zakoian (1990) can be defined as:
rt = + t t2
t2 = 0 + 1 t2−1 + 2 t2−1 + t2−1d t −1
where d t-1 = 1 if ε t -1 0 and 0 otherwise
ε t -1 0 good news, ε t -1 0 bad news
good news has an impact of 1 t2-1 while bad news has an
impact of (1 + ) t2-1
if 0, the news impact is asymmetric .
If 0, there is leverage effect.
The EGARCH Model
The EGARCH or exponential GARCH model proposed by Nelson (1991)
takes the form:
t −1 t −1
log = 0 + 1 log
2 2
t −1 + 3 +
t −1 t −1
t
This implies that the leverage effect is exponential than quadratic.
If t-1 t-1 0 the impact is ( 3 + )
if t-1 t-1 0 the impact is ( 3 − )
if 0, the impact is asymmetric
If 0, there is leverage effect.
Assessing the Fit of the Model
We can use the AIC and SBC criteria to examine the fit of the model.
However, we need to be careful with the statistical properties of such
criteria in ARCH context. For instance, define
AIC = T ln (RSS) +2n
Note that RSS = 2; hence, the AIC assesses the fit of the mean
model. Instead, define
T
RSS = ( t2 − t2 ) 2
!
t =1
and select the model that provides minimum RSS!.
Diagnostic Checks for Model Adequacy
An estimated GARCH model must capture all the dynamics of mean and
variance model. The estimated errors (t) are serially uncorrelated and do not
display any remaining conditional volatility.
Standardize the residuals:
St = t 2
Construct the Ljung and Box Q statistics:
s
Q = T (T + 2) rk2 (T − k )
k =1
T is number of observations; r is autocorrelation coefficient. If estimated Q
exceeds critical value of 2 with degrees of freedom s then at least one of the
autocorrelation coefficients is statistically significant; implying that mean of the
model is not properly specified.
Construct Q statistics for s2 and if the calculated statistics exceeds critical value
then it implies that the GARCH effect is not completely captured by the model.
