Extensions of ARCH Models

Dr. Sunita Arora

Associate Professor
Government College for Girls
Sector 14, Gurugram
Email: [email protected]
 Brief overview of ARCH and GARCH models
In linear regression model it is assumed that variance
in error term is unconditional or constant but in ARCH
models variance is conditional to the information
available up to previous period. Thus in ARCH family of
models variance equation is also to be specified.
In ARCH(1) model variance equation is :

 2
  0   1 u t 1
2

ARCH models assumes that conditional variance is a

function of past error terms. This model can be
extended to more past error terms, if required.
 Limitations of ARCH Models
• Value of q i.e. that the number of lagged
squared error required to capture the
dependence of conditional variance may be very
large.
•Non negativity constraint may be violated, as
with more parameters in conditional variance
equation, more probability is of having one or
more parameters to have negative estimated
value.
 GARCH Model
In GARCH models variance is a function of
previous error terms and fitted variance
from previous periods.
GARCH(1,1) Model:

   0  1 ut 1    t 1
2 2 2
 Estimation of ARCH and GARCH
model with eviews
Open the file CNX Nifty with eviews
Generate new variable : Return=dlog(close)
1. We are interested in modelling return and
2. Return is a stationary variable
 Open return series
View Correlogram
Select level and lags to include 36
Correlogram will look like this.
If we note the prob column, it is clear that probability is
less than 0.05 at lag one only, so this may be the case of
AR(1) MA(1)
Quick Estimate Equation
Return ar(1) ma(1)
select method ARCH.
For ARCH(1) model put value 1 in ARCH column and 0 in GARCH column
For GARCH(1,1) model, put 1 in ARCH Column and 1 in
GARCH column
Click OK
Results of GARCH(1,1) will be
Interpreting the results:

looking at p values, it can be interpreted that in mean

equation, coefficients of ar(1) and ma(1) are not significant
but in variance equation all the values i.e. Constant, ARCH
term
RESID(-1)^2 as well as GARCH term
GARCH(-1) are significant
Coefficient of ARCH term depicts whether there are spikes
in the return series higher the coefficient more and larger the
spikes in return series
Sum of ARCH term and GARCH term shows whether
volatility is persistent or not, higher the sum more the
persistence of volatility.
Diagnostics of the model
We have two equations in the model, mean equation and
variance equation. To diagnose whether these equations are
correctly specified or not, steps are
From equation window
View Residual Diagnostic Correlogram, if the mean
equation is correctly specified, all Q values will be
insignificant.
Diagnostic for variance equation
To check whether the variance equation is correctly
specified or not, the steps are:
From equation window View
Correlogram Squared residuals. If the variance equation is
correctly specified, all the Q statistics will be insignificant.
Diagnostics of CNX Nifty Return-
first one is for mean equation i.e. Correlogram of
standardized residual and second one is for Correlogram of
squared residuals.
Need for Extensions of GARCH model

In case of CNX nifty diagnostics of the mean as

well as variance equations show that the model is
correctly specified, then what is the need of
extensions of the model. Reasons for the same
are:
•Non negativity condition may be violated by the
GARCH model.
•To account for leverage effect, i.e. whether there
is same of different effect of good and bad news
from previous period, on the volatility of a period.
Need for Extensions of GARCH model

•To examine the feedback relationship

between conditional variance and
conditional mean.
•To estimate the higher powers or lowers
powers of ARCH and GARCH
•To differentiate between permanent and
transitory effect of volatility.
Models that solve the above mentioned problems
•TARCH and EGARCH model – to account for leverage
effect.
•GARCH-M model – to examine the feedback
relationship between conditional variance and
conditional mean.
•PGARCH- to examine the higher and lower powers of
ARCH and GARCH
•CGARCH- to examine the permanent and transitory
effect of volatility.
GARCH in Mean or GARCH-M model to see whether the
popular commercial saying “More risk more reward” is
applicable to CNX Nifty or not
Mean equation for CNX Nifty will be
Return  c1  t  c2 ar (1)  c3 ma (1)
2

and variance equation will be

σ  c 4  c 5 u t 1  c 6 σt 1
2 2 2

In the mean equation if the coefficient c1 is

significant that means there is feedback relationship
between conditional variance and conditional mean.
Results for GARCH-in-Mean model for CNX Nifty
P value of GARCH coefficient in mean equation is 0.0326 i.e
there is feedback relationship in conditional mean and
conditional variance.
TARCH Model to account for leverage effect

Leverage effect : Good and Bad News have

different effects on the volatility of the next
period. Volatility increases more in the period
following the bad news than the period
following the good news.
Bad News: when the realized value is less than
the expected mean, this will be the case when
error term is negative. We know error term is:
Y Y
Good News: when
TARCH Model in eviews: TGARCH model is also
known as GJR model, known after the authors-
Glosten, Jagannathan and Runkle.

To apply the TARCH model in eviews, steps are

•Quick Estimate equation
•Specify mean equation
•Method: ARCH
•Model: GARCH/TARCH
•ARCH: Specify number, generally one
•GARCH: Specify number, generally one
•Threshold order: specify number, generally one
Mean and variance equations for TARCH
model
Mean equation for TARCH Model for CNX Nifty
data will be:
Return  c1ar (1)  c2 ma(1)
And variance equation will be:
 2  c3  c4 ut21  c5 ut21 * lt 1  c6  t21
where : lt 1  1, if , ut 1  0; otherwise  0
If c5 in the variance equation is positive and
significant the leverage effect is present in the
data.
TARCH Model window in eviews
Results of TRACH model for CNX Nifty data interpretation
of results: as c5 is positive and statistically significant
coefficient =0.084657 with p value 0.000, it shows that
leverage effect is present in the data.
EGARCH model to account for asymmetric or
leverage effect in data. EGARCH is also applied
to account for leverage effect.
To apply EGARCH model in eviews steps are:
• Quick Estimate equation
•Specify mean equation
•Method: ARCH
•Model: EGARCH
•ARCH: Specify number, generally one
•GARCH: Specify number, generally one
•Asymmetric order: specify number, generally one
Mean and variance equations for EGARCH
model
Mean equation for EGARCH Model for CNX Nifty data
will be:
Return  c1ar (1)  c2 ma(1)
And variance equation will be:
Log( t2 )  c3  c4  ABS (u t 1) /@ sqrt ( t21 ) 
c5 * ut 1 /@ sqrt ( t21 ) 
c6 * log( t21 )
For asymmetrics or leverage effect to be present in
data, c5 in the variance equation should be negative
and significant.
EGARCH Model window in eviews
Results of EGRACH model for CNX Nifty data interpretation
of results: as c5 is negative and statistically significant
coefficient =-0.062519 with p value 0.000, it shows that
leverage effect is present in the data.
 Regressors in Variance Equation

Sometimes we are interested in knowing

• Whether turnover on a day has any impact on the
volatility of asset return on subsequent trading day; or
• Whether there is any month of the year effect on
volatility; or
• Whether there is any day of the week effect on
volatility
• Or any other factor has any effect on volatility
In all the above cases we introduce related regressor in
variance equation.
Equation window with regressor in variance
equation

regressor in variance equation

Results for impact of turnover on subsequent day
volatility on CNX Nifty- as in the following table coefficient
of turnover(-1) is negative but not significant, it can be
said that turnover of a day has no impact on volatility of
subsequent day.
Result of impact of turnover on volatility
 A word of Caution
Participants are requested to perform diagnostic
checks of the model applied before quoting the
results.
As already discussed diagnostics for ARCH family of
model are
•Correlogram Q statistics
•Correlogram squared residuals
•ARCH LM test