Re: [R] TSeries GARCH Estimates accuracy

From: Patrick Burns <>
Date: Tue 12 Apr 2005 - 18:18:05 EST

GARCH models are notoriously hard to optimize, so I'm not terribly surprised.

The first thing is to make sure that your reference results are really better
than what you are getting in R. Perhaps they have improved it, but the last time I looked at garch in S-PLUS, it did not necessarily give good results. I don't know anything about the SAS routine.

The first-aid approach to getting better estimates is to start the optimization
at various locations and pick the best one you get. One starting point might
be near to (.05, .9) -- that is, .9 times the lagging conditional variance.
Another would be to restart from where the routine ended. From the help file for 'garch' it says that the default starting point is with values close to
zero -- that is not a very good starting point.

A more involved approach would be to change the routine so that the intercept
is derived from the desired asymptotic variance (usually the unconditional variance) and the other parameter estimates. Optimizers tend to be much happier with this problem.

The R-sig-finance list is a more likely spot for a discussion like this.

Patrick Burns

Burns Statistics
+44 (0)20 8525 0696
(home of S Poetry and "A Guide for the Unwilling S User")

Sanjay Kumar Singh wrote:

>I am trying to fit a GARCH(1,1) model to a financial timeseries using the 'garch' function in the tseries package. However the parameter estimates obtained sometimes match with those obtained using SAS or S-Plus (Finmetrics) and sometimes show a completely different result. I understand that this could be due to the way optimization of MLEs are done, however, I would appreciate any help to obtain consistent results using R.
>Also is there any garch simulation function available other than garchSim from fseries package?
>Thanks in advance,
> mailing list
>PLEASE do read the posting guide!
> mailing list PLEASE do read the posting guide! Received on Tue Apr 12 18:23:32 2005

This archive was generated by hypermail 2.1.8 : Fri 03 Mar 2006 - 03:31:07 EST