error in GARCH estimation

[kenshiro]

I will. Thanks Startz!

[kenshiro]

Hello Startz,

I checked the estimated conditional variance series where the p-values (as well as adj.R Sq. and DW statistics) from the regressions are NA. Here is what I found --- the estimated conditional variance series is an explosive process.

(i) Would multicollinearity lead to such a problem?
(ii) Could it be anything like model-misspecification? I mean, can I get rid of it by adjusting GARCH(r,m) choices?

Thank you very much.

Best,
Kenshiro

[startz]

Kenshiro,

It sounds like you're checking the right things. I'm not sure what the outputs mean when you're getting the NA results. You may need advice from the EViews folks.

[kenshiro]

just emailed support@eviews.com. Thank you Startz.

[trubador]

You should make sure that the residual from the mean equation is stationary and serially uncorrelated. It seems your equations violate these. I believe this is the reason behind the changing variance symptoms. Try adding AR(p) lags to the mean equations and see if (G)ARCH effects still remain.

Please keep in mind that "garch-in-mean" specification is quite complicated and you may not always get interpretable empirical results even if such effect is theoretically valid. You should try all three forms of "in-mean" specifications (i.e. variance, stdev and log), since we usually have very little (if any) idea as to the functional form of this relationship.

[kenshiro]

Hi Trubador, thanks a lot for the advice. Let me try.