Possible bug in ARCH/MV-ARCH estimation

[rousing]

Hi

I have found something that appears to be a bug. I noticed it when looking at long-term convergences in an estimated diagonal BEKK model for variances, covariances and correlations which did not correspond to what I was expecting to find. I am not fully sure where the issue comes from but just a few observations that might help to identify the problem. For instance using the example file: intl_fin.wf1 and the series sp500 I noticed the following: Comparing the estimated constant term of a pure OLS of the form: dlog(sp500) c reveals a difference to the constant (mean) estimated when adding a garch(1,1) specification to the variance part. Furthermore, when comparing the sample variance of the residuals of the garch(1,1) model - this does not correspond to the model implied long-term convergence level defined as omega/(1-alpha-beta). This might also translate into the problems in the MV specification.

I am using Eviews 8 with the 2 July 2014 patch.

Thanks in advance for looking into this.

Best regards
Rasmus

[EViews Gareth]

I'm not sure I follow.

Are you saying that you expect the estimate of the constant term from the following two equations to be the same?

Code: Select all

equation eq1.ls y c equation eq1.arch y c

[rousing]

yes, I would not expect the mean estimate, here c, to be impacted by the choice of variance specification

[EViews Gareth]

Unfortunately, it is.

[rousing]

ok, maybe I am just getting rusty - never noticed it changed in the past (which is not to say it did not...) the estimated mean, however, is not where I see the major problem, but that the unconditional variance of the residual do not match the level prescribed by the garch parameters, or maybe I am getting that wrong as well...

[EViews Gareth]

Which residuals?

[rousing]

those produced by the garch model, so y-c in your specification: equation eq1.arch y c

[EViews Gareth]

Probably just means that your model specification is wrong. It is an identity relationship.

Use the TSDGP add-in to generate a truly GARCH based variable, then do the same comparison. You should see that the two values are pretty darn close.

[rousing]

if you don't see an issue that is fine, thanks for your time :)