Strange EGARCH-M forecast bug.

[CharlieEVIEWS]

Hi guys,

Using EViews 8 May 6 build (although I suspect that nothing recent has changed the equation which i'm trying to estimate), I'm running forecasts of EGARCH-M models. I've run this successfully to estimate about 80 out of sample forecasts or so, so far, but one series gives me a strange 'unable to compute due to missing data' error. However, there is no missing data, and this is only a uni-variate forecasting exercise.

Attached is the work-file with the series with an equation with the egarch(1,1,1) with log variance in the mean equation. Forecasting from this equation seems fine throughout the whole sample, but if you change the start date to just run out of sample forecasts (in fact, anything from about 2007:q1?) on-wards, it gives the error, which seems strange....

Any suggestions as to why this is happening would be great (maybe it isnt a bug, and I'm just using a very naive model specification, which I am, but I am still confused as to why it gives me this specific error message).

Best wishes, and thanks again for your time,

Charlie

[EViews Glenn]

I'm looking into this one.

[EViews Glenn]

I think the issue is that that the equation estimates are not very good.

My guess is that you are using a convergence tolerance that is set to a high value (possibly a default of 1e-3) so that we get convergence reported when it really shouldn't. So what's happening is that you are getting NAs in the GARCH series which feed into the forecast for forecasts that start at the late date.

If you set your convergence tolerance to, say 1e-8, and reestimate you'll be able to perform a forecast, even starting at the latter periods. I caution you though, that examination of the gradients suggests that this model really hasn't converged in the usual sense. The model is stalling out the final parameter values because the GARCH series is generating NAs in later periods.

The bottom line, I think, is that the EGARCH-M spec just doesn't work well for these data.