Another (?) BVAR .forc bug

[CharlieEVIEWS]

When using the .forc command for bayesian vars, the forecasts are generated, but then the equation appears to 'reset' (?) itself with 0.0000 for all coefficients and standard errors in the VAR object. I've tried this with a couple of different datasets, so it isn't a function of that (for whatever reason). When you re-estimate the BVAR, all coefficients and standard errors 're-appear'....

Is this related in some way to the .bvar(prior=NW) crash bug from earlier this week and it will therefore be fixed upcoming patch?

Best wishes

[EViews Gareth]

The .forc command?

[CharlieEVIEWS]

Apologies! I meant .fcast from the VAR forecast add-in!

[EViews Gareth]

Could you provide an example workfile, I cannot replicate this.

[CharlieEVIEWS]

Certainly! I didnt bother in the first instance as it didnt appear to be data dependant.

I have also realised that as a consequence of this, forecasts for all variables after the first variable are also subsequently 0, for the entire forecast sample (so the 'resetting' appears to occur after the forecast for the first endogenous variable is created)

I will however be the first to admit that my knowledge of bvar theory is fairly flimsy!

Best wishes, and thanks for your time

[EViews Gareth]

How do you replicate the bug?

If I open up the workfile, I can see that the VAR test is a little screwed up, but how did it get to that state? If I re-estimate it, it looks fine. If I then run the VAR forecast add-in on it, it appears to work fine.

[CharlieEVIEWS]

I believe this code will show the bug (for my build? at least):

Code: Select all

wfcreate q 1990 2010 series y = nrnd series x = nrnd series z = nrnd var bvartest smpl 1990q1 2004q4 bvartest.bvar(prior=sznw) 1 2 x y z smpl 2005q1 2008q4 bvartest.fcast _f smpl @all

The problem (if indeed it is one) is that y_f (and z_f) has zeros for the forecast sample, and that the object bvartest is 'a little screwed up' as you said before. The forecasts for x_f have appeared to generate correctly however, and we can recover the coefficient estimates through simply estimating the equation again.

Apologies if I am being extremely dense somewhere!

[EViews Gareth]

Thanks. That looks like a bug.