Bayesian VAR

[EViews Esther]

This thread is about the new version of BVAR add-in that perform Bayesian VAR estimation based on analytic results.

[Three ways the new 'BVAR' can be different from the old 'BVAR'] (2010/11/11)

• The new version does not require R package *MSBVAR* installation.
• Unlike the previous BVAR which provides the Sims-Zha (IER, 1998) priors only, the new BVAR provides the 5 different prior;
  Normal-Wishart (default), Normal-flat (Sims-Zha priors), conjugate Normal-Wishart, Minnesota/Litterman (Koop-Korobilis; 2009), and Diffuse priors.
• Added marginal likelihood computation.

[Change log]

• 2010/11/30 Added model object option (command option "m=model name"; e.g. bvar(m=model) y x1 x2 x3)
• 2011/01/10 Corrected IRFs
• 2011/01/20 Make "the list of endogenous variables" box bigger
• 2011/03/22 Updated command lines
  • E.g. bvar 2 gdp interest inflation
    E.g. bvar(prior=1) 2 gdp interest inflation
    Note: the first argument is the number of lags!
• 2012/02/23 Updated the options for the Ko-Ko priors
  • Selection for the mean and cov of coefficient parameters (BETA ~ N(B0*iota, V0*Identity))
    Selection for the degrees of freedom and scale matrix of residual parameters (SIGMA ~ W(NU0,S0*identity)).
    Please note that B0,V0, NU0 and S0 are all scalars.
• 2012/03/23 Edited documentation
• 2012/05/24 Added a new error message

[donihue]

Once again, let me thank you for this superb addition to EViews. Already, the previous version using the R package was of great help; this one is even better!

At the risk of appearing ungrateful for this exceptionally helpful new BVAR implementation, might I suggest that the next step, at least for those of us who work with emerging-country central banks, would be to expand the set of priors to include DSGE-generated priors, as in Del Negro, M., and F. Schorfheide (2004), "Priors from General Equilibrium Models for VARs", International Economic Review, 45(2), 643–673. This has been implemented in Dynare (see Adjemian, S., Darracq-Pariès, M. and S. Moyen (2008), "Towards a monetary policy evaluation framework", ECB Working paper 942) and it would be a major step forward if it could be done in EViews as well.

Again, congratulations and thanks for a major improvement to EViews.

Regards
Donihue

[donihue]

Hello Esther/Gareth,

Again, many thanks for the new BVAR add-in, which is a huge advance.

Could I request a small improvement? The IRF graphs display is too small to read text characters when several variables are used (as is typical for a BVAR related to a DSGE), even when the graph is "extracted" and displayed full-size. Would it be possible to have an option to show the IRF graphs by equation? (This would also deal with a small problem in the current version wherein the IRF graphs do not display the name of the response variable ("Response of .. to Y" has a blank space for ..))

Many thanks and regards
Donihue

[donihue]

Esther/Gareth,

Using the excellent new BVAR add-in leads me to another request, perhaps rather harder to incorporate: would it be possible to have Proc tab on the add-in results function as does that for the ordinary VAR object?
In particular, would it be possible to set it up to create Systems and Models directly from the BVAR results? These options do not seem to be present in the current implementation, and they would greatly facilitate post-estimation life!

Regards
Donihue

[EViews Gareth]

That isn't possible.

[donihue]

Gareth,

Looking into the BVAR and associated programmes, I see that one can extract the matrix of coefficients "cn" and the matrix of standard errors "covn" by simply not deleting them at the end of the BVAR programme (line 520). From these, it seems (to a naive non-programmer!) that one could build up a simulation model à la the standard VAR command. Undoubtedly it is more complicated than this ...

Regards
Donihue

[EViews Gareth]

I'm not sure I understand the question there.

[donihue]

Gareth,

Sorry if I wasn't clear: this was in reference to my earlier query on this topic:

In particular, would it be possible to set it up to create Systems and Models directly from the BVAR results? These options do not seem to be present in the current implementation, and they would greatly facilitate post-estimation life!

to which you replied

That isn't possible

Not being a programmer, I wondered if the fact that the necessary matrices (cn and covn) can be made available would make it possible, albeit with some further programming. I guess the operative word in my earlier post was "directly".

Regards
Donihue

[EViews Gareth]

Yeah, I see what you mean.

The nature of Add-ins are that you cannot add things to the menus, nor do anything once the add-in has executed. Thus it would not be possible to add stuff to the Proc tab. However, I take your broader point. It would be possible to, at the time of estimation, say "estimate this BVAR, and please make a model from the results". I have Esther looking into this possibility at the moment.

A system cannot be done, since a system will re-estimate the VAR. You cannot create a system object and specify the coefficient values.

[donihue]

Many thanks, Gareth. The "model from results" option would be very helpful indeed.

Regards
Donihue

[EViews Esther]

As requested, the "model from the results" option is added.

Upon running the add-in from the menus, a dialog which contains a new "Model object" option will appear.
Note that the "Model Object (optional)" box lets you specify the name of the model object.

[donihue]

Wonderful! Thank you, Esther. I have tested the new "model object" option and it works perfectly.

I am also grateful to you for correcting the display issues mentioned in other messages; both tables and graphs of IRFs now display names correctly. There remains one small issue with display: as I mentioned in an earlier post

The IRF graphs display is too small to read text characters when several variables are used. Would it be possible to have an option to show the IRF graphs by equation?

This would also be useful for publications, where one would like to extract just certain IRFs.

And whilst still on this subject (sorry!): is there any chance of having the IRF graphs display a +/- 1 (and/or 2) std. error band?

Many thanks again.
Regards
Donihue

[EViews Esther]

Dear Donihue,

The IRF graphs display is too small to read text characters when several variables are used. Would it be possible to have an option to show the IRF graphs by equation?

The smart EViews gives you an option, "Extract selected graph", by right-clicking on the IRF graphs.

And whilst still on this subject (sorry!): is there any chance of having the IRF graphs display a +/- 1 (and/or 2) std. error band?

I will put it on my to-do list.

Best,
Esther

[donihue]

Many thanks Esther for the std error to-do listing.
Regarding the graphs, I had previously extracted the complete IRF graph, using the very "smart" extract option, but had not seen that I could then also extract the individual graphs. Thank you for that tip!
Regards
Donihue

[Sanke Charmer]

Hello Eviews people.

It seems to be a little and unimportant mistake in the BVAR add-inn: on the "Koop-Korobilis Minnesota" options it says: "a2 - Relative CORSS variable weight". I guess it should be "CROSS" (?)

Thanks for always solving my questions.