Re: [Gretl-devel] VARs as system special case ($sysA, $sysB, etc.)
On Fri, 30 Oct 2015, Sven Schreiber wrote:
Hi,
I propose to treat VARs more fully as a special case of a structural system
of equations. This would imply to make the accessors $sysA and $sysB
accessible, and probably a few other things as well. If you want I could
write down a complete list.
$sysA is already quite easy I guess, as $compan[1:N,]. However, I don't see
why users must be expected to know this trick and not directly provide the
endogenous lagged coefficients. And for $sysB it's more involved.
And BTW, in the built-in help the definition of $sysA is not totally clear,
whether the ordering is by variable or by lags. Only the section in the user
guide explains it 100%.
I could imagine that such stuff need not be implemented at the C level in
principle, but it's not obvious (to me) how the interface should look like to
come as close to the $-accessors as possible.
For VECMs the situation is maybe less clear, but given that you specify it by
using the levels of the variables and that $compan also exists there, I tend
to think it could be covered just as well.
Opinions?
A tiny linguistic point first. I guess English differs from many
languages in this respect, but in English usage "I propose to X"
basically means "I plan to X", but with the implication that you're
seeking approval first, hence: "I plan to X unless there are any
objections to my doing so". I don't have any objections to what
you're proposing ;-)
Anyway, what you're suggesting makes sense. It would help if you
could come up with some sort of hansl prototype or even reasonably
specific pseudo-code. Then I'd be happy to try to make C out of it.
Vaguely related: I've recently been thinking that our "system"
apparatus could do with a revamp. I've got interested in Ray Fair's
arguments about macroeconomic modeling and I think it would be good
if gretl could handle a "realistic" macro system such as the Fair
Model. (I notice that he offers a version of FM that runs on
Eviews.) We're not a million miles away from being able to do that,
but there are some gaps, notably the fact that we currently handle
identities only in the FIML case (and it seems that practical macro
structural modelers typically use 2SLS in preference to fancier but
possibly more fragile estimators).
Allin