Am 05.02.2024 um 14:22 schrieb Artur T.:
Am 05.02.24 um 14:19 schrieb Sven Schreiber:
>
> In any case, from what I read in the first couple of pages of the
> NBER WP version of the paper above, all you seem to need is the
> forecast error variance decomposition (FEVD) of a Choleski-identified
> VAR. (Please correct me if I overlooked something there.)
>
> So in principle you just would have to estimate the VAR (gretl's
> 'var' command), and then you can retrieve the FEVD values in the
> accessor '$fevd' (after setting the calculated forecast horizon; 'set
> horizon <whatever>'). With the FEVD values, you then have to do some
> summing and normalizing, AFAICS. The $fevd matrix apparently already
> contains the respective contributions as fractions (as per the help
> text, which you should read in any case), but if you sum over several
> horizons, I guess you still have to do that part of the normalization
> step.
>
> All in all, I guess it's a very doable exercise in VAR-oriented hansl
> scripting, because the difficult parts are already taken care of
> natively.
Sven is right, the main apparatus already exists. Only the generalized
fevd (Pesaran & co authors) -- which is very popular in this
literature -- is not available as a package even though some people
(including me) might have some functions on their machine.
Yes, I'm guessing that the Diebold&Yilmaz approach puts the target
variable on one end in the Choleski ordering. So if you want to
calculate the indices for all the N contained variables, you would have
to estimate N different FEVDs, changing the ordering each time. This is
pretty much what the generalized IRFs do, too.
So you basically put an N-loop around the 'var' estimates.
But again, I haven't read the paper until the end...
-s