Re: [Gretl-devel] RFC: $sigma & co.
On Sat, 26 Apr 2008, Riccardo (Jack) Lucchetti wrote:
I've been working on the manual to update the matrix chapter.
Only now I realise that there is some inconsistency in the way
dollar accessors are used, with respect to variances/covariances
after estimation.
Typically, after a model is estimated, we use $sigma for the
residuals and $vcv for the parameters. However, this is not
always the case; some exceptions are well justifed, some less
well.
After estimating a system, $sigma holds the covariance matrix of
the residuals and $vcv the covariance matrix of the parameters
of the structural form; this is nice and consistent with, say,
ols, but is _not_ consistent with VAR/VECM estimation, where
$vcv holds the cov. mat. for the residuals and $sigma is unused.
Yup, the VAR/VECM case should be made consistent with the others.
the only question is the timescale. Sven commented not so long
ago on backward-incompatible changes; let's wait to hear from him.
In the meantime, however, I've made a change that I presume should
be uncontroversial: after estimating a VAR/VECM, $sigma now
retrieves the covariance matrix of the residuals (so does $vcv,
which we'll change at some point).
Moreover, we have an accessor ($h) only used for garch
estimation, which does the same job as $sigma (which is unused
for garch models), when you take into account that the estimate
of the conditional variance is not a scalar but a series in
these models by their very nature.
One point here: in the garch case $h gets the conditional
_variance_, not standard deviation, while $sigma gets the
unconditional or steady-state standard deviation. This is
unusual, but then garch is unusual; I'm not so sure it's
inconsistent.
Allin.