Concerning problem #1 (empirical autocorrelations), Johannes is right in
the sense that - with the current code - one can get an autocorrelation
that is not within [-1,1], because the covariance on the numerator does
not span the same sample than the variances on the denominator; the fix
that he suggests would fix this. On the other hand, estimators of the
autocorrelation coefficients are known to be biased in finite samples
(and there is no easy way of fixing this), so applying Johannes’ change
would just consist in moving from a biased estimator to another biased
estimator (and I don’t know how the bias is changed). These estimators
are only asymptotically consistent, and getting a weird value is the
indication that the sample is too small and that more simulations should
be done.
What is your opinion about this issue, and about the two suggestions
(options for no IRF files, and NaN/Inf treatment in estimation) ?
Let's compute the covariance on the minimal sample as suggested by
Johannes