Hello Allin,

If we're still talking about the ARCH(1) model for the NYSE "close"
value, it ain't goint to converge, it's too badly misspecified. With
some datasets and specifications the limiting factor may be
calculation of the parameter covariance matrix, but that's not the
case here.

​I see.

In case of trouble with GARCH convergence I recommend checking the
box marked "Show details of iterations" (or using the --verbose
option on the command line). This gives you some diagnostic
information. Gretl claims "convergence" only if it finds a high
point in the admissible parameter space where the norm of the
gradient of the loglikelihood is acceptably small, and that's not
happening in the NYSE example. (Eviews apparently doesn't care about
the admissibility of the GARCH parameter values; I don't know
whether it bothers to calculate the gradient, or just stops when the
loglikelihood "stops changing".)

​This is exactly what I was suspecting. Either there was a problem with 
Gretl to converge, or Eviews spits out whatever or any outcome instead
of a meaningful one.​

Anyway, in the NYSE case you'll see this message:

"Convergence was not reached.  One possible reason for this is
autocorrelation in the error term.

After estimating the model by OLS, the following result was
obtained for a test of autocorrelation of order 52:
LMF = 2.30988, with p-value 4.68768e-07"

​I will try to reproduce this.​

Gretl will not try to "keep you happy" by claiming convergence when
it is not obtained.

​This is great Allin! thank you lots!​


Gretl-users mailing list