2012/3/23 Allin Cottrell <cottrell(a)wfu.edu>:
On Fri, 23 Mar 2012, Riccardo (Jack) Lucchetti wrote:
> On Fri, 23 Mar 2012, Daniel Bencik wrote:
>> However, Im gettin into a problem (again!). The estimation does not
>> converge. For some simpler specifications ( I mean simpler than
>> ARMA(7,5)-GARCH(1,1)-t ) I was able to help the algo to converge by setting
>> different parameter starting values. However, with this complicated model,
>> even when I set the initial values equal to "true" values found by
>> the algo still does not converge.
> With all due respect:
> 1) Unless you have over one billion observations, then estimating an
> ARMA(7,5) model is asking for trouble
> 2) I wouldn't put my life in the hands of Eviews' optimising algorithms.
To expand on Jack's point 2) a little: it's possible that
Eviews' algorithm is more effective in finding the MLE in this
case, but it's also possible that Eviews is using a relatively
sloppy criterion for reporting a claimed MLE. Before
concluding that Eviews is giving you "true" values, you should
look at the gradients at "convergence".
In gretl, use the --verbose switch with the mle command to see
what's happening in detail. You might -- or might not -- see
that gretl is finding a log-likelihood comparable to Eviews'
maximum, but is not reporting convergence based on the
gradient criterion. If you find that Eviews is in fact
reaching a proper solution with a greater log-likelihood than
gretl is able to find, that would be worth reporting back to
Daniel, are you sure about 'true-values' in EViews?
I mean, did the algorithm converge? In other words,
did it say 'convergence achieved'? If that is the case,
you could evaluate your LL with zero iteration and
see the value of gradients (Allin's suggestion).