On Wed, 20 Jun 2012, Skipper Seabold wrote: > On Wed, Jun 20, 2012 at 1:16 PM, Skipper Seabold <jsseabold(a)gmail.com> > wrote: >> On Wed, Jun 20, 2012 at 12:29 PM, Allin Cottrell <cottrell(a)wfu.edu> wrote: >>> >>> On Wed, 20 Jun 2012, Skipper Seabold wrote: >>> >>>> I was playing around and I think I noticed a bug in ARMA prediction, or >>>> I >>>> am missing something. You can find the data to replicate here >>>> >>>> http://www.pastie.org/4120330 [...] >>> >>> (Annual sun activity data, 1700 to 2008, attached in gretl >>> format for reference) >>> >>> Thanks, you're right. There was an off-by-one error in the >>> code that decides when a dynamic ARMA forecast can start using >>> previously computed forecast values in place of the observed >>> data. In your case, with an AR order of 9, gretl didn't start >>> using computed values until the forecast for 1711, whereas it >>> should start with the forecast for 1710. >>> >>> With this point corrected (in CVS), gretl agrees with Stata. > > Out of curiosity does this only affect AR models estimated with ARMA > or should I go back and check my ARMA(p,q) dynamic forecasts that > start in-sample as well?
I believe the problem would affect all ARMA models with a non-null AR component, where the --dynamic flag was passed to request a dynamic forecast from the earliest possible date: such forecasts would go dynamic one period late. Allin Cottrell
