[Gretl-users] Re: Recursive forecasts for Arima models

On Sun, 24 Nov 2024, Artur T. wrote: [...] The basic problem here is the same one I indicated in my response to Sven. The independent variable is expressed in such a way that it's not apparent to gretl that it's just the first lag of the dependent variable -- so gretl doesn't know that it's a dynamic model. But in addition (unlike Sven's case), in this example the generation of log differences is confined to a subsample, so you can't get any forecasts out of sample. You seemed to indicate that if the following statements were added to the script above <hansl> smpl 2017:1 $tmax series fc = $fcast print fc -o </hansl> you would see two forecast values. But I think you'll find if you append those statements to the original above and run the whole script from scratch you get ? series fc = $fcast The statistic you requested is not available Terminated on error You also said you'd expect to see "only a single h=3 forecast value at date 2017:3". I can understand why that might be expected (modulo the issue over the independent variable in the regression). But actually if you request a k-step recursive forecast with the --out-of-sample option, gretl moves the end of the first estimation sample back such that the first out-of-sample observation is k steps ahead of the last data used in estimation (if possible). That's no doubt debatable, and may not be totally in agreement with the help text, but that's what it does. Allin