[Gretl-users] The "fcast" command and panel data

This is a follow-up to Artur's comment/question in http://lists.wfu.edu/pipermail/gretl-users/2019-February/013703.html Use of "fcast" with panel data is now enabled (to varying degrees) for pooled OLS, fixed effects, random effects and dpanel. By "now" I mean in git and the snapshots. Preliminary note: if you want to produce out-of-sample predictions using panel data you should ensure that any required transformations (lags, differences, etc.) are generated for the entire dataset before a model is estimated on a sub-sample. Within sample forecasts Producing within-sample "forecasts" is unproblematic, if in general not very interesting since the $yhat accessor for fitted values is already available. However, there's modest added value from "fcast" in a couple of respects: you can get a printout of the forecast evaluation statistics; and for random-effects estimation the values produced by fcast include an estimate of the individual effect, which is omitted from $yhat. Out of sample forecasts In the panel data context "out of sample" could mean either in the time dimension or in the cross-sectional dimension. As of now, gretl's automatic --out-of-sample option operates strictly in the time dimension, but it's possible to forecast out of sample in the cross-sectional dimension via the stobs/endobs arguments to "fcast" or by setting the sample range appropriately before calling "fcast". I'll deal with the --out-of-sample option (time dimension) first. This is subject to one restriction and one caveat. Restriction: the model on which forecasts are to be based must be estimated using all individuals in the dataset (i.e. sub-sampled only with respect to time). This is not a very severe limitation: one can always save, then open, a reduced copy of the dataset (sub-sampled in the cross-sectional dimension) first, if necessary. Caveat: It's not clear what we should do if the model specification includes time dummies. At present we impute an out-of-sample time effect equal to the mean of the in-sample time effects. This is obviously debatable; feel free to debate it. Perhaps we should ban such prediction altogether. In addition, a refinement: for models estimated via "dpanel", the --out-of-sample option produces a dynamic forecast by default (as with dynamic estimators on straight time-series data). This can be interdicted by use of the --static option to "fcast". Now to out-of-sample in the cross-sectional dimension. What are we going to do with individual effects, for the fixed- and random-effects estimators? For now, we set the predicted individual effect to the global constant (for fixed effects) or to its expected value of zero (for random effects). This is obviously another debatable point, analogous to the out-of-sample time effect case. Besides debatable policies there are no doubt some bugs lurking in the new fcast facilities. Comments welcome. Allin