Jack, just to clarify, my query was about using the forecast option with a panel dataset, for forecasting within the sample time period range in each panel unit, in which the forecast is the fitted value. The graphics function treats contiguous observations in adjacent panel units and plots them as such when calling up Analysis; Forecast from a pooled OLS model with DVs, thereby making the graph line for the dependent variable continuous. The forecast function is not available when specifying FE or RE Panel models. Brian
On Wed, 28 Feb 2024, 20:34 Riccardo (Jack) Lucchetti, < p002...@staff.univpm.it> wrote: > > OK, this message is meant to summarise a few things we discussed in > today's meeting for the benefit of Artur and Sven and also, so that we > don't forget. > > The starting point was B. Revell's message on out-of-sample forecasting > in panel datasets, which spurred a discussion on several different > points (all panel-related): > > * out-of sample forecasting for panel estimators _in the time dimension_ > may be tricky for FE and/or dynamic models, so maybe we could start from > providing this facility for static models estimated with random effects, > and proceed from there. > > * forecasting _a unit_ may pose different problems. Again, this should > not be a problem if the estimator is RE (and the model is static); what > we should do if the estimator is FE is less clear and we may want to put > this aside for the moment. > > * related to out-of-sample forecasting is the issue of extending the > "dataset addobs" command to panel datasets. At present, this works by > appending empty units, provided the parameter is an integer multiple of > $pd, but it may be interesting to provide a way to add time periods. It > could be nice to make this the object of a collective coding session > next week. This would also have the benefit of illustrating some > characteristics of the DATASET struct, that appears pretty much > everywhere in libgretl. > > * we may provide a new option to the panel command to have CRE > estimator, which is basically RE with cross-sectional means added and > returns the FE estimator for time-varying regressors. I find it of some > pedagogical value (J. Wooldridge is quite a fan), and besides, it > implies very little extra computational effort from what we do already. > In fact, I'm providing an example in my book that the attached script > illustrates. > > ------------------------------------------------------- > Riccardo (Jack) Lucchetti > Dipartimento di Scienze Economiche e Sociali (DiSES) > > Università Politecnica delle Marche > (formerly known as Università di Ancona) > > r.lucche...@univpm.it > http://www2.econ.univpm.it/servizi/hpp/lucchetti > > -------------------------------------------------------_______________________________________________ > Gretl-users mailing list -- gretl-users@gretlml.univpm.it > To unsubscribe send an email to gretl-users-le...@gretlml.univpm.it > Website: > https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/ >
_______________________________________________ Gretl-users mailing list -- gretl-users@gretlml.univpm.it To unsubscribe send an email to gretl-users-le...@gretlml.univpm.it Website: https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/
