A quick clarification. With the command “first_obs” in the estimation line, the manual says this is “The number of the first observation to be used”. Does this mean that all previous observations are simply discarded? Or is it the first period in which filtered/smoothed/forecast variables are estimated, but all previous observations are available for training sample? And how does the latter relate to the presample command?

Code below (largely inherited), if that helps. Thanks in advance.

estimation(optim=('MaxIter',200),datafile=usmodel_data_rec,mode_compute=4,first_obs=71,presample=4,lik_init=2,prefilter=0,mh_replic=0,mh_nblocks=2,mh_jscale=0.20,mh_drop=0.2,bayesian_irf,irf=0,forecast=1,nograph) labobs robs pinfobs dy dc dinve dw ;

Yes, all observations before this are discarded. The presample option then specifies the number of observations after firstobs that are used for training the Kalman filter

A further - probably unrelated question. Using similar code to before but with filtered_vars added

gives me two sets of forecasts for the observables, one **filtered **(parameters estimated over full sample 1:T but using only observables 1:t) and one **recursive **out of sample (parameters and data both using only subsample 1:t).

In principle then, as t=>T the two sets of forecasts ought to be the same, in particular the final observations when the two sets of estimated parameters and information sets are identical. However, this is not the case: if anything they are moving further apart. Can you explain how this can possibly be the case?