[Gretl-users] Re: multi-step-ahead-forecasting

On Thu, 7 Jan 2021, Sven Schreiber wrote: The problem here is that "fcast t1 t2" doesn't start the forecast machinery until t1. So when attempting a dynamic forecast according to fcast 100+i+h 100+i+h # with h > 1 gretl first looks for a prior forecast of y at obs 100+i+h-1, but doesn't find one. Here's a version of the script which works properly: <hansl> open djclose.gdt set verbose off series frcst = NA h = 5 loop i=1..50 -q smpl 1+i 100+i ols djclose const djclose(-1) fcast 100+i+1 100+i+h frcst[100+i+h] = $fcast[h] endloop </hansl> Now you might say that gretl should be clever enough to realize that it needs to forecast the intermediate values (at least if the --dynamic option is given), even if they will be discarded in the output, but right now that doesn't happen. Why did the original version "work" with the --dynamic option and only complain about missing values if that option was not given? A bug: with --dynamic, gretl fell back on using actual y in place of the missing forecast lag (so in fact giving a static forecast) but in the "auto" case that didn't happen -- which is probably the reverse of what should occur. Stuff needs fixing here, but note that the problem is specific to the case where there's a temporal gap between the estimation sample and the start of the dynamic "fcast" range. Allin