Am 06.05.2017 um 03:03 schrieb Allin Cottrell: > I've recently had cause to hit on gretl's fcstats function and I found a > couple of limitations which are now fixed in git and snapshots. > Nice incremental features, thanks. While we're at it, don't have time to check the status quo right now, but there was this thing with fcstats() that somehow the way some of the measures were calculated (Theil?) they weren't defined for some data constellations (if the target was all zeros perhaps???). I don't think this was changed in the last one or two years, was it?

Am 09.05.2017 um 01:28 schrieb Allin Cottrell: > On Mon, 8 May 2017, Sven Schreiber wrote: >> While we're at it, don't have time to check the status quo right now, but >> there was this thing with fcstats() that somehow the way some of the >> measures were calculated (Theil?) they weren't defined for some data >> constellations (if the target was all zeros perhaps???). I don't think >> this was changed in the last one or two years, was it? > > No, the laws of arithmetic have not changed in this universe in the last > year or two ;-) I suppose we could lop off the undefined terms in fcstats > (return a shorter vector) rather than recording NaNs, if we thought that > was preferable. Well, I have re-freshed my memory now, but it actually wasn't so bad, it really did have to do with zeros: Since gretl uses the "U2" measure, the y_t appear in the denominators so it breaks down if some y_t are zero. This is of course not gretl-specific. What I meant is that this IMHO is an unfortunate shortcoming, because realizations of zero are reasonable for many applications (GDP growth...). U2 really seems to make sense only for positive variables. So my question is, perhaps gretl's fcstats() should be extended with the U1 measure? (U1 being basically, using some obvious notation: numerator = sqrt(mean((my-mfc).^2)) denom = sqrt(mean(mfc.^2)) + sqrt(mean(my.^2)) out = numerator/denom -- This only breaks down if all observations are zero and if those are all correctly and exactly forecast.)

On Mai 5 2017, at 10:03 pm, Allin Cottrell <cottrell(a)wfu.edu&gt; wrote: (...) Problem 2: OK, but I'd like to provide several columns of forecast values -- can't you give me a return matrix with several columns of forecast evaluation statistics? Fix 2: We now support this. For the actual observed values, y, we still require a single series or vector, but if you give as the second argument, f, a matrix with n columns, we return a matrix with n columns, each of which contains forecast evaluation statistics for the corresponding column of f. (The number of rows of f must, of course, match the number of values in y). (...) Dear Allin, very nice implementation! Is it possible to include lists as thesecond argument? [68cbb5b0f4586c0c8c889750918ab411ba7c6fd0474c8443ff21cfc4e75cf00d]