First denis joubert wrote: > I was playing with gretl gui when i got an "undefined error" on > arch (and some other model too) with this serie of lags : > 1 2 3 4 5 6 7 8 28 37 41 42 50 52 58 72 87 97 129 > with the data i joined with this mail. > I think it's a bug.
Then Jack Lucchetti replied: > Could you please be more specific about the problem and give us > a reproducible example? The following works ok here... Then denis said: > when you go into arch model you select d_v1 as the dependent var > then you click on the lags button you check the dependent > variable for lags then you check the alternative way to put lags > you paste my given lags into the list of lags then you validate > and run the model. > You got : "unspecified error" (on the last built cvs version under windows) > > Extra : if you find a way to predict the sign of any next value > with great accuracy (when learning from any previous) i would be > very grateful (and perhaps very generous if it could be scaled > to similar problem) because i think it's possible but very very > hard. I now reply: I agree it's a bug that you got an "unspecified error" message; it means that gretl's error-reporting system broke down. That should now be fixed in CVS. However, one can see why this might have happened. Your list of ARCH lags is quite extravagant. The 129th lag is an influence on the current error variance? It's hard to imagine a situation in which that would make sense. Also please note that ARCH estimation has nothing to do with predicting the sign of a forecast error. ARCH is purely concerned with the error variance; it is totally agnostic with respect to the sign. Allin Cottrell
