2010/5/4 Allin Cottrell <cottrell(a)wfu.edu>: > > On Tue, 4 May 2010, denis joubert wrote: > >> 2010/5/3 Allin Cottrell <cottrell(a)wfu.edu>: >> > The test libgretl was using for "weights all zero" was not really >> > appropriate: it was that the sum of squares of the weight series >> > equals machine zero. I've now modified this in CVS and your >> > example runs. (The weights are extremely small, but not literally >> > zero.) >> >> thanks, i will get the cvs version :) >> > >> > However, the example is rather odd and this may explain gretl's >> > complaints. The auxiliary regression for hsk estimation produces a >> > perfect fit with the original dataset (T = 226), since you have >> > 151 usable observations and 151 regressors (75 lags, 75 squares of >> > lags, plus constant). Adding another observation gives one degree >> > of freedom to the auxiliary regression, which seems to help >> > matters. >> >> the 151 observations with 151 regressors is only an example of the >> problem i got, with 210 observations or more i got the same problem (I >> added 1 observations at a time, and more I add observations more the >> problem shows up) >> >> does it mean i don't need to use hsk and should use only the auxiliary >> regression function in order to predict next values ? > > No, the auxiliary regression is predicting the error variance, not > the dependent variable. I wonder, though, why you reckon you need > 75 lags in this model (other than, perhaps, to stress-test > libgretl!).
75 was a bit to stress libgretl (i stressed libgretl with 100 lags the first time i used it). But the datas are very hard to predict correctly, i was trying predicting using 100 lags with gretl gui, the datas are made from others and can be partially reversed. (the informations alone are very hard to be correctly predicted) But i successfully predicted some informations about the informations which made the datas correctly This attemps was not serious. So now i do the tries as a clockwork (but i tried 75 in order to reduce the amount of informations if it does work too i will reduce the number of lags). this try is connected this time to machine learning algorithm in order to recover the basis informations (the first time it was made with my brain and hands). > > Allin Cottrell > _______________________________________________ > Gretl-users mailing list > Gretl-users(a)lists.wfu.edu > http://lists.wfu.edu/mailman/listinfo/gretl-users >
