Re: [Gretl-devel] Possible ARMA CMLE bug

On Fri, 11 Apr 2014, GOO Creations wrote: > I'm not sure if this is a gretl or cephens bug. I've attached my test data of > 1024 samples, in case the problem needs to be reproduced. > > Using an ARIMA(17,0,2) model with conditional MLE and the given test data, > gretl gets stuck in an infinite loop. Just before the infinite loop starts, > the following stderr is printed: > > /arma: MA estimate(s) out of bounds > cholbeta: test[18] = 0// > //gretl_matrix_QR_rank: rcond = 9,94709e-20// > //dropping redundant variable 18 (e_1)// > / > > > After further investigation, I discovered that the infinite loop starts in > plugin/arma.c:167 were the following cephens function is called: > > /cerr = polrt(b->temp, b->tmp2, qtot, b->roots);// > > / > > It seems that the while-loop in cephens/polrt.c is executed over and over > again. Inside the loop there is a "goto tryagn;" statement which makes the > function go back before the loop and then execute the while-loop again. This > happens over and over again (stuck in infinity). > > This problem is specific to the given test data, AR order 17 and MA order 2. > If I have different data or different orders, the problem disappears. Does > anyone know what's going wrong? There does appear to be a potentially infinite loop in cephes' polrt.c, and that's now modified in CVS. It seems the problem was triggered only by a "broken" model, where the roots of the AR polynomial are on or inside the unit circle. Allin Cottrell _______________________________________________ Gretl-devel mailing list Gretl-devel(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-devel