Re: [Gretl-users] About the automatic ADF lag length selection

On Fri, 20 Feb 2009, Ignacio Diaz-Emparanza wrote: > I remember that in the time when I was using RATS, there was an > interesting procedure "lagselect" which served to decide how to > determine the lag. You can have a look at it in > http://www.estima.com/procs_perl/500/lagselec.src Thanks for the reference. I have hacked up a little script to play AIC, BIC and HQC against gretl's built-in "testing down" procedure (attached). Perhaps some people might like to play with it. Here's one set of results, for the case where the original error process is AR(3) and we test down from a maximum lag of 6. Specifically: y = 0.02 + 0.9*y(-1) + u u = 0.5*u(-1) - 0.3*u(-2) + 0.2*u(-3) + e e = normal() test regression: ols dy const y(-1) dyL where the list dyL contains successively 6, 5, ..., 1, 0 lags of dy. Results: 1000 replications, test down from 6 lags Lag order selected: mean on row 1, s.d. on row 2 AIC BIC HQC z-test 5.85 3.38 4.97 3.16 0.46 1.48 1.26 1.39 The built-in z-test procedure looks pretty good, as does BIC, but AIC and HQC seem to select too many lags. AIC selects the maximum lags available most of the time. This is emphasized if we put the max lag order (from which we test down) to 8 while leaving everything else unchanged: 1000 replications, test down from 8 lags Lag order selected: mean on row 1, s.d. on row 2 AIC BIC HQC z-test 7.85 3.95 6.64 3.98 0.45 2.07 1.73 2.12 The script is set up so that you can easily tweak the parameters governing y and u. Allin._______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users