Re: [Gretl-users] Using bfgs

On Sat, 24 Dec 2011, clarodina(a)lycos.com wrote: This is my last contribution on this point. It's not surprising that BFGS failed to converge since there's no guarantee that there's any finite maximum for this problem. Nor is there any guarantee (or even likelihood) that the "d1hat" arrived at in this way can be considered as a "fitted value" or prediction for d1. It may bear no resemblance to d1. I guess that in fact you want to minimize, not maximize, the ADF t-ratio (that is, maximize the strength of the rejection of the unit-root null hypothesis for the series in question, which requires a negative t-value). But that doesn't alter my comment above. Here's a sample script, but as I said this is my last shot on this topic: <hansl> function scalar adf_max (matrix *b, series d1, series d2, series d3) series u = d1 - (b[1]*d2 + b[2]*d3) adf 0 u --nc --quiet # maximize negative tau scalar tau = -$test return tau end function nulldata 100 series d1 = normal() series d2 = normal() series d3 = normal() matrix b = {1,1} # set a sloppy tolerance set bfgs_toler 1e-6 tau_max = BFGSmax(b, adf_max(&b, d1, d2, d3)) print b series d1hat = b[1]*d2 + b[2]*d3 series u = d1 - d1hat adf 0 u --nc # are d1 and d1hat related? corr d1 d1hat </hansl> Allin Cottrell