On Sat, 16 Jan 2016, Riccardo (Jack) Lucchetti wrote: > On Sat, 16 Jan 2016, Sven Schreiber wrote: > >> Am 15.01.2016 um 20:39 schrieb Allin Cottrell: >>> Following up Jack's comment at >>> >>> http://lists.wfu.edu/pipermail/gretl-devel/2016-January/006467.html >>> >>> in current git there's a basic "preview" of Julia support in gretl. >> >> exciting! >> >> >>> # NIST's certified coefficient values >>> matrix nist_b = {-3482258.63459582, 15.0618722713733, >>> -0.358191792925910E-01, -2.02022980381683, >>> -1.03322686717359, -0.511041056535807E-01, >>> 1829.15146461355}' >>> >> >> Since I don't have it installed yet, could you comment on whether the >> results match (between gretl/Julia/NIST)? > > These are the results I get > > <output> > Log-relative errors, Longley coefficients: > > gretl julia > 12.228 8.0224 > 10.920 7.5300 > 11.797 7.5697 > 12.528 8.1421 > 13.169 8.3801 > 11.770 7.2368 > 12.235 8.0333 > > Column means > 12.092 7.8449 > > </output> > > So it would seem that the MultivariateStats julia module leaves a > bit to be desired for the moment, at lest in terms of precision.
My test was admittedly kinda silly, in that there's not really any reason to delegate to a "foreign" program stuff that gretl handles well natively. One would be more likely to get Julia to do MCMC or the like. That said, among the various "foreign" programs on which I've tried the notorious Longley exercise, only numpy comes close to gretl for numerical precision. However, Anders makes a fair point in saying that the statistical error (and I would add, data error) swamps the numerical error for this sort of linear problem. Allin
