Re: [Gretl-users] Out-of-sample forecasting

On Fri, 9 Nov 2012, Marta Szymańska wrote: This should be intended as a reply to Tomasz too. Here's a variation on your script that should work as intended: <hansl> include gig.gfn open djclose.gdt RETURN = ldiff(djclose) model = gig_setup(RETURN,1,const) gig_set_dist(&model, 2) gig_estimate(&model) series e = model["uhat"] series hfc = model["h"] matrix coef = model["coeff"] a0 = coef[2] a1 = coef[3] # coef[4] is reserved for the asymmetry coefficient b1 = coef[5] # forecast the variance dataset addobs 50 setobs 5 1980/01/02 series hfc = ok(hfc) ? hfc : a0 + (a1 + b1) * hfc(-1) smpl 1989/09/1 ; print hfc --byobs gnuplot hfc time --time-series --with-lines --output=display smpl full </hansl> A few comments: * we use djclose in this example so everyone has it. * gig is an addon, so its "products" are not accessible through "$" variables. Instead, it uses bundles, so you may fetch them by ordinary bundle syntax; see the User's Guide and the gig documentation * when forecasting the variance, you don't want to use the square of the expectation of e as a predictor of e squared (Jensen's lemma): what you need is a predictor of e^2. If you use the expectation as your predictor, that's precisely what h is. As a consequence, in the simple case of the garch(1,1) model with normal errors, you just forecast h by its past values (for more complicated models, it's not so easy). -------------------------------------------------- Riccardo (Jack) Lucchetti Dipartimento di Economia Università Politecnica delle Marche (formerly known as Università di Ancona) r.lucchetti(a)univpm.it http://www2.econ.univpm.it/servizi/hpp/lucchetti --------------------------------------------------