Re: [Gretl-users] The "fcast" command and panel data

> fcast --out-of-sample --quiet

On Tue, 26 Feb 2019, Sven Schreiber wrote: > Am 26.02.2019 um 20:06 schrieb Allin Cottrell: >> BTW, I see that Stata is very restrictive in this department. With >> "predict" after "xtreg" you can't get an out of sample forecast even in the >> time dimension when no time dummies are present -- other than plain "xb" >> which omits individual effects, whether fixed or random. > > Well for small-T (standard micro panel assumption) the unit > effects are not well estimated, are they? That might be the reason > they ban it. (And probably a good reason.) Interesting point. I doubt that's the rationale in Stata's case, since the "ban" is not conditional on the time-series length of the panel. I suspect it's because Stata doesn't really make a distinction between out-of-sample in the cross-sectional dimension and out-of-sample in the time dimension: it's all just in or out of sample. But whatever Stata's up to, it's worth assessing whether individual fixed effects estimated with small T are actually any use in forecasting out of sample in the time dimension. I show below a Monte Carlo experiment, comparing out-of-sample MSE for forecasts computed with and without individual effects. (In the "without" case I use the global constant, as in Stata's "xb" forecast for fixed effects.) I simulate 100 individuals observed over T periods, T-1 of which are used in estimation leaving the last for an out-of-sample forecast, with T = 4 as baseline. Executive summary: As Sven suggests, if T is small enough (and the fixed effects are moderate in size, even if statistically significant at the usual level on a joint test), then the estimated effects are just noise so far as forecasting is concerned; the forecast MSE tends to be smaller without them. But if T is a bit larger (even just 5 as opposed to the 4 in the script), or if the scale of the fixed effects passes a certain threshold[*], then inclusion of the estimated fixed effects enhances the forecast accuracy quite substantially. [*] in the script below, with fixed effects simulated by drawings from U(0,1) * fe_scale, an fe_scale value of 2.5 is sufficient to make inclusion of the estimated fixed clearly beneficial, even with a scanty T = 4. <hansl> set verbose off # set seed 87564427 T = 4 N = 100 NT = N*T nulldata NT --preserve setobs T 1:1 --stacked-time-series genr time series x = normal() list X = 0 x series u = normal() # generate scaled uniform random fixed effect fe_scale = 2.5 matrix am = fe_scale * muniform(N,1) series a = pexpand(am) K = 10000 matrix MSE = zeros(K, 2) matrix fepv = zeros(K, 1) loop i=1..K -q series u = normal() series y = 5 + x + u + a smpl time < T --restrict panel y 0 x --quiet fepv[i] = $model.fixed_effects_F.pvalue matrix b = $coeff fcast --out-of-sample --quiet smpl time == T --restrict --replace series fca = $fcast series fc0 = lincomb(X, b) MSE[i,1] = sum((y - fca)^2) / $nobs MSE[i,2] = sum((y - fc0)^2) / $nobs smpl full endloop printf "Forecast MSEs\n" means = meanc(MSE) sds = sdc(MSE) printf "Including fixed effects: mean %g, sd %g\n", means[1], sds[1] printf "Excluding fixed effects: mean %g, sd %g\n", means[2], sds[2] printf "memo: average p-value for fixed effects: %g\n", meanc(fepv) </hansl> Allin _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users