[Gretl-users] Re: fitted actual values over time problem in CVS

On Tue, 30 Dec 2008, John Paravantis wrote: Urgh, yes, and get a big mess! This is now fixed in CVS and the Windows snapshot. (This was a newly introduced bug, and nothing specifically to do with nonlinear least squares -- it would have messed up any fitted vs. actual plot -- so thanks for spotting it!) A little more on what I've done with NLS and forecasting. We now check whether or not an NLS equation is dynamic (i.e. contains one or more lags of the dependent variable on the right-hand side). If so, we offer a dynamic forecast out of sample, but we do not produce any forecast standard errors / confidence intervals, since I don't know how we could produce defensible values without bootstrapping (which would require a fair amount of work). If the model is _not_ dynamic (as in your cars example), we produce forecast standard errors as simply the standard error of the NLS regression (abstracting from parameter uncertainty), and confidence intervals that are point forecast plus/minus t(df, 0.025) * SE. Allin.