Re: [Gretl-devel] constant mis-specification and arima convergence problem

I have found a sizable quantity of mis-specified non-converging models due to vicious feedback loop: to detect seasonal unit root I use OSCB test inside auto_arima{lagreg} this test 1) construct some artifical variables 2) use armax(3 0 0; 1 0 0) to compute some statistics if the latter do not converges it uses armax(3 0 0; 0 0 0) compute some statistics etc the statistics from the last convergent model is compared with simulated critical value Non-convergence tends to cause detecting seasonal unit roots But 197 may non-converge due to redundant constant So it detected seasonal unit roots for season-adjusted data With redundant seasonal differencing again a problem of redundant constant evinced I propose a palliative : if in catch arima ... $error > 0 to create a new extractor, say $badmodel containing: final parameter values, lnl, gradient norm, iterations number, abs values of polynomial roots corresponding to final parameters and standard errors if possible This information can be useful for finding better specification Oleh P.S. I'd like to have abs of pol. roots as a member of $model for standard situation 22 жовтня 2018, 08:59:50, від "Riccardo (Jack) Lucchetti" <r.lucchetti(a)univpm.it&gt;: