Forecasting out of sample using ARMAX model with a specific lag (eg. lag 2, anything but one) saved from a previous gretl session produces mumbo-jumbo. To replicate build a model then save to session as an icon. Save session file. Close Gretl. Open Gretl and previous session file. Forecast ahead and you will get wrong predictions. I have attached a session file that show the results. Model 3(lag 1) works but Model 4 and 5 do not. If you redo model 4 or 5 you will get correct forecasts.