Hi! Thank you very much for your answers. I will try to forecast first differences and I also started some experiments with forecasting log transformed data. This mor or less solved he problem with values below zero. I think you are right about my model isn't fitted as good as I thought. I managed to model sales for 4 years very accurate (std dev of error was like 8% with avg error 0%) with ARIMA(2,1,2)(1,0,2) but then the forecast I got was below zero. When I changed one parameter ARIMA(1,1,2)(1,0,2) I've got almost as accurate model but forecasts for the next 3 months twice as big as the last values. So... something is definitely wrong. The problem I've got is that i have to forecast something like 1000 different products, which are in 60 different groups (collections). So i figured out, that I can forecast those collections and then overlay forecasts for collections on particular products because percentage share of products in collection are more likely the same in time. Thanks to that i can use more sophisticated methods of forecasting (I wouldn't been able to forecast 500-1000 items in gretl because of lack in time)... Once again, thanks for your answers! 2010/9/30 MICHAEL BOLDIN <mboldin(a)temple.edu&gt;: