Holdout MAPEs in SAS PROC ARIMA and SAS Forecast Studio don't match

[Xman]

I have a Time Series (ARIMA) model in SAS, modelled using proc ARIMA which I am trying to replicate in SAS Forecasting Studio. What I see is that The parameter estimates in both are very similar (which is surprising, since my colleagues haven't had such luck, but I digress), but the MAPEs calculated with a Holdout of 7 months does not match. I would understand the parameters not matching, with the numerical optimization algorithms giving different results with different initial values, but the MAPEs not matching is very puzzling. Has anyone else here come across this, or has any idea what might be the case? DO you need any more information to address this issue?

Thanks, V

[udo_sas]

Hello -

Different to PROC ARIMA SAS Forecast Studio supports both concepts of hold-out sample and out-of-sample data. In a nutshell: hold-out sample data are used during the model selection process, out-of-sample data are not.

I presented at SAS Global Forum 2009 some slides, which may help to explain the 2 concepts in more detail (http://support.sas.com/resources/papers/proceedings09/316-2009.pdf). Of course you will find similar information in SAS Forecast Server documentation.

PROC ARIMA only supports out-of-sample data using the BACK option - so you will need to make sure to compare similar statistics.

If in doubt please contact Techical Support for some guidance.

Thanks,

Udo

[Xman]

HI Udo - Thank you for your helpful comments!

I guess I have covered my questions on the back= option in another post so will not duplicate it here.

The link you provided has only the abstract of the talk... could you please guide me to the full presentation? That would be very helpful!

Many Thanks again!

Regards,

Xe

[udo_sas]

Hello -

Yes, I should have done this before but I was fighting with the upload options of our communities.

Here you go.

Thanks,

Udo