01-30-2013 08:18 AM
I use SAS Forecast Studio 4.1. I have highly seasonal weekly time series, with a seasonal cycle of 52, covering 3 entire years. When I run a project, restricting the model choice to ESM, the clear seasonal pattern is rarely picked up.
I then played with the Diagnostics settings, notably the Seasonality test. According the to help file, "a sensitivity value of 0 will imply always seasonality". I thus activate the test and set the value to 0, however, this does not work. I again obtain many non-seasonal ESM models. If I create myself the "Best Seasonal Model" then I get a good looking forecast, often with lower MAPEs.
I then read on the HPF user guide that "a smaller value of the SIGLEVEL= options means that stronger evidence of a seasonal data is required before PROF HPFENGINE uses seasonal models". Isn't this a contradiction with the Forecast Studio help ? As I want to force seasonality, I thus should set the value to something close to 1. But this doesn't work neither.
Does anybody has more detailed info and instructions to force the use of seasonal models (ESM, ARIMA, UCM) in SAS Forecast Studio ? Thanks !
02-01-2013 04:16 PM
Hello Marcel -
For your consideration: in "Forecast Settings" add "SASHELP.HPFDFLT.BESTS" as "models from an external list":
This will force a couple of seasonal ESM to be run as part of your project.
For details on BESTS please have a look at: http://support.sas.com/documentation/cdl/en/hpfug/63959/HTML/default/viewer.htm#hpfug_hpfdiag_sect01...
02-04-2013 02:44 AM
Thanks, Udo, for the quick reply. This works perfect to me.
Curious however to hear more from you on the seasonality test, and why I struggle to make it work without using the BESTS repository.
02-11-2013 11:41 AM
Hello Marcel -
Please excuse for delay in responding - in order to test for seasonality Forecast Studio fits different autoregressive models behind the scenes and picks the most appropriate one based on AIC. For weekly data (seasonality of 52) the AIC penalizes models having too many parameters and a non-seasonal model will be preferred.
As I have mentioned before - if you have knowledge about the seasonal behavior of your data, you are better off including seasonal models in a custom repository (for example BESTS).
Happy to continue this discussion with you directly.