Forecasting using SAS Forecast Server, SAS/ETS, and more

SAS/FS and SMAPE fit statistic

Reply
Occasional Contributor
Posts: 9

SAS/FS and SMAPE fit statistic

Hi guys,

I want to use SMAPE statistic to assess fit quality of models.

This statistic should be bounded between 0 and 200, cf. for instance

Symmetric mean absolute percentage error - Wikipedia, the free encyclopedia

Nonetheless, in SAS/FS, I do have values greater than 200; and this is not an issue about reconciliation as this also happens in a list project.

SAS_FS_SMAPE.png

Any idea?

Eric

SAS Employee
Posts: 416

Re: SAS/FS and SMAPE fit statistic

Hello Eric -

Some thoughts for your consideration:

There seem to be many different definitions of SMAPE available (SMAPE - Monashforecasting.com) - please note that SAS Forecast Server is not using the SMAPE definition which is published on the WIKI page you pointed to. You will find a details about SMAPE here: http://support.sas.com/documentation/cdl/en/hpfug/63959/HTML/default/viewer.htm#hpfug_hpfdet_sect034... - look for MASPE (unfortunately some outdated terminology seems to be used still).

Looking at your data I'm not sure if SMAPE an appropriate accuracy statistic, since your data seems to feature 0 values, which could cause numerical issues.

Also your actual values seem to feature large negative values, which could cause the SMAPE value be larger than 200.

I would suggest to report this behavior to Technical Support in order to track down the root cause of the problem.

Thanks,

Udo

Occasional Contributor
Posts: 9

Re: SAS/FS and SMAPE fit statistic

Hi Udo,

Sorry for delays, went to hollidays.

Thanks for this input - I was not aware of the various possible definitions of SMAPE.

Something strange is that I do effectively have some TS with a value of 200 (upper bound of "my" definition).

I don't find on google something on MASPE so I guess this is indeed old or erroneous info for a certain definition of SMAPE

Anyway, I may go to another statistic -- come back to more classic MAPE I guess...

Thanks again,

Eric

Ask a Question
Discussion stats
  • 2 replies
  • 495 views
  • 3 likes
  • 2 in conversation