I was able to get better results by reviewing, and then dropping, some harmonics. I hadn't thought to do this the first time around as I've normally used type=dummy. In particular, I dropped 6, 8 and 10 even though the p values were not particularly high (but were higher).
This would complicate matters if I introduced this into our current model process, as we automatically specify the model. I would need to build logic that would figure out which harmonics to drop/keep according to certain rules (some p value cutoff, allowing it to iterate several times to check the results and remove again); and have those rules consistently result in modeling the seasonality the way it should be modeled (this is just one example, will I even have this problem with production data? If so, will this be the solution?)
I'm curious now as to whether or not this would be common, to be able to drop a few harmonics on the season and have it make a dramatic difference in the blockseason component.
... View more