I want to do an automatic forecast for a bunch of time series using Singular Spectrum analysis in R-Studio (RSSA package). The main problem is that I can't figure out which window length is better to chose for each of the time series.
I tried to do a forecast of the last 15 values for each of the time series: iteratively (within the interval from N/6 to N/2, where N - time series length) I picked a window length, then I tried to forecast values, then I measured the forecasting error (after that I've got about 300 values of the window lengths with corresponding errors). Then I picked a window length on which a forecasting error was minimal. This gave me pretty good result of the forecast.
But the idea is to perform a forecast of the 'unknown' values. So, I tried to do that this way:
I cut off the last 15 values of the time series. Then I picked the best window length (by the procedure I described earlier) on the part of TS that remained after deleting 15 values. Then I tried to use that window length for forecasting those 15 values, that I cut in the first place (just like if they were unknown). And I've got horrible results.
Does any of you have an idea what should I try? I'm stuck.