Parameter Choice for Holt Winters: Synchronizing
We looked at the Excel moving
average for differing numbers of points to find the best agreement
(synchronization, i.e. the peaks and valleys aligned in time) of the moving
average of the ABwE Cap measurements made at 2.5 minute intervals with the
HW forecast (using 95% of the weights occur in the last ~10 hours = 250
points). The graphs below show the effects. It appears that using a moving
average over the last 125-150 points is probably near the optimum. The
change from 250 (for HW) to 125-150 (for moving average) is probably due to
the exponential effects. I think the best synchronization is actually closer
to 125, i.e. ~ a factor of 2. This factor of two is probably due to
exponential smoothing and probably changes as one uses different weighting.
