At some point I played with this functionality as well. As I observed things similar to you (flat line at the past average, and an uncertainty band encompassing pretty much al previous variation), I did a simple experiment: I took (complete) data for 1 year an replicated that 5 times, creating 5 identical years, hoping that the forecasting tools would pick up certain trends in the past (like the dip that occured in August every year). Note that I did not use one numbe rper month, but a very large set of points of sales at different dates and that what I tried to predict is the total sales per month. Even with perfectly replicated data for many years in a row, obvious seasonal variation is not well fitted. Maybe overfitting can be an issue in some cases, the VA forecaster seems to underfit seasonality. If the same (>30%!) dip occurs every year in the same month, in my humble opinion a forecasting tool should at least show some of this in the forecast for next year. I think a simple forecast would be a trend (linear over time, for example), plus seasonality, plus random 'noise'. The latter cannot be predicted, of course, but the first two can be, to some extent. Given that even with a series of identical years VA refuses to do so it seems to me that only a trend is predicted and that the uncertainty range is just obtained from all variation around the trend in the time span that is in the data. Would Forecast Studio give a more complete package to do forecasting in some more detail? For now, my self-written base sas code does better than VA. Far better.
... View more