Forecast Pro works by fitting a statistical model to your historic data and extrapolating it via the fitted model. Thus, your data must be long enough to provide reasonably stable estimates of the most important features of the data. Very short or very noisy historic records usually yield very simple models because the data are too short to support statistical estimates of important features like seasonality.
If the data are very short, say four points or fewer, Forecast Pro can pick up neither seasonality nor trend and reverts to the Simple Moving Average model.
For more than four points but less than two years’ worth of data, Forecast Pro can fit and forecast trends but not seasonality. If your data are in fact nonseasonal, your forecasts are likely to be adequate. If your data are in fact seasonal, the forecasts are likely to be poor—Forecast Pro cannot extract or forecast the seasonality.
However, seasonal forecasts from short data sets are feasible using some of Forecast Pro’s customized approaches. For example, when the short data sets are nested within aggregate product groups with longer histories, top-down forecasting can be used. If there are other longer series which exhibit the same seasonal pattern then the pattern can be estimated from the longer data set and applied to the short data sets using the INDEXES modifier, a custom component model or the forecast by analogy model. Finally, you can map history from one or more items to your short data history item to create a longer history to use for forecasting.
Seasonality can be estimated and forecasted from two to three years of data, but this amount of data is marginal, especially when your data are noisy or intermittent. Patterns in the noise may be mistaken for seasonality, yielding inappropriate “seasonal” forecasts.
Robust capture of seasonality requires three or more years of data. Four to seven years is even better since there is more information from which the program can separate seasonality and trend from the noise.
There is little additional payoff in accuracy beyond about seven years of data, and the cost in computer time can be substantial.