The multistep forecasts generated by the preceding steps are summed from the STARTSUM= number to the LEAD= number. For example, if STARTSUM=4 and LEAD=6, the 4-step-ahead through 6-step-ahead forecasts are summed. The predictions are simply summed. However, the prediction error variance of this sum is computed by taking into account the correlation between the individual predictions. The upper and lower confidence limits for the sum of the predictions is then computed based on the prediction error variance of the sum.

The forecast summation is particularly useful when it is desirable to model in one frequency yet the forecast of interest is another frequency. For example, if a time series has a monthly frequency (INTERVAL=MONTH) and you want a forecast for the third and fourth future months, a forecast summation for the third and fourth month can be obtained by specifying STARTSUM=3 and LEAD=4.

Variance-related computations are computed only when no transformation is specified (TRANSFORM=NONE).