Model Identification in Exponential Smoothing

Model identification has traditionally been ignored in forecasting via exponential smoothing. The usual practice is to apply the same model to every time-series in a collection. This paper develops a procedure for model identification in large forecasting applications based on an examination of variances of differences of the time-series. The order of differencing yielding minimum variance suggests an appropriate model for the series. Empirical results show that this procedure selects models that give reasonable ex ante forecast accuracy.