Simple vs. complex extrapolation models

This paper describes an empirical study of time series data for consumer products in the food processing industry at the stockkeeping unit level. The focus is on whether or not simple forecasting models adequately fit this type of data. The Box-Jenkins methodology is used to select forecasting models. The results show that simple models are identified by this procedure. In particular, two types of models prevail: (1) models that require information from only the previous time period and (2) simple seasonal models. The forecasting results reinforce the choice of simple models and show that better fitting models do not, in general, give better forecasts. Finally, the simple exponential smoothing model is shown to be a robust forecasting model for data of this type.