Lumpy demand is a phenomenon encountered in manufacturing or retailing when the items are slow-moving or too expensive, for example fighter plane engines. So far, the seminal procedure of Croston's, with or without modifications, has been the preference for forecasting lumpy demand. Nevertheless, Croston and others, such as Venkitachalam et al., have suggested the use of zero forecasts when the demand contains many zeros. In this paper, we put to the test this idea by doing a full factorial study comparing five forecasting methods, including all-zero, under several levels of demand lumpiness, demand variation and ordering, holding and shortage cost. We evaluate the forecasting methods by three measures of forecast error and two measures of inventory cost. We find that all-zero forecasts yield the lowest cost when lumpiness is high; is it also best for mid-lumpiness, if the shortage cost is much higher than the holding cost. We also find that the lowest forecasting error does not necessarily lead to the lowest system cost. And contrary to the assertions of Chen et al. and Dejonckheere et al., our factorial experiment reinforces the intuition that simple exponential smoothing is superior to an equivalent moving average.
