• weighted moving average;
  • exponential smoothing;
  • Croston's method;
  • Syntetos–Boylan approximation;
  • neural networks


Forecasting for inventory items with lumpy demand is difficult because of infrequent nonzero demands with high variability. This article developed two methods to forecast lumpy demand: an optimally weighted moving average method and an intelligent pattern-seeking method. We compare them with a number of well-referenced methods typically applied over the last 30 years in forecasting intermittent or lumpy demand. The comparison is conducted over about 200,000 forecasts (using 1-day-ahead and 5-day-ahead review periods) for 24 series of actual product demands across four different error measures. One of the most important findings of our study is that the two non-traditional methods perform better overall than the traditional methods. We summarize results and discuss managerial implications. Copyright © 2011 John Wiley & Sons, Ltd.