Applying modeling and simulation to evaluate statistical accuracy and inventory control perfomance of lumpy demand forecasting methods

A number of time series methods – 13-month simple moving average (SMA13), single exponential smoothing (SES), Croston’s method, and the Syntetos- Boylan approximation (SBA) – are well-referenced methods in the literature on intermittent or lumpy demand forecasting. We apply these four methods to an industrial dataset involving more than 1000 stockkeeping units (SKUs) in the central warehouse of a firm operating in the professional electronics sector. Earlier studies have argued that the negative binomial distribution (NBD) satisfies both theoretical and empirical criteria for modeling intermittent demand. We have found that the NBD often does not provide a good fit. We apply an alternative approach, using a two-stage distribution involving the uniform and negative binomial distributions, in modeling actual demand. We use modeling and simulation to evaluate the four methods in terms of statistical forecast accuracy and, more importantly, inventory system efficiency.