MAS 2015 Proceeding

Better statistical forecast accuracy does not always lead to better inventory control efficiency: the case of lumpy demand

Authors:   Adriano O. Solis


Neural network (NN) modeling has been applied to forecasting of lumpy demand (Gutierrez, Solis, and Mukhopadhyay 2008; Mukhopadhyay, Solis, and Gutierrez 2012) and empirically compared with a number of well-referenced methods traditionally applied in studies on lumpy demand forecasting ? simple moving average, weighted moving average with optimal weights, simple exponential smoothing, Croston?s method, and the Syntetos-Boylan approximation. The overall superiority of NN over the other methods, in terms of forecast accuracy based on a number of scale-free error statistics, was demonstrated. However, demand forecasting performance with respect to standard accuracy measures may not translate into inventory systems efficiency. Applying a (T,S) inventory system, we consider fill rate (FR) as service criterion. We conduct simulation searches to find order- up-to levels required to meet a target FR of 0.90 or 0.95. We find that significantly higher levels of on-hand inventory are required when using the more statistically accurate NN forecasts.

I3M  Scientific Sponsors

I3M  Industrial Sponsors

I3M  Media Sponsors