A comparison of methods for forecasting demand for slow moving car parts

This paper has a focus on non-stationary time series formed from small non-negative integer values which may contain many zeros and may be over-dispersed. It describes a study undertaken to compare various suitable adaptations of the simple exponential smoothing method of forecasting on a database of demand series for slow moving car parts. The methods considered include simple exponential smoothing with Poisson measurements, a finite sample version of simple exponential smoothing with negative binomial measurements, and the Croston method of forecasting. In the case of the Croston method, a maximum likelihood approach to estimating key quantities, such as the smoothing parameter, is proposed for the first time. The results from the study indicate that the Croston method does not forecast, on average, as well as the other two methods. It is also confirmed that a common fixed smoothing constant across all the car parts works better than maximum likelihood approaches.