Forecasting-based SKU classification

Different spare parts are associated with different underlying demand patterns, which in turn require different forecasting methods. Consequently, there is a need to categorise stock keeping units (SKUs) and apply the most appropriate methods in each category. For intermittent demands, Croston's method (CRO) is currently regarded as the standard method used in industry to forecast the relevant inventory requirements; this is despite the bias associated with Croston's estimates. A bias adjusted modification to CRO (Syntetos–Boylan Approximation, SBA) has been shown in a number of empirical studies to perform very well and be associated with a very ‘robust’ behaviour. In a 2005 article, entitled ‘On the categorisation of demand patterns’ published by the Journal of the Operational Research Society, Syntetos et al. (2005) suggested a categorisation scheme, which establishes regions of superior forecasting performance between CRO and SBA. The results led to the development of an approximate rule that is expressed in terms of fixed cut-off values for the following two classification criteria: the squared coefficient of variation of the demand sizes and the average inter-demand interval. Kostenko and Hyndman (2006) revisited this issue and suggested an alternative scheme to distinguish between CRO and SBA in order to improve overall forecasting accuracy. Claims were made in terms of the superiority of the proposed approach to the original solution but this issue has never been assessed empirically. This constitutes the main objective of our work. In this paper the above discussed classification solutions are compared by means of experimentation on more than 10,000 SKUs from three different industries. The results enable insights to be gained into the comparative benefits of these approaches. The trade-offs between forecast accuracy and other implementation related considerations are also addressed.