On the performance of adjusted bootstrapping methods for intermittent demand forecasting

A plethora of methods have been developed in the last decades to deal with the inventory forecasting of intermittent demand items. These methods belong to the parametric and non-parametric approaches, the artificial neural networks approach and the fuzzy logic-based techniques among others. Most of the parametric methods represent variations of the Croston method and the non-parametric ones are based on bootstrapping. When the inventory performance is considered, these methods often result in an under-achievement of the target service level. A study by Teunter and Duncan (2009) has shown that when the lead-time demand forecast is adjusted by assuming that the first period in the lead-time bucket corresponds to a non-zero demand, the service performance improves considerably. The study has been conducted by considering two Croston type methods and a parametric (with Normal distribution) bootstrapping method. However, the study has not included two well performing bootstrapping methods that are commonly considered in the literature to deal particularly with the inventory forecasting of intermittent demand items. A first method that samples demand data by using a Markov chain to switch between no demand and demand periods and a second method that samples separately demand intervals and demand sizes. In this paper, we propose variations of the two bootstrapping methods where the lead-time demand is adjusted by considering that a demand occurs in the first period of each lead-time bucket. A service driven inventory system is considered to evaluate the performance of the proposed forecasting methods with two objective service measures: the cycle service level and the fill rate. Through an empirical investigation based on the spare parts demand of more than 9000 stock keeping units, we show that the proposed adjusted methods result in a higher service-cost efficiency compared to the original methods.