The interaction between forecasting methods and inventory management policies for intermittent demand

Abstract

Inventory systems have to cope with uncertainty in demand. The inventory control literature mostly makes use of the Normal or Gamma distributions for describing the demand in the leadtime. The Poisson distribution has been found to provide a reasonable fit when demand is very low (only a few pieces per year). Less attention has been paid to irregular demand. This type ofdemand is characterised by a high level of variability, but may be also of the intermittent type, i.e. demand peaks follow several periods of zero or low demands. In such a situation forecasting demand is considered difficult. This research investigates the performance of several forecasting methods for intermittent demand and their impact on inventory management policies. A simulation model is built in order to investigate these effects and to serve as a guide for parameter settings of both the forecasting models and the inventory management policies. The experimental design includes three forecasting methods: simple exponential smoothing, moving average and Croston’s method. The inventory management policies make use of fixed review-time periods. Two alternative policies are studied: reorder point method with a fixed order quantity policy, and reorder point method with an order-up-to-level policy. Intermittent demand frequency is generated according to a Bernouilli process or a first-order Markov process. Individual order sizes are generated using a Gamma distribution. Preliminary experiments show that there is an impact of the forecasting method and of the review-time period, and to a lesser extent of the inventory management policy. Croston’s method has shown best performance in terms of service but expensive, while the Moving Average method shows low cost but also low service performance. Total costs are low when the review-time period is set equal to the lead-time, but high service performance is achieved when the review-time period is set equal to twice the lead-time in the experiment. This paper investigates further how robust these conclusions are for various settings of the model data and the various parameters of the methods.