Forecasting Slow Moving Demand (for the Royal Air Force, UK)

In this talk I compare methods for forecasting slow moving demand using a large dataset from the Royal Air Force in the UK. The methods included in the analysis are: moving average, exponentially weighted moving average, Croston's method and a simple variant of bootstrapping. As a benchmark, the zero method that simply always predicts zero demand is also included. I first argue that the traditional performance measures (MAD, MSE) cannot be used, even though they often are in the literature, and show that they actually identify the zero method as the best method. I then compare the methods based on their ability to obtain the right service level, assuming for all methods that lead time demand is Normal. The main conclusions are the following. (1) Croston and bootstrapping are the winners. (2) These methods can be considerably improved by realizing that each lead time starts with a demand. (3) Further improvement is possible by dropping the assumption that demand is Normal. For information: Robin Nicolai, rnicolai@few.eur.nl