Intermittent Demand Forecasting

13Non‐parametric Methods

13.1 Introduction

In previous chapters, we have adopted a parametric approach to the characterisation of uncertain demand. This presupposes that future demand can be well represented by a distribution with parameters that are unknown but may be forecasted using past data. This approach is reasonable if a distribution can be found that satisfies the criteria introduced in Chapter 4 :

1. Empirical evidence in support of its goodness of fit.
2. Intuitive appeal in terms of a priori grounds for its choice.
3. Flexibility to represent different types of demand patterns.
4. Computational ease.

The most basic requirement is the first. Unfortunately, this requirement is often not met for a significant proportion of stock keeping units (SKUs). This problem was illustrated in Chapter 5, where we reviewed some empirical evidence on goodness of fit. Demand over a single period was well fitted by the stuttering Poisson distribution for over 90% of SKUs (Table 5.3). However, this figure dropped to 72% for the Royal Air Force (RAF) dataset when considering demand over lead time (Table 5.4), leaving 28% of SKUs not well fitted by this distribution.

In a situation like this, we can look to other distributions to fill the gap, but need to take care not to overcomplicate the allocation of SKUs to probability distributions. (Indeed, the flexibility criterion was introduced to avoid this type of complication.) Another possible solution would be to use a three‐parameter distribution, ...