Intermittent Demand Forecasting

14Model‐based Methods

14.1 Introduction

In Chapter 13, we saw how simple bootstrapping can be adapted to take into account dependence between demand occurrences by using a Markov chain approach. This can be further augmented by ‘jittering’, to allow for the generation of demand values not previously observed.

The advantage of bootstrapping is that there is no requirement to assume that demand conforms to any of the standard distributions. However, if the demand does conform in this way, for example to the Poisson distribution, then parametric methods may be preferred. No jittering is required because these methods allow estimation of cumulative distribution functions (CDFs) across the entire range of values, including values higher than those previously observed.

So far, we have not examined how parametric methods should be used when there is correlation between successive observations, known as autocorrelation. As noted in Chapter 13, there is empirical evidence of autocorrelation of demand for some intermittent series, and so this topic is worthy of further attention.

In this chapter, we start with a general discussion on models and methods, with a particular focus on models for faster‐moving demand. This is followed by five sections on integer autoregressive moving average (INARMA) models, the most well‐developed models for low‐volume and intermittent series. After defining the models, we move on to parameter estimation and model identification. Then, we focus on forecasting ...