John E. Boylan - Intermittent Demand Forecasting

For optimisation of control parameters, the results obtained for the system can be easily transferred to an system by substituting for , for , and for (where is the lead time and the annual demand). The combination does not take into account the variability of demand and hence should not be applied in a probabilistic demand context.

In Chapter 2, we reviewed inventory rules that may be used to manage the stock of intermittent demand items, paying particular attention to the and policies. In both of these policies, the inventory position is reviewed every periods, and enough stock is ordered to raise it to the order‐up‐to level, , also known as the OUT level. We noted that is often used for intermittent demand items because of its simplicity and robustness.

The policy requires the determination of the review interval and OUT level for each individual stock keeping unit (SKU). In practice, the review interval is usually set to be the same for all SKUs or for whole classes of SKUs, for reasons that were discussed in Chapter 2. The setting of the review interval varies according to industry sector. In grocery retail, this may be every day or half day, whereas in automotive spare parts, the review may be weekly or monthly.

The OUT level, , should be set separately for each SKU, to take account of its demand uncertainty. The determination of an OUT level for an individual SKU is an important issue for ‘mission‐critical’ items, for example spare parts without which a grounded plane cannot fly. For other SKUs, the determination of OUT levels may be less critical but is still important because of its effect on aggregate inventories. As discussed in Chapter 1, a whole range of SKUs may account for significant stock holding, the level of which is influenced by the OUT levels.

The setting of the OUT level in service‐driven inventory systems depends on three main factors:

1 Service measure.

2 Demand distribution.

3 Forecasting method.

The first factor, the service measure, is analysed in this chapter. Chapters 4and 5focus on the second factor, with discussion of various demand distributions and the criteria they should satisfy. The following two chapters are concerned with the third factor: Chapter 6concentrates on methods to forecast the mean demand, while Chapter 7is devoted to forecasting the variance of demand and its associated forecast error. All of these elements are brought together in Chapter 8, which explains how, for a given service measure, the OUT level can be found for intermittent demand items.

In this chapter, we begin by arguing against using rules of thumb for setting OUT levels, and by stressing the strategic significance of aggregate level financial and service targets. The choice of SKU‐level service measures is examined, noting their links to inventory costs, before moving on to the calculation of the two operational service level measures that are most commonly employed in inventory systems. Then, we return to the setting of aggregate service targets, emphasising the importance of ‘what‐if’ modelling capabilities. The chapter concludes with comments on the use of judgement and points to the need for reliable demand distributions to assess the service implications of different ordering policies.

In this book, we argue for a systematic and analytical approach to forecasting and inventory management. This should be based on inventory replenishment rules and forecasting methods that are well grounded statistically and have solid evidence of good performance in practice. From our work with a variety of organisations, we are aware that practitioners may use ordering rules that are ad hoc, or may adjust computer‐generated orders using their own judgement. In this section, we make some brief remarks on these practices.

Suppose that the review interval is one week and the lead time is two weeks, giving a total protection interval of three weeks. Suppose, further, that our forecasted mean demand is two units per week, and the demand is non‐trended and non‐seasonal. It may be tempting to set the OUT level as the forecasted mean demand over the protection interval, namely six units. This would be correct if it were certain that demand would be for the exact mean demand predicted, but this is rarely the case. More commonly, the demand will be fluctuating. Not taking account of these fluctuations can lead to frequent stockouts.

An alternative calculation, which attempts to address this issue, is to multiply the forecasted demand per week by a period that is longer than the protection interval, to allow for demand uncertainty. For example, we could multiply the forecasted demand, of two per week, by four weeks, instead of three, to give an OUT level of eight units. The problem now is that the setting of four weeks is arbitrary. Why not use five or six weeks instead of four? There is no guarantee of hitting service level targets using this type of calculation. For highly unpredictable demand, we may set the OUT level too low and, for more predictable demand, we may set it too high. Therefore, although this rule of thumb has the merit of simplicity, it risks service level targets being missed or targets being achieved with excessive stocks.