Forecasting Intermittent Demand: Why Moving Averages Fail on Slow-Selling Parts

A large share of items in many organizations, especially spare parts, specialized items and special products, have intermittent demand: in most periods sales are zero and occasionally an order, sometimes a large one, arrives. Usual methods such as moving averages and exponential smoothing do not suit this pattern; their output is a small, steady number that neither covers shortages nor supports a good decision.

The first step is diagnosing the pattern. Two simple indicators help: the average demand interval (ADI), showing how many periods typically pass between demands, and the squared coefficient of variation of demand size (CV²), measuring the variability of order sizes. Based on these two, items split into four groups: smooth, intermittent, erratic and lumpy. The lumpy group, both irregular and with variable size, is the hardest.

Croston's method was designed for these items. Instead of smoothing the sales series with its many zeros directly, it estimates two things separately: the demand size in periods where demand occurred, and the interval between demands. The demand rate is their ratio. Improved versions such as SBA correct the original method's bias and are used more in practice.

But stay realistic: no statistical method gives high accuracy for highly lumpy demand. Here the goal is not point-forecast accuracy but a good inventory decision. Instead of forecasting each period, estimate the distribution of demand over the lead time and set stock from that distribution at an appropriate service level, not by assuming normality, which is wrong for this pattern.

For very low-usage, high-cost items, a question comes before any modeling: should stock be held at all? Options include make or buy to order, a supplier agreement for fast delivery, sharing stock with other units or customers, and using substitute parts. Often the best answer for a CZ item is to hold no stock and accept the lead time.

A common mistake is applying one high service level to all intermittent items. Because the forecast is unstable, the safety stock needed for 99 percent service becomes very large and expensive. Set the service level by each item's true importance (impact of a stoppage and holding cost), and for critical items consider non-stock solutions such as fast-delivery contracts.

Another mistake is dropping intermittent items from planning because they are 'unforecastable'. These items often tie up significant capital and, if short, can halt operations. A different, simpler planning path should be defined for them, not just forgotten.

A practical starting point: classify your items by ADI and CV² and see the count and capital in each group. For the lumpy group, decide item by item whether to hold stock; for the others, test Croston or SBA and compare with your current method.