Intermittent Demand

The standard treatment splits the series into two components – the interval between demand occurrences and the size of each occurrence – and forecasts them separately. Croston's method and its bias‑corrected variants work this way, and their output is a demand rate per period rather than a prediction for a specific day. Items are classified by average inter‑demand interval (ADI) and the squared coefficient of variation of demand size (CV²), which separates smooth, erratic, intermittent and lumpy patterns. Spare parts, slow‑moving assortment in small stores and B2B order books with few large customers are the usual cases.

• ADI = average number of periods between two non‑zero demands.
• CV² = (standard deviation of non‑zero demand size / mean non‑zero demand size)².
• Common cut‑offs (Syntetos–Boylan–Croston): ADI 1.32 and CV² 0.49.

Intermittent demand in practice

Intermittency is a property of the item–location–period combination, not of the item. The same SKU is smooth at a distribution centre and lumpy at store × day, so you have to classify at the level at which you order. Aggregating to weeks or months makes the series look well behaved without changing anything about the order that has to go out tomorrow.

The bigger trap is that a zero is ambiguous. It can mean nobody wanted the item, or that it was not on the shelf, and without stock‑on‑hand history the two are indistinguishable – the model learns that the item does not sell, orders less and confirms itself. Veritico STOCK takes stockouts into account and forecasts actual demand instead of recorded sales, and cleans the history of outliers, stockouts and promotional effects before the model is fitted.

Because the output is a rate, accuracy measured period by period says little. What decides availability is the demand distribution over lead time plus review period and the safety stock quantile derived from it. For part of the assortment the answer is not a better model at all: an item that sells four pieces a year with an order unit of twelve is decided by the pack size and the listing decision, not by the forecast. Sales‑frequency segmentation – how many months out of twelve the item sold – separates those items before anyone tunes a model.

An item selling in single units per week has more zero days than non‑zero ones, and a network of hundreds of small outlets multiplies the number of such series. Dr. Max deployed Veritico STOCK across 490 pharmacies to automate ordering and optimise delivery schedules, saving around two hours of manual work per pharmacy per day – see the Dr. Max case study and replenishment and allocation management.

Related terms

• Demand Forecasting
• Safety Stock
• ABC Analysis
• MAPE
• XYZ Analysis
• Time Series Forecasting