The Croston Method for Slow-Moving Items

Standard smoothing keeps telling you to reorder a SKU that sells once a month, because it reads the long silences as "low steady demand" instead of "no demand most weeks." Croston's method was built in 1972 for exactly this product, and it starts from the one insight the standard methods miss.

Croston's method is a forecasting technique for intermittent demand that splits the pattern into two parts: the average size of a sale when one happens, and the average interval between sales, then forecasts each separately. By separating how much from how often, it produces a steadier, more usable forecast for slow-moving and irregular SKUs than simple averaging.

Key takeaways

• The insight is refusing to average the zeros: the silences become spacing information, not demand observations.
• Two series instead of one: sale sizes get smoothed, intervals between sales get smoothed, and the forecast is one divided by the other.
• The output reads differently: not "2.2 units every week" but "around 10 units, roughly every 4 to 5 weeks," which is what a stocking decision actually needs.
• You should never run it by hand: modern systems classify intermittent SKUs and apply Croston (or its SBA correction) automatically.

What is Croston's method?

Croston's method forecasts intermittent demand by refusing to average the zeros. Instead of one demand series, it tracks two: how big a sale is when it occurs, and how many periods pass between sales. Each gets smoothed on its own, and the forecast combines them into a demand rate per period. The zeros stop diluting the picture, because they're no longer treated as demand data; they're treated as spacing.

Where it came from

J.D. Croston published the method in 1972, working on exactly the inventory problem it still solves: spare parts and slow movers whose reorder math kept misfiring under exponential smoothing. It has since become the standard baseline for intermittent-demand forecasting in planning tools. (What makes demand "intermittent" in the first place, and how to spot those SKUs in your catalog, is owned by intermittent demand; this page is about the fix.)

How does it split demand size and interval?

The method maintains two smoothed estimates, updated only when a sale actually happens: the demand size estimate (running average of non-zero sale quantities) and the interval estimate (running average of periods between those sales). Divide size by interval and you get the forecast: an average demand rate per period that's honest about the sparsity.

A compact walk-through. Suppose a SKU's last stretch of sales looks like this: 12 units in week 3, 8 units in week 7, 10 units in week 12. The size series is 12, 8, 10: average 10 units per event. The interval series is 4 weeks then 5 weeks (week 3 to 7, week 7 to 12): call it 4.5 weeks between events. Croston's forecast: 10 ÷ 4.5 ≈ 2.2 units per week. Compare that with a 12-week moving average, which would report roughly 2.5 units a week while also implying that a normal week sells something. Croston's output reads differently: expect an order of around 10 units roughly every 4 to 5 weeks. Same arithmetic neighborhood, radically more useful for stocking decisions, because it tells you the shape of what's coming, not just the diluted rate.

When should you use Croston's method?

The dividing line is the shape of the demand, not the importance of the product:

• Use it for: SKUs with frequent zero periods and irregular order timing. Slow movers, spare parts, accessories, specialty items, the deep tail of the catalog.
• Skip it for: steady sellers, where ordinary smoothing is simpler and at least as accurate.
• Watch the false positives: a SKU that looks sparse because of stockout gaps, or one with strong seasonality, isn't truly intermittent. Those causes are knowable, and a pattern-aware model should handle them explicitly instead.

Variants worth knowing about

Croston's original math runs slightly high as a long-run rate estimate; the Syntetos-Boylan approximation (SBA) applies a small correction factor and is what many modern systems actually run under the Croston name. There are further refinements for obsolescence and for erratic sizes, but for a working planner the headline is simpler: if your planning tool offers Croston or SBA for slow movers, that's the intended tool for the sparse tail.

In practice you shouldn't be running Croston by hand at all. Modern planning systems apply it (or its corrected variants) automatically once a SKU classifies as intermittent, and that classification is where the real gain sits. Conative AI detects the demand pattern per SKU and routes each one to a method that fits, with deep learning models that can also take inputs Croston never sees, like stockout history and marketing activity, where the data supports it. AI-powered demand forecasting matched to the demand pattern, instead of one formula stretched across the whole catalog. Book a call to see how your slow movers forecast.

Frequently asked questions

What problem does Croston's method solve?

The failure of average-based forecasting on intermittent demand. When a SKU sells rarely, averages and smoothing dilute occasional spikes across long silences, producing a forecast that's wrong in both directions. Croston separates how much sells per event from how often events happen, giving slow movers a demand rate that reflects their real pattern.

How is Croston different from exponential smoothing?

Exponential smoothing updates one estimate (the demand level) every period, zeros included, so long silences drag the forecast into no man's land. Croston applies smoothing twice, and only when a sale occurs: once to the sale sizes, once to the intervals between sales. The zeros become spacing information rather than demand observations.

What is the SBA (Syntetos-Boylan) correction?

The Syntetos-Boylan approximation corrects a known bias in Croston's original method, which slightly overstates the long-run demand rate. SBA applies a small downward adjustment based on the smoothing parameter. Many planning systems labeled "Croston" actually run SBA. For practical purposes it's the refined default for intermittent-demand forecasting.

Can a planning tool run Croston automatically?

Yes, and that's the normal way to use it. Planning platforms classify each SKU's demand pattern, apply Croston or SBA to the intermittent ones, and re-run the estimates as sales arrive. Nobody should maintain interval series in a spreadsheet across a real catalog; the method's value shows up when classification and application are automatic.

Is Croston the best method for spare parts?

It's the established baseline, and for most spare-parts and slow-mover forecasting it beats averages and standard smoothing clearly. Refinements like SBA usually edge out the original. Machine learning approaches can do better where useful extra signals exist, but for sparse demand with little context, Croston-family methods remain the standard.

What are the limits of Croston's method?

It assumes the pattern is stable: steady average sizes and intervals. It doesn't anticipate trend, seasonality, promotions, or the moment a slow mover starts dying entirely (obsolescence), and it updates only when sales occur, so it reacts slowly to change. Treat its forecast as a baseline rate, and layer judgment or richer models where those forces matter.