What Is Intermittent (Lumpy) Demand Forecasting?

This SKU sells zero units for six weeks, then someone orders forty at once. Your forecast, built on the average, expects six a week, so it's wrong every single week: too high during the silence, hopelessly low when the spike lands. Nothing is broken except the assumption that this product behaves like your best-sellers.

Intermittent (lumpy) demand is sporadic, irregular demand with many zero-sales periods and occasional spikes, typical of slow-moving or specialty SKUs. Standard forecasting methods fail on it because averaging across the zeros understates the spikes and overstates the quiet weeks. These items need methods built for sparse, uneven demand.

Key takeaways

• The average describes a week that never happens: most weeks sell zero, occasional weeks spike far above, and the forecast lands uselessly between.
• Intermittent, lumpy, and slow-moving aren't synonyms: intermittent means many zeros, lumpy adds erratic order sizes, slow-moving just means low volume.
• Two metrics sort the whole catalog: the average interval between demand events, and the variability of the non-zero order sizes.
• The fix treats "when" and "how much" as separate questions: which is exactly what purpose-built methods like Croston's do.

What is intermittent demand?

Intermittent demand arrives in scattered, irregular bursts separated by stretches of nothing. The pattern isn't a trend, isn't a season, and isn't noise around a steady average: it's a fundamentally different shape of demand, where "when will someone buy?" and "how much will they buy?" are two separate uncertainties. Spare parts, high-end specialty items, B2B add-ons, and deep-tail catalog products all live here.

Intermittent vs lumpy vs slow-moving: the vocabulary

The terms overlap but aren't synonyms. Intermittent means many zero periods: sales happen occasionally, in whatever quantity. Lumpy means intermittent and irregular in quantity: not only do sales arrive rarely, the amounts jump around (three units one time, forty the next). Slow-moving just means low total volume, which can still be perfectly regular (one unit every week is slow but steady, and easy to forecast). Lumpy is the hard case, because both the timing and the volume misbehave.

Here's the shape difference in numbers, eight weeks of two SKUs with the same total sales:

• Regular SKU. 1: 6; 2: 5; 3: 7; 4: 6; 5: 5; 6: 6; 7: 7; 8: 6
• Lumpy SKU. 1: 0; 2: 0; 3: 40; 4: 0; 5: 0; 6: 0; 7: 8; 8: 0

Both average six a week. One average describes its product; the other describes nothing that ever happens.

Why do standard forecasting methods fail on lumpy demand?

Because moving averages and smoothing assume demand is a steady level plus noise, and lumpy demand is neither. A moving average of the lumpy SKU above cheerfully forecasts six units every week: it over-forecasts seven quiet weeks and misses the forty-unit spike by a factor of nearly seven. Smoothing does the same with more math. The methods aren't bad; the assumption is wrong. How much a series bounces around is its own subject, demand variability, but lumpy demand is beyond bouncy: it's structurally sparse.

The zero-inflation trap

The zeros are what poison the math. Most weeks contribute nothing to the average except dilution, so the "expected" demand lands in a no man's land: far above what happens in a normal week (zero), far below what happens when demand actually arrives (the spike). Reorder triggers built on that number fire at the wrong times in both directions, holding stock through long silences and getting caught flat by the burst. Any fix has to treat "how often" and "how much" as separate questions, which is exactly what the purpose-built methods do.

How do you spot slow-moving and irregular SKUs?

Classify by two measurable dimensions: how often demand arrives, and how much the order sizes vary. In practice, screen every SKU for its share of zero-demand periods and the spread of its non-zero order quantities. High zero-share plus stable order sizes is intermittent-but-tame; high zero-share plus wild order sizes is lumpy, the segment that needs different handling entirely.

The metrics that classify them

Two numbers do the sorting:

• Average interval between demand events: how many periods pass, on average, between sales. This separates steady from sparse; a common working line puts intermittent at an interval above roughly 1.3 periods.
• Variability of the non-zero order sizes: how much the quantities jump around when sales do happen. This separates smooth from erratic.

Cross the two and every SKU falls into one of four demand patterns: smooth, intermittent, erratic, or lumpy, each with its own appropriate forecasting treatment. The named method built for the sparse quadrants is the Croston method, which forecasts the interval and the size separately.

Running that classification by hand across a real catalog is the part nobody sustains, and it's exactly the kind of work AI-powered demand forecasting should absorb. Conative AI classifies every SKU's demand pattern automatically and applies a method that fits it, so your lumpy tail stops being forecast like your best-sellers, and reclassifies as behavior changes. The planner sees which segment each SKU landed in and why, instead of discovering it through a quarter of bad triggers. See a demo on your own catalog.

Frequently asked questions

What's the difference between lumpy and intermittent demand?

Intermittent demand has many zero periods but reasonably consistent order sizes when sales do happen. Lumpy demand is intermittent and erratic in quantity: long silences plus order sizes that swing wildly. The distinction matters because intermittent-but-stable items are moderately forecastable, while lumpy items carry double uncertainty (timing and size) and need the most careful buffering.

Which products have intermittent demand?

Spare parts and accessories, high-price specialty items bought rarely, B2B products ordered in occasional batches, deep catalog variants (unusual colors), and anything whose buyers are few and sporadic. In a typical eCommerce catalog it's the long tail: SKUs kept for range or service reasons whose individual sales arrive weeks apart.

Why is intermittent demand hard to forecast?

Because the average, the workhorse of standard forecasting, describes a week that never happens: most weeks sell zero and occasional weeks spike far above it. Methods that assume a steady level plus noise over-forecast the silences and miss the bursts. The problem splits into two separate questions (when, and how much) that standard methods answer as one.

How do you classify a SKU as slow-moving?

Measure two things per SKU: the average interval between demand events (how many periods typically pass between sales) and the variability of non-zero order sizes. Long intervals mark the SKU as intermittent; add erratic order sizes and it's lumpy. Screening the catalog on those two dimensions sorts every product into a demand pattern with a matching method.

Can you hold safety stock for lumpy-demand items?

Yes, but the standard formula fits poorly because the demand distribution isn't the smooth bell curve it assumes. Practical approaches lean on service-level judgment: decide per SKU whether a stockout is tolerable, hold a buffer scaled to the typical burst size for the critical ones, and accept longer replenishment waits on the rest.

Does AI handle intermittent demand better than averages?

Meaningfully better, on two fronts. Classification: AI-powered platforms detect which SKUs are intermittent or lumpy automatically instead of letting them hide inside average-based forecasts. Method: models built for sparse demand forecast the interval and the size separately, and machine learning approaches can add signals beyond sales history. The gap versus a plain average is largest exactly here.