Intermittent demand forecasting: methods that actually work

By Elie Dufeu, CTO & Co-Founder, Metreecs. Published 22 September 2026.

Intermittent demand forecasting is the practice of predicting demand for products that sell irregularly, with long stretches of zero sales between orders. Standard forecasting methods assume demand shows up every period in some quantity. Intermittent items break that assumption, which is why they need their own models and their own accuracy metrics.

If you plan inventory for a catalog with any long tail, spare parts, or seasonal one-offs, you already know the feeling: a product sells eleven units in March, zero in April and May, then four in June. A moving average smooths that pattern into a number that describes no real week. This article covers why that happens, which methods actually handle it, and how to set safety stock and reorder points once the forecast itself is not misleading you.

Key Takeaways

• Intermittent demand is defined by frequent zero-demand periods mixed with irregular nonzero orders, not by low volume alone.
• MAPE is mathematically undefined when actual demand is zero, which makes it the wrong metric to grade intermittent demand forecasts by.
• The Croston method, published in 1972, forecasts demand size and the interval between orders separately, then combines them.
• According to McKinsey, AI-driven forecasting can reduce demand-forecast errors by 20 to 50% and cut lost sales by up to 65% when embedded properly into planning.
• Safety stock formulas built for steady demand systematically overstock or understock intermittent products unless they account for the probability of an order occurring at all.

What makes demand intermittent in retail

A demand pattern counts as intermittent when a meaningful share of periods have zero sales, interrupted by irregular positive demand. That is different from a slow-moving product that sells a small, steady amount every week. A product selling two units every Tuesday is slow but not intermittent. A product that sells eight units in one week and nothing for the next six is intermittent, regardless of its total annual volume.

Retailers run into this constantly: new product tails still building a sales history, accessory and replacement parts, seasonal one-offs that only move during a specific week, and the long tail of any catalog above a few thousand products. Jewelry and footwear retailers see it at the variant level, where a single style splits into a dozen size or metal combinations, some of which barely sell. Electronics distributors see it on replacement parts. Grocery and pharmacy networks see it on slow specialty items that still need to be on shelf when a customer asks for one.

If your catalog already shows this pattern, see what product-level forecasting looks like for a mixed catalog before restructuring how your team plans it.

Priya buys for a multi-brand beauty retailer carrying about 600 niche and limited-run products alongside the core catalog. Her team’s weekly forecast treated a product that sold four units in one week and nothing for the next five as roughly one unit of steady weekly demand, which looked fine on paper. On the shelf, it meant the product was either out of stock the week a customer actually wanted it or sitting untouched with a full case in the stockroom the other five weeks. Neither outcome showed up clearly in a weekly sales report, which is part of why the pattern went unaddressed for two full seasons.

One mistake we repeatedly see is treating every low-volume product the same way, whether it is steadily slow or genuinely intermittent. The fix is not more data, but recognizing that these are two different statistical patterns that call for different forecasting logic, not the same formula turned down.

Why standard forecasting methods break on intermittent demand

Most demand planning tools default to some form of moving average or exponential smoothing. Both work by averaging recent periods, which assumes each period carries a comparable signal about demand. Intermittent demand violates that assumption on both sides: most periods show zero, and the nonzero periods can vary widely in size.

Run a simple moving average over a pattern like 0, 0, 9, 0, 0, 0, 4, 0, and the forecast lands somewhere around one or two units a period. That number never actually happens. It understates the true order size when demand does occur and it overstates demand in every zero period, which quietly builds excess stock across the catalog.

See how AI demand forecasting handles this at the product and location level instead of applying one smoothing curve to every item regardless of its demand shape.

The accuracy metric compounds the problem. MAPE, the most common forecast accuracy measure, divides the forecast error by actual demand, and when actual demand is zero, that division is undefined. Most systems either throw an error or silently treat it as 100% wrong. Rob Hyndman’s widely cited paper on forecast-accuracy metrics for intermittent demand makes this point directly: MAPE cannot be computed reliably once zero periods enter the data, and scale-free alternatives like MASE and RMSSE were built specifically to sidestep that division-by-zero flaw.

Forecasting methods built for intermittent demand

The Croston method

J.D. Croston published the foundational method for this problem in 1972. Instead of forecasting demand period by period, Croston’s method splits the series into two separate signals: the size of demand when an order occurs, and the interval between orders. Each signal gets its own exponential smoothing, and the forecast is the smoothed demand size divided by the smoothed interval.

That split matters because it stops zero periods from dragging the average down. A product that sells 10 units every eight weeks gets a forecast that reflects an order of roughly 10 units happening on that rough cadence, not a flat 1.25 units every week.

Syntetos-Boylan Approximation (SBA)

Later research found that Croston’s original method carries a small positive bias, meaning it tends to forecast slightly higher than actual demand over time. The Syntetos-Boylan Approximation corrects that bias with an adjustment factor applied to the Croston output. For catalogs with thousands of intermittent products, that correction adds up to a measurable difference in aggregate overstock.

Parametric and probabilistic models

Some planning teams model intermittent demand as a probability distribution rather than a point forecast, using distributions like Poisson or negative binomial to estimate the likelihood of an order of a given size in a given period. This approach is more common for spare parts and industrial catalogs, where the question is less “what will next week look like” and more “what is the probability we stock out before the next replenishment window.”

Machine learning approaches

Newer approaches train models on features beyond the product’s own sales history: category-level patterns, promotional calendars, and similarity to other products with comparable lifecycle stages. These models do not replace the logic of Croston or SBA so much as generalize it, learning the size and interval relationship across thousands of products at once instead of tuning it product by product. In our experience deploying forecasting across catalogs that mix steady and intermittent items, a single model architecture rarely performs best across both. The stronger setups, including how Metreecs approaches this, select a different model per product based on its demand pattern, rather than forcing one method to cover the whole catalog.

Comparing the main methods

How intermittent demand changes safety stock and reorder points

Standard safety stock formulas assume demand follows something close to a normal distribution around an average. Intermittent demand does not look like that. It looks like mostly zeros with occasional spikes, which means a formula built for steady demand will either flag a stockout risk that is not real or miss the actual risk of running out right when an order finally does land.

The practical adjustment is to build safety stock around the probability that a nonzero order occurs during the lead time, not just the average size of that order. A product that orders rarely but in large batches needs a different buffer than one that orders often in small batches, even if their average weekly demand looks identical on paper. The safety stock formula for seasonal retail covers the standard calculation; for intermittent products specifically, that formula needs the demand-occurrence probability built in rather than assumed away.

Across Metreecs’ work with jewelry and footwear retailers managing thousands of product and variant combinations, the products causing the most planning friction are rarely the top sellers. They are the mid-tail items that order unpredictably and get either buried in a category-level average or manually overridden by a planner guessing at the pattern from memory.

Marisol runs replenishment for a home décor chain with about 40 stores. Her catalog includes roughly 1,800 seasonal and accent products that sell in bursts tied to specific weeks, holidays, or local events. Before she separated intermittent items from her core assortment, the same weekly reorder cycle applied to everything, and her end-of-quarter markdown pile was dominated by products that had ordered heavily once, then sat untouched for eight weeks while the system kept reordering off a smoothed average. Splitting the intermittent tail into its own review cadence, checked every two weeks instead of every week, cut the number of products carrying unjustified safety stock without adding manual review time to her core assortment.

A practical approach for planning teams

Getting from a generic forecasting setup to one that handles intermittent demand correctly does not require rebuilding the whole planning process at once.

1. Classify the catalog first. Separate products by demand pattern, not just by volume. A product with steady low volume needs different treatment than one with irregular bursts, even at similar annual totals.
2. Stop grading intermittent items on MAPE. Switch to MASE or a fill-rate-based metric for this segment so forecast quality gets measured honestly instead of penalized for a metric flaw.
3. Apply Croston or SBA to the intermittent segment specifically. Leave the standard forecasting method in place for steady-demand products; there is no need to change what already works.
4. Rebuild safety stock around order probability, not just order size, for the intermittent segment.
5. Review cadence should match the demand pattern. Weekly review on a product that orders every six to eight weeks wastes planner time and encourages reactive overrides.

For catalogs with more than a few hundred products carrying this pattern, AI-powered product x location forecasting can apply the right method per product automatically instead of requiring a planner to classify and tune each one by hand.

Elias manages forecasting for a franchise network of quick-service restaurants that also stock a rotating set of limited-time merchandise items, mugs, seasonal cups, branded apparel, alongside the core food and beverage catalog. Those merchandise items are classic intermittent demand: a launch spike, then long flat stretches, then occasionally a second small bump when a location runs a local promotion. Treating them with the same replenishment logic as core beverage inputs had been producing chronic overstock on discontinued designs. Moving the merchandise tail to interval-based forecasting reduced the number of products sitting past their promotional window without materially changing labor on the planning side, since the system flags them for review instead of requiring a planner to track each item’s launch date manually.

FAQ

What is intermittent demand in supply chain planning? Intermittent demand describes a pattern where a product has frequent periods of zero demand interrupted by irregular nonzero orders. It differs from steady low-volume demand, which sells consistently in small amounts every period rather than in irregular bursts.

Why does MAPE not work for intermittent demand? MAPE calculates error as a percentage of actual demand. When actual demand is zero, that calculation is undefined, which either breaks the metric or produces misleadingly harsh accuracy scores. Scale-free alternatives like MASE or RMSSE avoid the division-by-zero problem and give a more honest read on forecast quality.

What is the Croston method used for? The Croston method forecasts intermittent demand by separately smoothing the size of demand when it occurs and the interval between occurrences, then combining the two into a single forecast. It was developed specifically because standard moving-average and exponential smoothing methods understate order size on intermittent items.

How do you calculate safety stock for intermittent demand items? Standard safety stock formulas assume demand follows a roughly normal distribution, which does not describe intermittent patterns well. The practical fix is to build the buffer around the probability that an order occurs within the lead time window, not just the average order size, so the buffer reflects real stockout risk rather than a smoothed average that rarely matches actual behavior.

Can AI improve intermittent demand forecasting? Yes. Machine learning approaches can learn demand-size and order-interval patterns across an entire catalog at once, applying a different model to each product based on its actual demand shape rather than tuning Croston or SBA parameters product by product. McKinsey research finds AI-driven forecasting can reduce demand-forecast errors by 20 to 50% and cut lost sales by up to 65% when properly embedded in the planning process.

Conclusion

Intermittent demand forecasting is a different statistical problem from standard forecasting, not a smaller version of it, and treating it with the same moving average and the same MAPE scorecard produces forecasts that look reasonable on a dashboard and wrong on the shelf. Separate the intermittent segment of your catalog, apply Croston, SBA, or a model built for this pattern, and rebuild safety stock around order probability instead of order size alone.

Book a demo to see how product-level forecasting handles a mixed catalog, steady and intermittent items together, without forcing one method to cover both.

Sources

• Rob Hyndman, “Another Look at Forecast-Accuracy Metrics for Intermittent Demand”, Foresight: The International Journal of Applied Forecasting. Supports the MAPE limitation and MASE/RMSSE discussion.
• McKinsey, “Succeeding in the AI-Powered Supply-Chain Revolution”. Supports the 20-50% forecast error reduction and up to 65% lost-sales reduction stat.