Croston's Method for Intermittent SKUs: Forecasting the Slow Movers Your Average Keeps Getting Wrong

Picture a foodservice-only pack, a regional flavor, or a club-channel one-off. It sells 40 cases one week, then nothing for six weeks, then 25 cases, then silence again. Your planning system runs a moving average or exponential smoothing across those weeks and hands you a forecast that spikes right after the sale and decays into a misleading trickle by the time the next order actually lands.

That is the slow-mover trap, and it burns planners every day. Averages assume demand happens every period. Intermittent SKUs don't work that way. To forecast them well, you have to forecast two different things separately: how much sells when it sells, and how often it sells. That is exactly what Croston's method does — and it's worth knowing because these sporadic items are usually a large slice of your SKU count and a surprising chunk of tied-up stock.

Why standard forecasting breaks on intermittent demand

Intermittent demand appears sporadically, with some periods showing no demand at all, and when a demand does occur the size may be constant or highly variable. That combination of zeros plus variable sizes is what wrecks conventional methods.

In practice, exponential smoothing is often applied to this type of demand. Because it places more weight on the most recent data, it produces estimates that are highest just after a demand occurrence and lowest just before the next one arrives. The result is a sawtooth: the forecast is at its peak in the weeks right after a sale — when you probably have plenty of stock — and at its lowest right before the next order lands, which is precisely when you most need replenishment signal. The forecast is mis-timed against reality.

Downstream, that mis-timing quietly corrupts everything you build on top of the forecast:

• Safety stock gets sized off a noisy per-period average that never matched the true demand pattern.
• Reorder points trigger at the wrong moments — high right after a sale, starved right before the next one.
• Overstock and obsolescence accumulate on the long tail, because a decaying average keeps recommending buffer for an item that may be fading out entirely.

If you've read Forecasting that survives the floor, this is the same lesson from a different angle: a forecast number that looks fine in a spreadsheet can be structurally wrong for the demand pattern it's describing.

What Croston's method actually does

Croston (1972) was the first to recognize the importance of separating demand size from the inter-demand interval. Instead of smoothing one raw series full of zeros, Croston's method splits the history into two streams:

1. The sizes of the non-zero demands.
2. The intervals — the number of periods between one demand and the next.

It then uses exponential smoothing to update each estimate separately, but only when a positive demand occurs. The per-period forecast is simply the ratio: estimated size ÷ estimated interval.

Croston's method is the standard approach used in industry to forecast intermittent requirements, and it's built into ERP-type systems such as SAP and specialized forecasting packages such as Forecast Pro. So if your planning tool has an "intermittent" or "slow mover" forecast profile, there's a good chance Croston (or a variant) is running under the hood.

A small worked example

Suppose a specialty SKU has this 12-week demand history:

0, 0, 40, 0, 0, 0, 0, 25, 0, 0, 50, 0

The non-zero sizes are 40, 25, 50. The intervals between demands are roughly 5, 5, and 3 periods. Use a smoothing constant of 0.2 and initialize the size estimate at 40 and the interval estimate at 5.

• After the second demand (size 25, interval 5):
  • Size estimate = 40 + 0.2 × (25 − 40) = 37
  • Interval estimate = 5 + 0.2 × (5 − 5) = 5
  • Per-period forecast = 37 ÷ 5 = 7.4 cases/week
• After the third demand (size 50, interval 3):
  • Size estimate = 37 + 0.2 × (50 − 37) = 39.6
  • Interval estimate = 5 + 0.2 × (3 − 5) = 4.6
  • Per-period forecast = 39.6 ÷ 4.6 = 8.6 cases/week

Notice two things. First, the Croston estimate is stable between demands — it doesn't sawtooth up and down with every zero week. Second, it separately tells you the expected order size (~40 cases) and the expected frequency (~1 order every 4-5 weeks), which is far more useful for sizing a replenishment than a blended 8.6 cases/week smeared across every period.

The catch: bias, and the fixes (SBA and TSB)

Croston's method is a big improvement, but it isn't perfect, and knowing its two weaknesses keeps you from over-trusting the number.

Weakness 1: positive bias. Syntetos and Boylan showed that Croston's method carries a positive bias, resulting in over-forecasting of mean demand. That matters directly for you: an inflated mean feeds inflated safety stock and reorder points, and on slow movers that means more cash and shelf space locked up in inventory that turns slowly — and more exposure to obsolescence.

The fix is the Syntetos–Boylan Approximation (SBA), which applies a deflating factor to Croston's estimate. In our example, an SBA-style correction (multiplying by roughly 1 − α/2 = 0.9) pulls the 8.6 down to about 7.7 cases/week. Independent empirical studies have shown SBA outperforms plain Croston on both forecasting accuracy and inventory performance. If your only lever is Croston vs. SBA, SBA is usually the safer default for the more irregular items.

Weakness 2: it goes stale on dying SKUs. Because Croston only updates when a positive demand occurs, its forecast doesn't change during long strings of zeros. That makes it unsuitable for estimating obsolescence risk — a discontinued item can keep showing a healthy forecast for months because nothing has "happened" to update it. The Teunter–Syntetos–Babai (TSB) method fixes this by updating the demand probability every period, including zero periods, so the forecast decays as an item fades out. If you manage seasonal, regional, or end-of-life SKUs that quietly die, TSB is the variant that will tell you the truth in time to stop buying.

Rule of thumb: Croston for reasonably behaved intermittent items, SBA for more erratic and lumpy ones, and TSB when obsolescence and fade-out are the real risk.

Which of your SKUs actually need this?

You don't want to apply Croston to everything — fast, steady movers are fine on ordinary smoothing. The standard way to sort SKUs comes from Syntetos et al. (2005), which classifies items using two numbers:

• ADI (Average Demand Interval): the average number of periods between demands. Higher = more sporadic.
• CV² (squared coefficient of variation): the variability of the demand size when it occurs. Higher = more erratic size.

Those two axes create four demand types, split at thresholds of ADI = 1.32 and CV² = 0.49:

Demand type	ADI	CV²	Recommended method	CPG example
Smooth	≤ 1.32	≤ 0.49	Ordinary SES / normal methods	Core SKU, steady weekly pull
Erratic	≤ 1.32	> 0.49	SBA	Frequent but wildly variable order sizes
Intermittent	> 1.32	≤ 0.49	Croston's method	Regional flavor, consistent case size but sporadic
Lumpy	> 1.32	> 0.49	SBA	Foodservice one-off, rare and variable

The method-selection logic follows the same grain: smoother series (lower CV² and ADI) are best modeled by the original Croston's method, while SBA is more appropriate for the more intermittent, erratic, and lumpy series. Lumpy demand — highly irregular in both timing and size — is the hardest quadrant and where a plain average does the most damage.

This pairs naturally with ABC-XYZ segmentation: ABC-XYZ tells you which SKUs are worth attention and how variable they are; the ADI/CV² split tells you which forecasting method each one needs.

From forecast to a schedulable number

A per-period intermittent forecast is only useful once it becomes an inventory decision. Historically, little attention was paid to linking Croston's forecasts to inventory control, but studies have since shown the generated forecasts can be used to calculate order-up-to levels that hit specific service-level targets. In other words, the size-and-interval estimate feeds directly into a reorder point and an order-up-to level rather than staying a floating forecast number.

A practical framing from GAINS is worth borrowing here: forecasting gives a point estimate of future demand, while demand patterns explain the recurring behaviors that shape that forecast. For irregularly ordered items, the workable approach is to set minimum and maximum inventory thresholds and monitor usage, so parts are available without overstocking. On the long tail, min/max control with active usage monitoring often beats chasing a precise number.

And the long tail is worth the effort. Even when individual units are low value, slow movers often represent a large percentage of the number of SKUs, so collectively they're an important element of the planning process — a large share of your item count, your warehouse slots, and your obsolescence exposure. For how service level targets translate into buffer, see cycle service level vs. fill rate.

Operator checklist

• Flag your zero-heavy SKUs. Any item with frequent zero-demand periods should not ride on a plain moving average or standard exponential smoothing — the sawtooth will mis-time your replenishment.
• Classify before you forecast. Compute ADI and CV² and drop each SKU into the four-quadrant grid. Smooth items stay on ordinary methods; intermittent, erratic, and lumpy items move to Croston-family methods.
• Split size from interval. Use Croston (or SBA) so you forecast expected order size and expected frequency separately — that's what makes safety stock and reorder points meaningful on sporadic items.
• Watch the positive bias. Croston over-forecasts the mean; use SBA's deflation before you size safety stock so you don't over-buffer the long tail.
• Use TSB for fade-out. For seasonal, regional, or end-of-life SKUs, switch to a method that updates every period so you catch obsolescence early instead of buying a dying item for months.
• Close the loop to the schedule. Translate the forecast into order-up-to levels and reorder points, set min/max guardrails, and re-plan on a regular cadence — the forecast is an input, not the answer.

Get these slow movers right and you stop the two silent failure modes at once: the stockout on the niche pack a big customer actually wanted, and the pallet of a fading SKU slowly aging out in the back of the warehouse. For the broader habit of catching drift before it costs you, pair this with tracking signal to detect forecast bias and stop firefighting stockouts.

Sources

• Teunter, Syntetos & Babai — Intermittent demand: Linking forecasting to inventory obsolescence — https://www.sciencedirect.com/science/article/abs/pii/S0377221711004437
• A new method to forecast intermittent demand in the presence of inventory obsolescence — https://www.sciencedirect.com/science/article/abs/pii/S0925527318300562
• Syntetos & Boylan — On the bias of intermittent demand estimates — https://www.sciencedirect.com/science/article/abs/pii/S0925527300001432
• Willemain et al. — Forecasting intermittent demand in manufacturing: a comparative evaluation of Croston's method — https://www.sciencedirect.com/science/article/abs/pii/0169207094900213
• Intermittency and obsolescence: A Croston method with linear decay — https://www.sciencedirect.com/science/article/pii/S0169207020301382
• Calculating order-up-to levels for products with intermittent demand — https://www.sciencedirect.com/science/article/abs/pii/S0925527308002508
• Snyder, Ord & Beaumont — Forecasting the intermittent demand for slow-moving inventories — https://www.sciencedirect.com/science/article/abs/pii/S0169207011000781
• ADI–CV² classification (beverage manufacturing application) — https://www.sciencedirect.com/science/article/abs/pii/S1568494625015674
• GAINS — Demand Patterns in Supply Chain — https://gainsystems.com/blog/demand-patterns-in-supply-chain-learn-how-to-plan-with-precision/