Demand Forecasting Wiki

Lumpy Demand: Meaning, How to Detect It, and How to Forecast It

By MLAIA Data Science Ltd. · Published 10 November 2026 · Updated 10 November 2026

Lumpy demand is demand that arrives only sporadically and in widely varying quantities: most periods record nothing, and the orders that do arrive jump from one unit to fifty. Formally, a series is lumpy when its average demand interval (ADI) is at least 1.32 periods and the squared coefficient of variation of its non-zero order sizes (CV²) is at least 0.49. The 12-month series worked through below — 0, 0, 20, 0, 1, 0, 0, 50, 0, 0, 4, 0 — has ADI = 3.0 and CV² = 1.07, so it is lumpy. It is the hardest of the four standard demand patterns to forecast: the usual answer is a Croston-family estimator (SBA) for the demand rate, and an empirical lead-time-demand distribution, not the normal-based safety-stock formula, for the reorder point.

What lumpy demand means

Demand patterns are usually described on two axes: how often demand occurs and how variable its size is when it does. The Syntetos–Boylan classification cuts each axis once, giving four patterns:

The four demand patterns. Lumpy is the bottom-right cell.
Stable sizes (CV² < 0.49)Variable sizes (CV² ≥ 0.49)
Demand most periods (ADI < 1.32)SmoothErratic
Many zero periods (ADI ≥ 1.32)IntermittentLumpy

A lumpy item therefore combines the two things that are individually hard: you do not know when the next order will come, and when it comes you do not know whether it will be for one unit or fifty. Spare parts for a small installed base, project-driven B2B orders, and the long tail of a distributor's catalog are typically lumpy.

How to detect it: a worked example

Take a part with the following monthly demand over one year:

Twelve months of demand for the example part. Four months have demand; eight have none.
Month123456789101112
Demand00200100500040

Step 1 — ADI. Twelve periods, four of them with demand:

ADI = 12 ÷ 4 = 3.0 periods between demands  (≥ 1.32, so the timing is sporadic)

Step 2 — CV² of the non-zero sizes. The sizes are 20, 1, 50 and 4. Their mean is (20 + 1 + 50 + 4) ÷ 4 = 18.75. The squared deviations from that mean are 1.5625, 315.0625, 976.5625 and 217.5625, which sum to 1,510.75; dividing by 4 gives a (population) variance of 377.69.

CV² = 377.69 ÷ 18.75² = 377.69 ÷ 351.56 = 1.07  (≥ 0.49, so the sizes are highly variable)

Step 3 — classify. ADI ≥ 1.32 and CV² ≥ 0.49: the part is lumpy. The free demand classifier performs exactly this calculation on a pasted series.

Two cautions. The classification depends on the time bucket: the same part summarised quarterly (20, 1, 50, 4 → every quarter has demand) is no longer intermittent at all, so classify at the bucket you plan in. And twelve periods is a short history; 24 or more gives a far more stable read, and an item near either cutoff should not be treated as fundamentally different from its neighbours.

Why lumpy demand defeats the usual methods

How to forecast lumpy demand

The standard estimator for lumpy items is the Syntetos–Boylan Approximation (SBA), the bias-corrected form of Croston's method. It smooths the demand sizes and the intervals between demands separately, updating both only when demand occurs, and forecasts the demand rate as (1 − α/2) · size ÷ interval. Applied to the example with α = 0.2, initialised from the first demand:

SBA on the example series, α = 0.2. Estimates change only in the four months with demand.
MonthDemandIntervalSmoothed size ẑSmoothed interval p̂Croston ẑ/p̂SBA 0.9·ẑ/p̂
320320.003.006.676.00
5120.2·1 + 0.8·20 = 16.200.2·2 + 0.8·3 = 2.805.795.21
85030.2·50 + 0.8·16.2 = 22.960.2·3 + 0.8·2.8 = 2.848.087.28
11430.2·4 + 0.8·22.96 = 19.170.2·3 + 0.8·2.84 = 2.876.676.01

The SBA forecast after month 11 is about 6.0 units per month, close to the long-run average but far steadier than exponential smoothing on the raw series would be. Notice what it does not say: nothing about which month the next order lands in, and nothing about whether it will be 4 units or 50. A demand rate is an input to the stocking decision, not the decision itself. The mechanics, the bias correction and the TSB variant for items that may be dying out are covered in SBA and TSB.

Two further techniques help specifically with lumpiness. Temporal aggregation (ADIDA, Nikolopoulos et al., 2011) forecasts at a coarser bucket, where the series is less intermittent, and spreads the result back. Bootstrapping (Willemain, Smart & Schwarz, 2004) resamples the observed history to build the whole distribution of demand over the lead time instead of a single rate, which is what the reorder point actually needs.

How to stock a lumpy item: the reorder point

Suppose the example part has a two-month lead time and a 95% cycle-service-level target. The classic formula uses the mean (6.25) and sample standard deviation (14.92) of monthly demand:

SS = 1.645 · 14.92 · √2 = 34.7    ROP = 6.25 · 2 + 34.7 = 47.2 → 48 units

Now build the lead-time-demand distribution from the history itself. Demand occurred in 4 of 12 months, so each month has a 1/3 chance of an order, and an order is equally likely to be 1, 4, 20 or 50 units. Over two months the probability of no order is (2/3)² = 0.444, of exactly one order 0.444, and of two orders 0.111. Listing every outcome and its probability gives the cumulative distribution below.

Cumulative probability that demand over the 2-month lead time is at or below each level (empirical compound model).
Lead-time demand ≤01420244048505470100
Probability0.4440.5560.6740.8060.8330.8400.8400.9510.9790.9931.000

At the formula's reorder point of 48, the probability of covering lead-time demand is 0.84, not 0.95: the formula sits just below the 50-unit lump that carries 11% of the probability. The first level that reaches 95% is 50 units, so the empirical reorder point is 50 and the safety stock is 50 − 12.5 = 37.5 units. The full step-by-step calculation, and the Poisson and bootstrap alternatives, are in Safety Stock and Reorder Point for Intermittent Demand.

Where lumpiness comes from, and how to reduce it

Lumpiness is usually a property of the customers, not the product: a few large accounts ordering in batches, project or fleet demand, maintenance campaigns, or order quantities set by the customer's own min/max rules. That suggests the levers for reducing it: aggregate across locations or time buckets so that independent lumps average out, share forecasts or schedules with the handful of customers who generate the lumps, and separate known project demand from the baseline before forecasting. Where none of that is possible, the honest response is to accept a lower forecastability and buy protection with stock or lead-time agreements rather than with a more complicated model.

Frequently asked questions

Is lumpy demand the same as intermittent demand?

No. Both have many zero-demand periods, but intermittent demand has fairly consistent order sizes (CV² of the non-zero sizes below 0.49), while lumpy demand has order sizes that vary widely (CV² at or above 0.49). In the worked example the sizes 20, 1, 50 and 4 give CV² = 1.07, which makes the series lumpy rather than merely intermittent.

What is the best forecasting method for lumpy demand?

The usual starting point is SBA, the bias-corrected Croston method, which forecasts a stable demand rate (about 6 units per month in the example). For the reorder point, use the empirical distribution of lead-time demand (bootstrapping or a compound model) rather than the normal safety-stock formula, because the normal formula misplaces the large, rare orders that define lumpy demand.

How much history do I need before calling an item lumpy?

ADI and CV² can be computed from as few as four periods, but with so little data the classification is only indicative. Twenty-four or more periods at the planning bucket (weeks or months) give a stable read. Items close to either cutoff — ADI near 1.32 or CV² near 0.49 — should be treated as borderline rather than switched to a different method on a small change.

References

See also

Try it on your own items