Demand Forecasting Wiki

Intermittent and Lumpy Demand

By MLAIA Data Science Ltd. · Published 13 July 2026

Intermittent demand is a demand pattern in which items are requested only sporadically, so that many time periods record zero demand. When the size of those occasional demands also varies greatly, the pattern is called lumpy demand. These patterns dominate spare parts, aftermarket components, and maintenance, repair and operations (MRO) inventories, and they defeat forecasting methods that assume regular, roughly normally distributed demand.

Definitions

Intermittent demand
Demand that occurs sporadically, with zero demand in many periods. The demand incidence (whether an order arrives) is irregular, while the sizes of the orders that do arrive are comparatively stable.
Lumpy demand
Intermittent demand whose non-zero order sizes are themselves highly variable. A part may sell nothing for months, then two units, then forty.
Erratic demand
Demand that occurs in most periods (low intermittency) but with highly variable sizes.
Smooth demand
Demand that occurs regularly with relatively stable sizes — the pattern most textbook forecasting methods implicitly assume.

The ADI–CV² classification

The standard way to classify a demand series uses two statistics, computed from its history:

Syntetos, Boylan and Croston (2005) derived cutoff values for these two statistics by comparing the theoretical forecast errors of Croston's method and its bias-corrected variant across demand patterns. The resulting scheme, with cutoffs at ADI = 1.32 and CV² = 0.49, has become the de facto standard in both research and practice:

Demand-pattern quadrants after Syntetos, Boylan and Croston (2005).
CV² < 0.49 (stable sizes) CV² ≥ 0.49 (variable sizes)
ADI < 1.32 (frequent) Smooth — regular timing, stable sizes Erratic — regular timing, variable sizes
ADI ≥ 1.32 (sporadic) Intermittent — sporadic timing, stable sizes Lumpy — sporadic timing, variable sizes

The cutoffs are pragmatic decision boundaries, not laws of nature: a series with ADI = 1.30 is not fundamentally different from one at 1.35. Their value is in giving a large item portfolio a consistent, automatable segmentation, which can then drive method selection per item.

Why intermittency breaks conventional forecasting

Most classical forecasting and inventory theory assumes demand per period is a well-behaved random variable — typically approximately normal, with observations in every period. Intermittent series violate this in several ways:

Where intermittent demand occurs

Intermittent and lumpy patterns are the norm, not the exception, in several settings:

Forecasting approaches

The established approaches to intermittent series fall into three families: dedicated statistical estimators (Croston's method and its SBA and TSB variants), empirical methods that resample the observed history (bootstrapping, following Willemain, Smart and Schwarz, 2004), and machine-learning models that pool information across many items. Which family is appropriate depends on the demand pattern, the data available, and the inventory decision the forecast feeds — a topic covered in Demand Classification for Forecasting.

Frequently asked questions

What counts as intermittent demand?

Demand is usually treated as intermittent when the average interval between periods with any demand (the ADI) is greater than about 1.32 periods — that is, when a substantial share of periods record zero demand. The exact threshold comes from the demand-categorization research of Syntetos, Boylan and Croston (2005).

What is the difference between intermittent and lumpy demand?

Both patterns have many zero-demand periods. Intermittent demand has relatively consistent order sizes when demand does occur (CV² of the non-zero sizes below about 0.49), while lumpy demand combines sporadic timing with highly variable order sizes (CV² at or above 0.49). Lumpy demand is the hardest standard pattern to forecast.

Why is MAPE a poor error metric for intermittent demand?

MAPE divides each error by the actual demand in that period. When many periods have zero actual demand, the division is undefined, and near-zero actuals produce exploding percentages. Scale-free alternatives such as MASE, or service-level-oriented measures, are generally preferred for intermittent series.

Can exponential smoothing or ARIMA be used on intermittent demand?

They can be computed, but simple exponential smoothing applied directly to an intermittent series is biased immediately after each demand occurrence and drifts toward zero in between, and ARIMA's Gaussian error assumption fits zero-inflated data poorly. Methods designed for intermittency — Croston's method, SBA, TSB, or bootstrapping — are the standard starting points.

References

See also