← Back to all tipsNovember 13, 2026

Why intermittent demand needs a probability, not a point forecast

What makes time series forecasting hard ends on the honest point that some variation in any series is genuine noise, not signal — and a good forecast says so instead of hiding it behind one confident number. Intermittent demand is the case where that honesty stops being optional, because the noise isn't a small correction around a stable average. Most periods are exactly zero.

When most periods have zero demand

Spare parts, low-turnover B2B products and anything ordered irregularly rather than routinely all produce series that look nothing like smooth retail sales — long stretches of zero, interrupted by occasional, often uneven spikes. Standard forecasting methods are built around smoothing noise around a trend; applied directly to a mostly-zero series, they tend to either predict near-zero forever (missing every spike) or chase the last spike into a forecast that never comes back down. Dedicated methods exist for exactly this shape of data — Croston's method and its refinements, including the Syntetos–Boylan approximation, separately model how often demand occurs and how large it is when it does, instead of forecasting the raw series directly.

Why a single number misleads here

Even with the right method, a forecast of "7 units next month" for a part that has sold 0, 0, 0, 22, 0, 0, 4 units over the last seven months isn't wrong exactly — it's just the wrong shape of answer. No single month in that history looked like 7. The number is a long-run average standing in for a genuinely uncertain outcome, and presented alone it invites exactly the wrong kind of confidence: ordering as if 7 is what will actually happen, rather than planning for the real range of what might.

point forecast
A single flat forecast number following a sparse, spiky demand history "7 units"
probabilistic forecast
A widening likely range following the same sparse, spiky demand history "2–14 units"
Same history, same model family — one output collapses the uncertainty, the other reports it.

What changes operationally

A range instead of a point forecast isn't a hedge, it's more information: a narrow range justifies ordering close to the median; a wide one justifies carrying safety stock or accepting occasional stockouts as a deliberate trade-off rather than a planning failure. Expressing demand as a distribution also composes better across a network — summing uncertain demand across several locations or SKUs is a well-defined operation on distributions, and a much blurrier one on a list of single numbers that were never meant to add up precisely in the first place.

This isn't new, just newly urgent

Probabilistic demand forecasting is decades-old statistical practice, not a 2026 invention. What's changed is the operating environment it's being asked to handle: more frequent demand shocks, shorter product lifecycles, and planning cycles measured in days rather than quarters all make a single static number stale faster than it used to go stale. The response isn't a fundamentally new method — it's applying the honest, range-based version of forecasting more broadly, and refreshing it more often, than a calmer market ever required.