Spare parts, seasonal niche items, and the long tail of any catalogue share a demand pattern that breaks most forecasting tools: lots of zeros, with an occasional spike. Feed that to ARIMA or Prophet and you get a smooth, confident, and wrong forecast. Knowing why, and what to use instead, is the difference between sensible stocking and chronic dead stock on slow movers.
Why the popular tools fail here
ARIMA and Prophet are built for series with a reasonably continuous signal: trend, seasonality, and noise around a moving level. Intermittent demand has none of that between the spikes, it is zero. Two failure modes follow:
- They smear the spikes into a low, steady average. A part that sells 10 units once a quarter becomes a forecast of “about 0.11 per day,” which is useless for a reorder decision.
- Their uncertainty is mis-shaped. The interesting question for a slow mover is “when will the next order come and how big,” not “what is tomorrow’s level.” Continuous models do not answer it.
The result is quiet: the forecast looks plausible, so nobody questions it, while you either stock out on the spike or carry the item for years.
What to use instead
The key idea, due to Croston, is to forecast two things separately: the size of demand when it occurs, and the interval between demands.
- Croston’s method smooths demand size and inter-demand interval independently, then combines them. It is the baseline for intermittent demand.
- SBA (Syntetos-Boylan Approximation) corrects a known bias in Croston’s that makes it over-forecast. For most intermittent series, SBA is the better default.
- TSB (Teunter-Syntetos-Babai) handles items whose demand is also obsolescing, updating the probability of demand each period, so it copes with parts going end-of-life.
# illustrative: Croston-style separation
# on a demand period: z = smooth(demand_size); x = smooth(interval_since_last)
# forecast per period = z / x
# (SBA multiplies this by (1 - alpha/2) to debias)How it feeds stocking
Because intermittent demand is lumpy and non-normal, the forecast hands off to a safety-stock
calculation that does not assume a bell curve, exactly the trap covered in
why your safety stock formula breaks on non-normal demand.
For the spiky tail, an empirical or distribution-specific buffer beats z * sigma.
The practical rule
Classify your items first: separate the smooth, fast movers (where ARIMA/Prophet or simple exponential smoothing are fine) from the intermittent tail, and route the tail to Croston/SBA/TSB. Trying to force one method across the whole catalogue is how slow movers quietly wreck both service and demand forecasting accuracy.
Implementing this at your scale?
The walkthrough above comes from production work. AvanSaber’s inventory practice has implemented variations of this pattern across multiple customer engagements.
If you are building this and want expert review of your design, or would rather have the team that built this build yours, book a discovery conversation or describe your situation at [email protected].