Data & drivers
Intermittent demand: when most days are zero
Spare parts, rare claims, cyber events, slow B2B orders — the series is mostly zeros, then a jump. Average error metrics lie. You often care about hit rate and size when a hit happens, not a daily mean of 0.4 units.
Updated Aug 26, 2026·8 min read
The shape of a rare event
Spare parts, export orders, certain claims, cyber incidents: most ticks are zero, then a burst. Grocery-trained intuition fails. A daily neural net that predicts zero will look brilliant on MAE and miss the only week that emptied the bin.
Most days are zero. Then a jump.
Metrics that lie
MAPE explodes or is undefined. WAPE on a mostly-zero series can make a flat-zero forecast look saintly. You often want a split: probability of a hit, and size when it hits. Or you change the grain until zeros are no longer the majority.
- Croston-style methods split 'when' from 'how much.' Put them in the race. They are one candidate, not the only method.
- MASE with a sensible seasonal period is less of a liar than MAPE.
- If ops can live with weeks, weekly grain is often the whole fix.
Grain is a lever
Forecasting daily gaskets is a flex. Forecasting weekly replenishment for that gasket is a product. Same for rare ED transfers, or fraud reports in a small region. Aggregate until the series has a spine, then disaggregate with a share if you must.
Do not force a grocery model on gaskets
Put a seasonal naive and a specialist in the same Predict.ai tournament as the booster. The goal will not insist on a dense daily model if weekly wins the folds. For maintenance parts, medical SKUs that move in bursts, or cyber event counts, that honesty is the product.
Driver discovery still helps: a production-run signal that leads spare demand, a weather extreme that leads certain claims. If nothing leads, you are in inventory-policy land (base stock, min-max) with a weak forecast. Say that. Do not decorate it.
FAQ
- What counts as intermittent demand?
- A series with a high share of zeros and occasional bursts. Spare parts are the textbook. Incident counts, certain claims, and export orders can look the same.
- Is Croston still used?
- Yes, as a candidate. It splits 'when' from 'how much.' It is not the only method. If weekly grain removes the zeros and ops can live with weeks, that is often the better product.
Keep going
Guide
WAPE, explained
WAPE (weighted absolute percentage error) is total absolute error divided by total actuals. It behaves when some periods are zero, and it does not let tiny SKUs dominate the score the way MAPE does.
Guide
Forecasting new products and locations
A cold start is a forecast with almost no history of its own. You borrow shape from similar items or places, use leading signals, keep the horizon short, and replace the borrowed prior as soon as real sales show up.
Guide
Claims and incident volume
Fraud queues and claim shops live on volume: how many will land next week, by type and region. That is a forecast. Who is risky is a score. Mixing AUC into a staffing plan is how you get a confident dashboard and an overtime surprise.
Use case
Equipment failure forecasting
Estimate failure risk and remaining useful life before equipment interrupts production.
Use case
Medical supply demand forecasting
Forecast medication, consumable, blood-product, and critical supply demand by facility and care setting.