Why the usual methods produce nonsense here
A spare part that sells nothing for five weeks then four in one day has an average weekly demand somewhere under one. A conventional forecast will happily output 0.8 units a week, which is not something you can stock, and will look wrong in every single week - too high in the zeros, too low in the spike.
The mistake is treating this as the same problem as forecasting a fast mover with a bit more noise. It is a different problem, and it needs a method that separates whether demand occurs from how big it is when it does.
Split the question in two
The practical approach is to model two things separately: the chance that any demand occurs in a period, and the typical size of an order when it does. Croston's method and its variants formalise this, and it is also a sensible way to think even if you never name the technique.
Separating the two questions gives answers you can act on. 'Roughly a one in six chance of demand in any given week, and when it comes it is usually two or three units' tells a planner far more than '0.4 units per week'.
Judge it on stock outcomes, not percentage error
Percentage-based accuracy metrics are close to meaningless on intermittent items, for the reasons set out in our piece on forecast accuracy metrics. The denominator is frequently zero and often one.
Score the thing you actually care about instead.
- Service level achieved - what share of demand was met from stock
- Stockout events per item per year, weighted by how urgent that part is
- Average stock held, in units and in cash
- Write-offs from obsolescence at end of life
These four together describe whether the policy is working. A percentage error does not.
Classify before you forecast
Not every slow-moving item deserves a model. A useful first step is to split the catalogue by how often demand occurs and how variable the size is when it does.
| Pattern | Description | Sensible approach |
|---|---|---|
| Smooth | Regular demand, modest variation | Ordinary forecasting methods work |
| Intermittent | Frequent zeros, consistent size | Croston-style split; simple reorder point |
| Erratic | Regular demand, wildly varying size | Focus on the size distribution, hold buffer |
| Lumpy | Frequent zeros and varying size | Hardest case - often a rule beats a model |
Lumpy items are where forecasting projects quietly fail. For those, a well-chosen reorder point and an honest conversation about service level usually beats anything statistical.
When a rule is the better answer
For a large share of slow-moving catalogues, the right answer is not a forecast at all. It is a reorder policy: hold this many, reorder when it drops to this level, review annually.
That is unglamorous and frequently correct. We have written separately about when a rule beats a model, and spare parts are one of the clearest examples. The value of the analysis is often in setting the reorder points sensibly by criticality and lead time, not in predicting next week.
On a lumpy catalogue, the win is usually a better policy, not a better prediction.