Time series and forecasting · insight-1246
Route intermittent series to a method by interval and size variability, then confirm by backtest
- Evidence type
- empirical finding
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- manufacturing, aviation, logistics
Summary: situation, action, reason, boundary
Situation
A spare-parts catalogue of 20,000 items is forecast with a single method, although some items sell steadily and others sell every few months in varying quantities.
Action
Compute, from data before the origin, the average interval between non-zero demands and the squared coefficient of variation of the non-zero sizes. Split series into classes, test a few candidates per class, and keep the routing only if it beats one method for everything.
Reason
Syntetos, Boylan and Croston derived demand-pattern classification rules in terms of exactly these two quantities to indicate where Croston-type methods differ in accuracy, and tested them on 3,000 real series.
Boundary
Spare parts, slow-moving retail and maintenance consumption.
Symptom or risk
One method suits one demand pattern, so the same setting is too reactive for the sparse items and too sluggish for the steady ones.
Exceptions
With few series a single, bias-corrected method is simpler and loses little.
Trade-offs
Routing adds a classification step that can flip as history accumulates and make forecasts jump.
Pitfalls
Taking cut-off values from a secondary page without recomputing the class boundaries on your own data.
Lessons and concepts
Claims and their sources
Syntetos, Boylan and Croston (2005) define demand-pattern categories using the average inter-demand interval and the squared coefficient of variation of demand sizes, derived by comparing forecasting methods and tested on 3,000 real intermittent series.