Methods
Croston separately smooths nonzero sizes and demand intervals. SBA multiplies the Croston level by 1 − α/2 to reduce its known upward bias. TSB smooths demand size and occurrence probability every period, allowing the forecast to decay during a run of zeroes. The historical mean is the simple baseline each method should earn the right to beat.
Use and limitations
The output is a demand baseline, not a reorder point, safety stock, or order quantity. It has no prediction interval and does not account for criticality, stockout-censored history, installed-base changes, planned overhauls, substitutions, MOQ, or lead time. Validate it with rolling origins and local consequence review.
Common mistakes
- Deleting zero periods or mixing weekly and monthly data.
- Tuning α and β on the same holdout used to report performance.
- Calling a declining TSB probability proof of obsolescence.
- Multiplying a point forecast directly into a stocking rule without uncertainty or consequence.