Start with these five checks.
- Match the metric to the planning decision.
- Compare against naive and seasonal baselines.
- Separate bias, scale and intermittent demand.
- Correct for stockout-censored sales.
- Audit events, overrides and hierarchy reconciliation.
Ask for only what can change the answer.
- ForecastsVersion, horizon and generated time
- Actual demandOrders, lost sales and substitutions
- AvailabilityStockout and fulfillment limits
- EventsPromotion, launch, holiday and price
- HierarchySKU, location and category relationships
- OverridesPlanner change, reason and result
Test one reversible move.
Freeze an older training period and an untouched later evaluation period. Change one element—event feature, hierarchy rule or override policy—and compare error, bias and downstream inventory decisions by segment. Reject it if average accuracy rises while priority items, service or working capital worsen.
A decision your team can use.
- 01A decision-grain baseline
- 02Bias and error segmentation
- 03Override and event analysis
- 04A time-correct holdout test
Common questions.
Which forecast metric is best?
Use metrics suited to scale and intermittency, and always connect them to the operational decision.
Why compare with a naive baseline?
A complex method has not added value if it cannot beat a simple recent or seasonal forecast.
How do stockouts affect actual demand?
Observed sales are censored when inventory is unavailable; lost sales and substitutions need explicit treatment.