Intent-informed forecasting groups comparable conditional requests by product, variant, quantity, market, and time window, then calibrates their historical conversion rate. It should report uncertainty and never treat every request as an order.
- Exact variant and quantity are more useful than aggregate waitlist count.
- Commitment and confidence should be separate dimensions.
- Expired, declined, converted, and refunded outcomes improve calibration.
- Forecasts support decisions; they do not guarantee future sales.
Build cohorts that match the decision
A replenishment question needs product, variant, quantity, market, and time-window evidence. A substitute decision also needs tolerance data. Combining unlike requests into one total produces a large number with little operational meaning.
Separate commitment from model confidence
A shopper can make a strong declaration in a cohort with little history, or a weak request in a well-calibrated cohort. Track the shopper's commitment level separately from statistical confidence so teams can see both the demand signal and the uncertainty around it.
Backtest and show uncertainty
Store versioned forecast snapshots, compare them with later outcomes, and report calibration metrics over rolling windows. When a narrow cohort lacks evidence, fall back to a broader relevant cohort and label that fallback. Minimum sample warnings are part of the product, not a footnote.
Connect forecasts to reversible actions
Use forecasts to prioritize supplier conversations, allocation, merchandising, or a controlled campaign. Avoid treating a model output as authorization to purchase inventory or promise delivery. Human review should remain available for high-cost and low-confidence decisions.
Frequently asked questions
Are conditional requests the same as purchase orders?
No. They are potential demand until a merchant accepts and a shopper completes checkout. Forecasts must apply observed conversion rates and uncertainty rather than summing requests as revenue.
What outcomes should improve the forecast?
Accepted, countered, expired, declined, converted, refunded, and cancelled outcomes all help explain how similar requests behave over time.