Time series and forecasting · insight-1240
In a direct model, align future-known drivers with the target date and history features with the origin
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- retail, e-commerce, aviation
Summary: situation, action, reason, boundary
Situation
A per-horizon boosted model for h = 1 to 14 days is built by shifting every column, including the promotion calendar, by h days.
Action
Build each horizon's row from three blocks: history features at or before the origin, future-known drivers at the target date (calendar, price plan, scheduled capacity), and the horizon itself. Audit that every feature was available at the origin.
Reason
A future-known covariate is legitimately available for the target date, whereas anything that is unknown at the origin must be limited to the information up to the origin.
Boundary
Direct and multi-output gradient-boosting or linear forecasters with promotion, holiday, price-plan or schedule inputs.
Symptom or risk
The promotion flag for the day being predicted is replaced by the flag from the origin day, so the model never sees the one driver it could know with certainty.
Exceptions
Drivers that are only planned and may change before the target date need the vintage treatment in the previous card.
Trade-offs
Joining at the target date needs the future calendar to be reliably populated in production.
Pitfalls
Shifting the entire feature frame by the horizon in one call, which silently misaligns the known-future block.