Time series and forecasting · insight-0240
In direct models, keep every lag at least as long as the horizon it predicts
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- Data scientist, ML engineer
- Industries
- retail, aviation
Summary: situation, action, reason, boundary
Situation
A direct model for 14-days-ahead demand uses lag 7 sales as a feature.
Action
For a direct model at horizon h, build lags relative to the origin, so the shortest usable lag is h. Generate features per horizon or add horizon as a feature with origin-based lags.
Reason
A feature is available only if it is observed by the origin; lags shorter than the horizon refer to unobserved periods.
Boundary
Direct and multi-horizon ML forecasts.
Symptom or risk
At forecast time the value seven days before the target is still in the future, so the feature is missing in production or quietly filled with stale data.
Exceptions
Future-known regressors such as calendar or schedule can be aligned to the target date.
Trade-offs
Origin-based lags lose the most recent dynamics for long horizons, which partly explains rising error with horizon.
Pitfalls
Building a single feature table keyed by target date and reusing it for all horizons.