Time series and forecasting · insight-0238
Choose recursive, direct or multi-output strategies by backtest at your horizons
- Evidence type
- empirical finding
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- logistics, finance, retail
Summary: situation, action, reason, boundary
Situation
A one-step-ahead ML model is fed its own predictions to produce a 28-day forecast.
Action
Backtest recursive, direct and multi-output versions with identical features and origins, and compare per-horizon error and the maintenance cost of each.
Reason
In one large comparison on NN5 series, multiple-output strategies performed best, but results depend on the data and models, so test rather than assume.
Boundary
Any ML or neural forecast beyond one step.
Symptom or risk
Recursive forecasting compounds errors over the horizon and needs forecasts of every exogenous input; a direct model per horizon avoids that but ignores dependence between horizons.
Exceptions
Classical state-space models such as ETS produce coherent multi-step forecasts natively.
Trade-offs
Direct needs one model per horizon (or a horizon feature); multi-output models are harder to explain.
Pitfalls
Evaluating the recursive model with true lags rather than its own predictions.
Lessons and concepts
Claims and their sources
A review of multi-step strategies on 111 NN5 series found multiple-output strategies performed best and that deseasonalisation uniformly improved accuracy.