Time series and forecasting · insight-0236
Scale each series by its own recent level when training a global model
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- aviation, retail
Summary: situation, action, reason, boundary
Situation
A global model across routes of 50 to 5,000 weekly passengers is accurate on big routes and poor on small ones.
Action
Divide targets and lag features by a per-series scale computed only from data before the origin (for example the mean of the last 8 weeks), train on scaled values, and rescale the forecast.
Reason
Scaling lets a shared model learn shapes that transfer across series of different sizes.
Boundary
Global ML and deep forecasting models over panels.
Symptom or risk
Unscaled targets make the loss dominated by large series, so the model learns their patterns and treats small series as noise.
Exceptions
If large series truly matter more for the business, keep a volume-weighted loss but still scale inputs.
Trade-offs
Series with a near-zero recent mean need a floor on the scale to avoid exploding values.
Pitfalls
Computing the scale over the full series, which leaks the future level into training.