Time series and forecasting · insight-1243
Forecast a new route or product with a global model on its attributes instead of waiting for local history
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- Data scientist, ML engineer
- Industries
- aviation, retail, e-commerce
Summary: situation, action, reason, boundary
Situation
A new route launches with three weeks of data, so a local seasonal model has nothing to estimate seasonal indices from.
Action
Train one global model on all established series with static attributes and calendar features, forecast the new series from its attributes and short history, and backtest the whole procedure by truncating each established series to its first few weeks.
Reason
A global model learns shared patterns across series, so it can produce a forecast from attributes where a local model cannot. Global methods also keep a constant complexity as the number of series grows.
Boundary
Launches, new stores, new routes and new SKUs that resemble existing ones.
Symptom or risk
Local models for new series either cannot be fitted or latch onto launch-period noise, and the planner falls back on a manual guess for the weeks that matter most.
Exceptions
A genuinely novel product with no analogue in the training set will be forecast like its nearest neighbours, and that may be wrong.
Trade-offs
Cold-start forecasts are wide and rest on the attributes you chose; they should be labelled and replaced as local history builds.
Pitfalls
Using attributes that were not known at launch, such as final capacity after a later upgauge.
Lessons and concepts
Claims and their sources
Montero-Manso and Hyndman show that global forecasting methods are not more restrictive than local ones, keep constant complexity as the number of series grows, and performed well even on heterogeneous collections.