Time series and forecasting · insight-0227

Represent a known level shift with a step regressor or a post-break training window

Situation
A competitor added about 25% more seats on three routes from a known date, and the route forecasts now run high.
Action
Add a step variable that switches on at the break date, or restrict training to post-break data once enough has accumulated, and compare both in a backtest with origins after the break.
Reason
A step variable captures a permanent level change, letting the rest of the model use the full history.
Boundary
Competitor entry, policy or tariff changes, product reformulations, new reporting systems.

Symptom or risk

The model averages pre- and post-break levels, so it over-forecasts after the break and its intervals are too wide or misplaced.

If the break changes seasonality or dynamics, not just level, a step dummy is not enough; retrain on post-break data or interact the dummy.

Post-break-only training is cleaner but has little data at first.

Detecting breaks after the fact and then backtesting as if the break date had been known at every origin.

The synthetic airline simulator steps up competitor capacity on IST-ADB, IST-AYT and IST-FRA on 2024-03-01.

A step variable takes value zero before an intervention and one after it, capturing a permanent level shift.
Forecasting: Principles and Practice (3rd edition)