Retraining isn't one decision: Airbnb splits refit, respecify, hold
Airbnb's forecasting team treats "retrain" as three distinct decisions, not one. Full refits carry hidden risk — refreshed data can actually degrade a model that was fine. The discipline that emerged from the pandemic unwinding: separate ordinary parameter drift from structural change, then choose between a cheap refit, a costly model respecification, or holding entirely. Their method is a field guide for any team whose forecasts others build decisions on top of.