What COVID did to our forecasting models (and what we built to handle the next shock)

Airbnb explains how COVID broke their booking-to-trip forecasting models and the architectural changes they built to handle structural data shifts.

• •Pre-pandemic models treated booking volume and lead-time distribution as an integrated process, which failed when both shifted simultaneously during lockdowns
• •The key fix was decomposing forecasts into two components: gross booking-date metrics and lead-time composition as a separate compositional time series
• •They developed B-DARMA (Bayesian Dirichlet Auto-Regressive Moving Average) to model lead-time compositions, enforcing simplex constraints and providing calibrated uncertainty bands
• •Post-pandemic discovery: lead-time distributions did not fully revert, especially for international guests whose booking horizons compressed and recovered more slowly