When history fails you, borrow from geography

This post details how Airbnb adapted its forecasting models to handle unprecedented, asynchronous shocks like the COVID-19 pandemic. When historical data became unreliable, the team leveraged sequential geographic recovery signals and prior propagation. This Bayesian approach uses the posterior distribution from an early-affected corridor as an informative prior for a later-affected corridor, enabling more accurate forecasts in markets with scarce local data. The system's effectiveness is attributed to Airbnb's global operational breadth and the structural similarity between origin-destination corridors, allowing for timely and reliable demand projections during periods of uneven recovery.