When history fails you, borrow from geography

Harrison Katz2 Jun 2026

Problem

A forecasting model trained on history fails confidently when the world changes: it gives tight, well-calibrated intervals around the wrong answer. At Airbnb this happened in the recovery from COVID, from late 2020 through 2022. Vaccines, border reopenings and reclosures hit each corridor (a pair of traveler origin city and destination city) at a different moment, so no corridor had a stable regime to learn from. Waiting for each market to collect its own data meant forecasting blind for months, just when Finance needed a read on demand.

Example from the article: booking lead times compressed in Europe in February 2020, and in North America four to six weeks later along the same path.

Solution in brief

Treat the markets hit first as leading indicators. After an early corridor observes its local data, its updated belief about its parameters becomes the prior for similar corridors that are hit later. A hierarchical Bayesian model does this naturally, and the transfer is weighted by how similar two corridors are.

Questions this case answers
  • How to forecast a group of related units when a shock or rollout reaches them at different times and the history of each unit is no longer a valid guide
  • How to transfer evidence from units that were affected first to similar units that have little or no data yet, weighting the transfer by similarity
  • How to forecast demand in markets that have little post-shock data when other markets were hit earlier?
  • How to reuse what an early-affected region already shows to forecast a region where the same change arrives later?

Tools and methods

  • Hierarchical Bayesian model — corridor-level demand parameters drawn from a shared population distribution
  • Prior propagation — the posterior of an early-affected corridor becomes the prior of a later one
  • Similarity weights between corridors — stronger transfer between structurally similar corridors
  • Mean booking lead time versus 2019 — the signal used to compare phases across regions
Continue reading — it's free

Google or a one-time email code. No password.