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

Harrison Katz30 Mar 2026

Problem

Airbnb finance forecasters predict metrics by trip date, but bookings arrive on booking date. A booking made today can become a trip in three days or three months, and the share of each gap length (the lead-time composition) turns today's bookings into future revenue. In March 2020, lockdowns and reopenings moved booking volume and lead-time composition at once. Joint models could not tell a volume surge from a behaviour change. Revenue and cash-flow planning inherited that error.

Example from the article: by late 2021 gross bookings had recovered, but long lead-time shares (4+ months) stayed below their pre-pandemic level through 2022.

Solution in brief

Forecast gross volume on the booking-date axis and the lead-time composition separately, then multiply them to get trip-date metrics. Model the composition with a Bayesian Dirichlet ARMA (B-DARMA) that respects the sum-to-one constraint. Handle a structural break with a directional-shift term: direction vector, amplitude and logistic gate. Track a normalized L1 distance to a baseline composition to see persistent shifts.

Questions this case answers
  • How to forecast a delayed outcome when the lag distribution between the triggering event and the outcome changes structurally after a shock
  • How to forecast compositional data (shares that sum to one) with calibrated uncertainty through a persistent structural break
  • How to monitor distribution shape drift that point statistics and top-line volumes do not reveal
  • How to forecast event-date revenue when the gap between booking and event keeps changing after a shock?

Tools and methods

  • Two-part forecast — gross metric by booking date times a lead-time composition vector.
  • B-DARMA — Dirichlet likelihood for proportions, ARMA dynamics, full predictive distributions.
  • Directional-shift term with logistic gate — learns direction, size and speed of a break.
  • Normalized L1 divergence — distance from a pre-shock composition, for monitoring.
  • Stan — estimation software (per the paper).
Continue reading — it's free

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