How COVID Broke Forecasting Models, and How to Design for the Next Shock

In March 2020, Airbnb's forecasting model was trained on the assumption of stable patterns, and it proved powerless in the face of the pandemic shock. It wasn't simply that the numbers changed — the relationship itself between booking and actual travel shifted, and the lead-time composition collapsed.

The core solution was to decompose the forecast into two layers. First, gross metrics were forecast as an ordinary time series along the booking date axis. Second, the lead-time composition was modeled separately as a compositional time series that gets allocated across each future trip window. The final trip-date forecast is reconstructed as the product of these two components, and since the composition ratios must always be non-negative and sum to 1, consistency cannot be guaranteed by simply forecasting individual buckets separately.

To address this, the team built B-DARMA (Bayesian Dirichlet Auto-Regressive Moving Average). It respects the simplex constraint via a Dirichlet distribution, learns temporal dependencies through a vector autoregressive structure in log-ratio space, and jointly estimates uncertainty within a Bayesian framework. Extensions such as B-DARCH were later added, along with a shrinkage prior for high-dimensional compositions and a structural break intervention mechanism.

After the pandemic, gross booking volume recovered, but lead-time composition never fully returned to its previous state. To verify this, the team used normalized L1 distance, comparing monthly distributions against a baseline, and confirmed across four US cities from 2018–2022 that after the sharp shock, the pattern settled not into full recovery but into a partial recovery followed by a new normal. In particular, international travelers experienced greater lead-time compression and slower recovery than domestic travelers.

This distributional shift was not merely a statistical phenomenon — it was an operational issue that directly affected revenue forecasting, cash flow planning, and capacity decisions. So instead of erasing temporary anomalies with dummy variables, the model needed to learn the direction and speed of the structural change itself.

The alternative introduced was directional-shift intervention. This mechanism consists of a direction vector indicating the direction of change, an amplitude determining the magnitude of change, and a logistic gate controlling the timing and speed of the transition. Unlike a binary dummy, the logistic gate can represent both sudden shocks and gradual transitions, and can also project forward into future periods.

In a rolling one-step-ahead evaluation from July 2020 to March 2021, the intervention model reduced Aitchison distance by 31% compared to the baseline, and its 80% prediction interval coverage came close to matching the target level. By contrast, the baseline achieved only about 54%, and the dummy-variable model about 71%. The conclusion is clear: the pandemic was not transient noise but a change to the structure of the data itself, and forecasting models had to be redesigned to learn that structural change directly.