Problem
A marketplace needs to know which listings are worth more to guests, to decide where to add supply, which hosts to coach and which campaigns pay off. Classic lifetime value (LTV) work assumes one seller and many buyers. With many sellers, a listing's bookings may simply move from other listings, so raw bookings overstate the value it adds. Predictions also take a year to verify, and a shock such as COVID-19 can break a model trained on earlier data.
Example from the article: Airbnb runs a six-month campaign that gives hosts tips for improving listings, and must compare the extra LTV it created with its cost.
Solution in brief
Airbnb splits listing LTV into baseline LTV (a forecast of bookings over the next 365 days), incremental LTV (baseline minus bookings taken from other listings) and marketing-induced incremental LTV (value created by an internal initiative). Cannibalization comes from a supply-demand production function fitted on segment data. Forecasts are corrected daily with actual bookings.
Questions this case answers
- How to estimate the incremental long-term value of a supply unit in a two-sided marketplace when its transactions may be taken from other units and no incrementality label exists
- How to evaluate a campaign whose outcome is a prediction over a long horizon, before the horizon has passed
- How to estimate the lifetime value of each supply unit (a listing) in a two-sided marketplace?
- How to separate the bookings a new listing adds from the bookings it takes from other listings?
Tools and methods
- Baseline LTV model — predicts bookings over the next 365 days from listing features
- LightGBM — gradient boosting that handles high-cardinality features such as geography
- Supply-demand production function — bookings as a function of segment supply and demand; gives incrementality
- Daily forecast updating — blends accrued bookings, current features and arrival patterns of similar listings
Google or a one-time email code. No password.