Problem
Expedia Group needs to know how much each customer will bring over the next year and beyond. The answer drives decisions on acquisition, retention, incentives, marketing messages and long-term investment. Cohort analysis, RFM and Buy-Till-You-Die models fall short: they work only for predefined segments, use only purchase history, and need several purchases per customer, so new customers get poor estimates. The company has hundreds of millions of customers across eight brands and five lines of business.
Example from the article: one EG-wide CLV model (in fact 30 segment models) scores customers of Expedia, Hotels.com, Vrbo and other brands. It covers Stays, Flights, Packages, Cars and Cruises, and the scores refresh every day.
Solution in brief
Treat CLV as supervised regression on customer-level features. Features are computed before a cutoff date, and the target is the cash flow after it. The authors train one CatBoost model per customer segment (30 segments: region by booking recency and frequency), evaluate with Gini, Lorenz curves and calibration plots, and run the pipelines on a batch ML platform.
Questions this case answers
- How to predict the long-horizon value of each customer at scale when cohort, RFM and probabilistic repeat-purchase models cover only predefined segments or need several purchases per customer
- How to evaluate a customer-level value model for both ranking quality and calibration when the true outcome only appears after a long delay
- How to predict the 12-month future value of every customer when the customer base has hundreds of millions of people?
- How to get customer lifetime value for new segments and for customers with few purchases without redoing the calculation?
Tools and methods
- CatBoost — one gradient-boosted model per segment
- Cutoff-date dataset — features before the date, target after it
- Gini coefficient, Lorenz curve, calibration plot — ranking and calibration checks
- Apache Spark, Apache Hive, Apache Airflow, GitHub Actions — data processing, scheduling and CI/CD
Google or a one-time email code. No password.