Preventing churn like a bandit
summary
Companies lose 2–16% of monthly revenue to churn and the practical solution ties uplift to money: predict uplift directly (use the transform outcome trick), multiply Δ retention by customer lifetime value minus treatment cost to rank actions, handle missing counterfactuals with masked multi‑output regression, remove assignment bias via inverse propensity weighting with clipping and calibration, and balance exploration versus exploitation with Thompson sampling plus top‑list randomization so the policy actually learns what works.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:08Churn and ML approaches
- 3:19Predicting churn alone fails
- 5:26Predict uplift instead
- 8:40Transform outcome trick
- 11:56Counterfactuals in treatments
- 14:29Reinforcement learning for churn
- 15:32Correlation vs causation
- 17:16Random trials limitations
- 18:40Inverse propensity weighting
- 21:50Exploration vs exploitation
- 22:52Thompson sampling
- 24:21Budget constraints
- 26:49Final solution
related talks
24mUnlocking revenue: newcomer classification for boosted profits
30mWinning (Some) Churners Back
13mOptimizing Subscription Services with Uplift Modeling
50mHow to build propensity model where churn is single digit
40mNext Best Action: Maximizing Customer Value
01h 00mHow to win at monetizing games as a service
38mHow to Enhance Customer Targeting in Marketing
30mDetecting Hidden Bias in Datasets Before Models Fail