Predictive Churn Model

Churn models score lapse probability from behavioral drift — session and deposit cadence changes, game-mix shifts, engagement decay — ahead of the inactivity threshold itself.

Deployment reality

Value comes from the intervention loop, not the model: scored risk feeding tested retention treatments (offers, content, contact cadence) with holdout measurement — plus the RG overlay this industry uniquely requires, since “about to churn” and “should perhaps churn” describe overlapping populations, and reactivating the wrong segment is a compliance finding waiting to happen. Feature honesty (no leakage), refresh cadence and per-segment calibration are the unglamorous 80%.