How to Validate you Client Churn Model
summary
Claims a survivals-based churn predictor built for ITV using the Cox Proportional Hazards model and the Telco dataset, fit with the Python lifelines package. Validation uses concordance (~0.86 raw, ~0.88 IPCW), Kaplan–Meier baseline (59% survival past 72 months), and residual checks: martingale and Schoenfeld residuals plus deviance and score residuals to spot nonlinearity, influential points and proportionality breaches, with stratification or interactions applied when assumptions fail.
AI-written summary · about
chapters (12)
- 0:00Client churn model intro
- 1:09Survival analysis for churn
- 3:14Data prep and features
- 6:05Survival analysis terms
- 8:57Demo: Kaplan-Meier
- 11:28Demo: Cox model
- 13:21Cox model assumptions
- 16:13Model accuracy with residuals
- 21:34Fixing model assumptions
- 23:45Demo: Concordance scores
- 27:00Advanced validation metrics
- 29:06Conclusion
related talks
35mUsing Survival Analysis to understand customer retention
36mCustomer Lifetime Value: Models, Metrics and a Multitude of Uses
50mHow to build propensity model where churn is single digit
32mClassifying the variety of customers online engagement for churn prediction
27mPreventing churn like a bandit
19mtidymodels for time-to-event data
34mModelling and predicting customer lifetime value
01h 24mBayesianism and Survival Analysis