How to build propensity model where churn is single digit
summary
A telecom with single-digit churn (around 3%) builds microsegments by Cartesian-producting three CDR-derived features: daytime work/home cell clusters, spending band, and recharge timing, labels them honeycomb, and models each segment’s 30-day usage via spectral analysis (sums of sine/cosine bases). A daily error heatmap on segment-level call-network graphs triggers interventions on yellow/orange groups; this group-level propagation approach lifts retention lead-time from ~8–9 days to 15–18 days and increases recoverable revenue from ~8% to ~60%.
AI-written summary · about
chapters (6)
- 0:00Start
- 6:26Confusion Matrix
- 10:01Induced Charge
- 23:04Spectral Analysis
- 30:29Call Pattern
- 42:11Call Center Data
related talks
27mPreventing churn like a bandit
31mHow to Validate you Client Churn Model
36mCustomer Lifetime Value: Models, Metrics and a Multitude of Uses
24mUnlocking revenue: newcomer classification for boosted profits
32mClassifying the variety of customers online engagement for churn prediction
35mUsing Survival Analysis to understand customer retention
30mWinning (Some) Churners Back
06mCellPred: A Behavior-aware Scheme for Cellular Data Usage Prediction