Churn Prediction

Churn prediction is the gateway application of predictive analytics. Every subscription business needs it. The data requirements are manageable. The impact is immediately measurable. And the difference between detecting churn risk 60 days out versus 15 days out is often the difference between saving the customer and writing the refund.

PATCH and I built our churn model on 18 months of customer data. The model identifies at-risk accounts 45 days before traditional warning signs appear. At 45 days, ANCHOR has time to intervene — a health check call, a value demonstration, an executive alignment meeting. At 15 days, it is usually too late. The customer has already decided.

CHURN PREDICTION MODEL — TOP FEATURES
========================================

FEATURE                          │ WEIGHT │ SIGNAL
─────────────────────────────────┼────────┼────────
Login frequency trend (30d)      │ 0.24   │ Declining usage
Support ticket sentiment trend   │ 0.18   │ Growing frustration
Feature adoption breadth         │ 0.15   │ Shallow usage = risk
Champion engagement velocity     │ 0.14   │ Going dark
Contract utilization rate        │ 0.12   │ Paying for unused value
NPS/CSAT score change           │ 0.09   │ Satisfaction decline
Payment behavior change          │ 0.05   │ Delayed payments
Competitor website visits*       │ 0.03   │ Evaluating alternatives

* Where intent data is available

MODEL OUTPUT:
  Risk score: 0-100
  Risk tier:  Green (0-30) / Yellow (31-65) / Red (66-100)
  Risk drivers: Top 3 features contributing to score
  Recommended action: Tier-specific intervention playbook

PERFORMANCE:
  AUC-ROC: 0.87
  45-day advance detection: 78% of churned accounts
  False positive rate: 12% (acceptable for intervention cost)

The key insight in churn prediction is that the model does not just predict who will churn. It explains why they will churn. The risk drivers — the top 3 features contributing to the score — tell ANCHOR exactly what to address in the intervention. Declining login frequency triggers a usage enablement campaign. Sentiment decline triggers a relationship repair meeting. Low feature adoption triggers a value demonstration.

The economics: a 5% reduction in churn rate, achieved by intervening on high-risk accounts identified 45 days in advance, is worth 3-5x the investment in the predictive model. LEDGER tracks this ROI quarterly. It is the most cost-effective analytics investment in our stack.