Churn prediction — start rule-based, mature into ML

A scoring job that runs daily over the warehouse, flagging accounts at elevated churn risk. V1 is a rule stack: transaction volume drop > 40%, support tickets up 3×, login frequency declining 2 weeks running. V2 layers a light gradient-boosted model trained on historical churn labels. Output writes to the CRM as a churn_risk field that triggers a retention workflow.

A mid-size payments group sized 1% churn reduction at ~$1.4M/year on an 80K-merchant base; even rule-based v1 is enough to capture most of it. Crucially, you do NOT need a warehouse to ship v1 — a few SQL queries against the production billing DB plus a CRM API write are enough to validate before investing in DWH.

1. List 3–5 signals you can compute today from billing + product DB.
2. Hand-tune thresholds with one CS lead over a week — calibrate to historical churners.
3. Schedule a daily SQL job that writes churn_risk_score into the CRM.
4. Wire a retention play per tier. Low risk = nothing. Medium = automated email + CS heads-up. High = call from named manager within 24h.
5. Track outcomes for 60 days. Compare retention of flagged-and-acted vs flagged-and-ignored.
6. Once you have 6+ months of labels, train a gradient-boosted model on the same features; ship the model output as an additional CRM field side-by-side with rules.
7. Retire rules selectively as the model proves stable per segment.

• "We can't build churn until we have a warehouse" is a six-month delay disguised as engineering hygiene. Start rule-based against the prod DB; the warehouse build runs in parallel.