Churn models need behavioral signals, not just demographics. The most predictive features tend to be changes in behavior — a customer who logged in daily and now logs in weekly is a stronger signal than a customer who always logged in weekly. Collect: login timestamps, feature usage events, support tickets (volume and sentiment), billing events, NPS/CSAT responses, and session duration trends. Store as time-series so you can compute rolling averages and deltas.

Gradient boosting models (XGBoost, LightGBM) are the industry standard for churn scoring. They handle mixed feature types, missing data, and non-linear relationships well. The model outputs a probability (0–1) for each customer. Key signals typically include: declining activity velocity, increasing support ticket frequency, reduced feature breadth, and payment failures. Train on historical churn labels with a rolling window approach to avoid data leakage.

Don’t just use arbitrary thresholds (e.g., > 0.7 = high risk). Instead, calibrate thresholds based on your intervention capacity. If your team can handle 50 save-calls per week, set thresholds so the “high risk” segment is roughly that size. Use precision-recall curves to find the threshold that maximizes caught churners while keeping false positives manageable. Revisit thresholds monthly as your model and customer base evolve.

Match intervention cost to risk tier. High risk: personal outreach (CSM call, executive sponsor), discounts or contract flexibility, urgent product feedback loop. Medium risk: automated nurture sequences, targeted feature education, proactive check-ins. Low risk: community engagement, success story sharing, upsell opportunities. Track intervention success rates by tier to continuously improve your playbook. Measure incremental retention — compare intervened customers vs. a holdout group.