> ML_RECIPE // SAAS-CUSTOMER-CHURN-PREDICTION_v1.0
SaaS Subscription Churn Prediction
Identify accounts at high risk of cancelling 60 days before contract renewal, enabling proactive CSM intervention.
Business Outcome
Identify accounts at high risk of cancelling 60 days before contract renewal, enabling proactive CSM intervention.
Model must achieve higher Top-Decile Lift than simple heuristic rule (days since last login).
Heuristic Baseline
Heuristic rule: Flag any account where primary admin login has been inactive for > 30 days.
Heuristic inactivity rule achieves PR-AUC 0.48 and Top-10% precision 0.38.
Phase 1: Prototype Path
Export 12 months of aggregated usage features from data warehouse. Train LightGBM baseline on developer laptop (CPU). Evaluate PR-AUC and Top-10% Capture Rate.
Phase 2: Production Path
Automate weekly feature transformations in dbt/Feast. Package LightGBM model as versioned artifact in MLflow. Run batch CPU inference on server with Evidently drift alerts.
Compute & Placement Topologies
Scheduled batch pipeline on CPU server (e.g. 4 cores, 16GB RAM)
Scheduled weekly batch job writing churn probabilities back to CRM
3-Plan Placement Alternatives
Local developer laptop cron script running monthly CSV export and scoring.
Multi-core CPU server running containerized batch inference on existing data warehouse cluster.
Scheduled dbt/Feast feature pipeline + MLflow artifact registry + containerized CPU batch inference with Evidently drift monitoring and Slack/CRM alerts.
Recommended Libraries & Tools
Governance, Safeguards & Risks
- Prohibit using protected demographic attributes as input features.
- Enforce model explainability (TreeSHAP) so customer success managers understand why an account was flagged.
- Maintain a shadow run for 30 days before triggering automated marketing emails.