Overview

Churn Prediction Model turns product telemetry into actionable risk scores so retention teams can focus on the right accounts at the right time. The emphasis was on leakage safety, explainability, and operational usability — not just offline AUC.

Goal: prioritize outreach while surfacing clear drivers. Constraint: build trust with stakeholders through transparent methodology.

Data & labeling

I designed a leakage-safe dataset with fixed observation windows and future outcome windows. Feature engineering focused on product adoption signals, usage decay, and account health patterns.

Windows: consistent lookback + prediction horizon
Features: frequency, recency, depth of usage, and trend deltas
Quality: missingness checks and outlier handling

Modeling

Trained a gradient-boosted classifier, calibrated probabilities, and built a simple “why this is risky” explanation summary for each account.

Training: time-aware splits to avoid future leakage
Calibration: reliability curves for usable probabilities
Explainability: driver summaries for success teams

Production & monitoring

The model ships with drift monitoring, threshold alerts, and weekly performance reporting. Documentation includes rich text examples with bold warnings, italic assumptions, underlined action steps, and links to artifacts.

Dashboards: risk dashboard and monitoring view
Notebook: analysis notebook

Outcome

Retention operations improved with prioritized outreach lists, clearer risk drivers, and fewer wasted touches. The team gained a repeatable pipeline for iteration.

Projects

Churn Prediction Model

Experimentation Metrics Pipeline