Customer analytics · 2025
Churn Prediction
A gradient-boosted churn model with explainable drivers that lifted retention by 18%.
Churn Prediction
- Role
- Data scientist
- Timeline
- 4 months
- Team
- 1 data scientist, 1 analyst, 1 customer success lead
- Focus
- XGBoost, SHAP, SQL
The problem
The customer success team reacted to cancellations after they happened. They needed an early signal of which accounts were at risk, and a reason they could act on.
My role
I framed the problem, built the feature set from product usage and billing data, trained and validated the model, and presented the results to the customer success team.
Key decisions
- Predict a useful window. The model flags risk 60 days ahead, early enough for outreach to matter.
- Explainability first. SHAP drivers were shown next to every score so the team knew what to say to each customer.
- Measure the action, not just the AUC. We ran a controlled outreach test to confirm the model changed outcomes.
Outcome
Retention improved by 18% in the targeted group compared with the control group. The model is retrained monthly and monitored for drift. I would add uplift modelling next to target the customers most likely to respond.