Churn prediction that pays for itself

Challenge

Retention spend was sprayed across the whole base. The team could not tell who was genuinely at risk, and worse, could not tell which of those an offer would actually move versus who would stay — or leave — regardless.

Approach

We trained a gradient-boosted churn model on usage, billing, and tenure features with strict time-based validation to avoid leakage, and used SHAP to surface the drivers behind each flagged account. On top of that we added uplift modeling — a two-model design that estimates the incremental effect of an offer — so retention budget targets the persuadable customers rather than the sure things or the lost causes.

Architecture

Results

• —Retention budget aimed at persuadable customers, not the whole base
• —Roughly 30% of predicted churners recovered by targeted offers in the pilot*
• —A top-driver explanation attached to every flagged account