When a perfect score is a warning

The financial project initially produced a perfect churn score. Its documented audit traced that score to features that encoded the outcome: data leakage, not extraordinary predictive ability.

Removing the leaking features and rebuilding with Gradient Boosting produced a more credible 0.756 ROC AUC. The useful story is the diagnostic process and honest evaluation, rather than the original headline.

This is synthetic SaaS data with a planted churn signal. The resulting risk rankings demonstrate an analytical workflow; they do not establish revenue recovered by a real business.