Acquiring new users costs far more than retaining existing ones, especially in digital banking where acquisition costs are high. This practical guide details a two-model ML system for smarter retention: a pre-churn model to flag at-risk users and an uplift model to pinpoint those actually responsive to retention offers. The author applied this to a debit card product, defining churn as no card payments for 30 days.

The pre-churn model uses gradient boosting (XGBoost, LightGBM, CatBoost) on transactional and profile features to predict payment probability over a 30-day horizon. Users with probability below threshold *t* enter the pre-churn segment. Validation relies on ROC-AUC for ranking quality and calibration curves with Hosmer-Lemeshow tests for probability accuracy.

Calibration drift over time is addressed by a lightweight logistic regression model atop the base predictor, transforming raw scores into reliable probabilities on fresh data. This modular design allows frequent recalibration without full model retraining, keeping the system stable and the retention budget efficient.