The client needed to reduce false positives and enhance the efficiency of alert reviews by developing an alert risk scoring model based on demographic, behavioral, and scenario-based features.

Developed and optimized an ML-based alert risk scoring model.

1. Feature Generation: Derived features from a feature mart encompassing demographic, behavioral, and scenario-based attributes.

2. Model Development: Developed multiple models (NN, ensemble-based), benchmarked them using Gini, Recall, and ROC-AUC scores. Selected the best-performing model based on stability and performance metrics.

3. Model Calibration: Calibrated and binned model output scores to reflect low-risk and high-risk alerts. Initially reviewed low-risk alerts and subsequently suppressed them.

4. Implementation and Monitoring: Implemented the model with ongoing performance monitoring and data drift measures (KS divergence, data density). Created a standardized approach and interpretable model in line with SR 11-7 guidelines..

Reduced false positives and improved model governance.

- The model reduced false positives by 20-30%, standardized the approach in line with SR 11-7 guidelines, and provided a high-quality, well-governed model.

- The bank benefited from an optimized and efficient alert review process.