Team: Larp it Reference: 07408B3D

Alert Escalation Risk Model

Predicts the probability that a financial monitoring alert gets escalated by a human reviewer, based on the customer's transaction history. Gradient-boosted trees (LightGBM) trained on 60 engineered behavioral features.

0.594
5-Fold CV ROC-AUC
Out-of-fold validation score computed across 5 folds. Hover to inspect curve below.
14,000
Training Alerts
Alerts with labeled ground truth human outcomes.
6,000
Test Alerts Scored
Unlabeled test partition evaluated during elimination task.
17.2%
Historically Escalated
Empirical positive class base rate in the training sample.
60
Engineered Features
Behavioral AML attributes aggregated over card, cash, and nocturnal transfers.

Model Performance & Diagnostics

Out-of-fold predictions across 5 validation folds · Hover any visual element for exact readings
ROC Curve AUC = 0.594
Hover over the ROC curve coordinate area to track FPR, TPR, and Youden's index.
Probability Distribution OOF Sample (N=200)
Hover any histogram bin to view split counts (Escalated vs Dismissed) and empirical risk.

What The Model Learned Matters Most

Top 15 engineered features ranked by LightGBM split gain · Hover row for exact metrics
SHOWING ALL 15 RECORDED TOP FEATURES (SORTED DESCENDING BY IMPORTANCE)

Live Examples From Hidden Test Set

Highest vs. lowest predicted escalation risk among the 6,000 scored alerts · Hover row for full precision float & baseline delta
Signal ID Predicted Probability Risk Tier Baseline Comparison
Signal ID Predicted Probability Risk Tier Baseline Comparison