Real-Time Fraud Risk Scoring

Leakage-safe fraud modeling across 590K transactions.

DATASET IEEE-CIS Fraud Detection Kaggle · 590,540 labeled transactions · transaction + identity tables

01

Validation design

TRAIN 413,378 · 70%

VALIDATE 88,581 · 15%

TEST 88,581 · 15%

PAST182 relative days → FUTURE

02

Model selection

Validation PR-AUC across model candidates

LightGBM feature-group ablation

03

Model performance

Model performance metrics

Final test confusion matrix

Top final-model features

Precision and confusion matrix use threshold 0.5.

Pseudo-entity ≠ verified customer TransactionDT = relative time Importance ≠ causality Batch latency ≠ API latency

Python
pandas
scikit-learn
LightGBM
CatBoost
PyArrow