How Stripe Detects Fraudulent Transactions Within 100 ms

Stripe's Radar fraud detection system evolved from Wide & Deep architecture (XGBoost + DNN) to Shield NeXt, a multi-branch neural design inspired by ResNeXt, reducing training time 85% while preserving a critical 1.5% recall advantage. The system processes 1,000+ signals per transaction in <100ms with 99.9% accuracy across billions of transactions. Competitive advantages include network effects (90% card reuse across merchants), automatic ground-truth labeling via payment disputes, and 100+ features extracted through fraud forensics and dark web monitoring.