How Stripe Detects Fraudulent Transactions Within 100 ms

Stripe's Radar fraud detection system evaluates over 1,000 signals per transaction in under 100ms with 99.9% accuracy. The post covers the architectural evolution from a Wide & Deep ensemble (XGBoost + DNN) to a ResNeXt-inspired multi-branch DNN called Shield NeXt, which cut training time by 85%. Key topics include Stripe's network-scale data advantage (90% of cards seen across multiple merchants), embedding-based feature representations for geographic fraud transfer, the precision-recall tradeoff and how merchants can configure thresholds and custom rules, explainability tooling built around deep neural networks, and production challenges like real-time feature computation, per-merchant model validation, and counterfactual analysis for blocked transactions.