Stripe’s Radar system watches over every card swipe, evaluating more than 1,000 signals in under 100 milliseconds to decide whether to approve a payment. The model, now built on a Shield NeXt architecture, cuts training time by 85% and maintains a 99.9% accuracy across billions of legitimate transactions.

Earlier iterations combined XGBoost and a deep neural network in a Wide‑&‑Deep ensemble. Removing XGBoost would have dropped recall by 1.5%, so the team replicated its power within Shield NeXt, a multi‑branch approach inspired by ResNeXt that preserves memorization while enabling fast retraining.

Radar gains a network‑wide view: 90% of cards appear on multiple merchants, giving the model cross‑merchant patterns that individual stores miss. Automatic label ingestion from disputes eliminates manual pipelines, letting Stripe update fraud signals in real time.

Feature engineering remains a core practice. Engineers mine fraud reports, monitor dark‑web activity, and prototype new signals—such as IP‑to‑card ratios or time‑zone mismatches—measuring impact before deployment. This iterative loop keeps the model razor‑sharp against evolving fraud tactics.