The practical implication of that design choice is significant. With over US $ 1.4tn in payments processed annually across 197 countries, Radar’ s machine learning models have access to a dataset few independent fraud tools can approach in breadth or recency.

Every transaction is assigned a risk score in real time, evaluated across hundreds of signals, device fingerprints, IP reputation, behavioural patterns, card network data and issuer responses before a decision is made to approve, flag or block.

There is a 92 % probability that any given card has been seen previously on the Stripe

network, a figure that materially improves the accuracy of those assessments.

Radar reduces fraud by an average of 38 %, according to Stripe’ s own data, while adaptive rules introduced in 2025 allow risk thresholds to incorporate issuer intelligence dynamically, reducing false positive rates and recovering revenue that stricter blanket rules would otherwise forfeit.

What distinguishes Radar’ s architecture is its iterative character. Models have evolved from logistic regression to deep neural networks as the training dataset has grown, with each architectural change producing measurable improvements in detection performance.

The system is also increasingly capable of protecting non-card payment methods, including ACH and SEPA transactions, relevant as businesses expand beyond traditional card rails and malicious actors adjust their targeting accordingly.

60 April 2026