Since launch, we’ve continuously invested in our suite of fraud prevention tools, and today, we’re excited to launch the result of those efforts.

The next generation of machine learning

As of today, we’ve rebuilt almost every component of our fraud detection stack to dramatically improve performance. In early testing, the upgraded machine learning models helped reduce fraud by over 25% compared to previous models, without increasing the false positive rate.

• Hundreds of new signals for improved accuracy: We’ve added new signals to better distinguish fraudsters from legitimate customers, including certain data from buyer patterns that are highly predictive of fraud. Some signals now use new, high-throughput data infrastructure to process hundreds of billions of historical events. (Even if a card is new to your business, there’s an 89% chance it’s been seen before on the Stripe network.)
• Nightly model training: Fraud evolves and changes rapidly. Radar can now adapt even faster by training and evaluating new machine learning models daily.
• Algorithmic changes for better recall and precision: We’ve optimized our machine learning algorithms in hundreds of ways—from boosting the performance of our decision trees to tweaking the minutiae of how we handle class imbalance, missing values, and more.
• Custom models for your business to maximize performance: Radar is constantly evaluating how to balance patterns from across the Stripe network with patterns that are unique to your business. Radar now trains and evaluates multiple models daily and determines which one achieves the best performance for you.