Stripe has developed a transformer-based payments foundation model, analogous to large language models, to move beyond the limitations of traditional machine learning models that rely on discrete features. By training on billions of transactions, this self-supervised network learns dense, general-purpose embeddings for each transaction, capturing complex relationships and subtle patterns. These rich embeddings have demonstrated significant success, such as improving real-time card-testing attack detection for large users from 59% to 97%. The model’s ability to distill transaction signals into versatile embeddings allows for better identification of nuanced, adversarial behavior and can be applied across various payment-related tasks like disputes and authorizations, suggesting that payments possess underlying semantic meaning and complex sequential dependencies.

@thegautam