Learning Richer Credit Representations with Transformers at Affirm | by Affirm Technology | Sep, 2026 | Medium
Every checkout on Affirm presents a different underwriting question. For consumers who haven’t yet established a repayment history with us, their credit reports are an especially valuable source of information. Those reports may contain years of borrowing and repayment history with other lenders. Our models combine that information with previous Affirm history (if a user is an existing Affirm customer) and merchant context to assess the credit risk for a particular transaction. For years, we captured these signals through hand-engineered features and tree-based models such as XGBoost. This approach delivered substantial improvements across successive generations of underwriting models. But as those models matured, further gains increasingly depended on finding and developing new features. Each addition required domain expertise and substantial time backfilling new features for training new models. Credit reports contain considerably more detail than any fixed collection of summary feat