Affirm has introduced a new machine learning model based on transformer architecture to enhance its real-time loan underwriting process. The system, which leverages an attention mechanism similar to that used in models like ChatGPT, was unveiled at the company's Investor Forum following 18 months of development.
The new model outperformed Affirm's previous system during experimental testing. It evaluates borrowers in real time by analyzing diverse financial signals, including data from linked bank accounts, which the company began incorporating in January 2026 to improve approval rates while managing delinquency.
Founded in 2012, Affirm has developed a proprietary dataset covering over 50 million consumers and more than $100 billion in originated loans. Unlike many competitors, the company maintains its machine learning and credit engineering teams entirely in-house.
The transformer-based approach allows the company to assess complex interdependencies between transaction details, such as merchant data, purchase amounts, and recent borrower behavior. This shift aims to optimize unit economics by balancing higher approval rates with lower default risks.
