Affirm turns to transformer models to expand lending

Affirm is incorporating transformer-based machine learning into its underwriting models to better predict consumers’ ability and willingness to repay. The lender hopes the technology, which also underpins large language models such as ChatGPT, Claude and Gemini, will help it approve loans it previously would not have funded.

President Libor Michalek said the approach analyzes sequences of credit, purchase and repayment behavior rather than relying on signals individually engineered for the model. Engineers can direct the system to examine historical behavior without specifying which patterns to find. That could help distinguish a temporary missed payment from evidence of a borrower’s deteriorating financial position.

Other fintech companies are pursuing similar methods. Stripe-owned Parafin uses transformer-based underwriting to assess small-business credit through cash flow, while Plaid also has a transformer-based underwriting model. Their adoption reflects broader interest in applying AI sequence prediction to lending decisions.

Aaron McPherson, principal at AFM Consulting, cautioned that opaque models can make predictions difficult to explain. He said using language models to identify characteristics that can then be assessed through more transparent techniques offers a way around that limitation. He also argued that consistent approaches across lenders could improve accuracy, while competitors could readily duplicate Affirm’s approach.