EXCLUSIVE: Affirm Rebuilds Underwriting to Approve Borrowers a Credit Score Can’t See

Affirm, a fintech company, has introduced a new real-time underwriting model at U.S. checkouts that approves borrowers it previously declined, including those with limited credit histories or no FICO score. This transformer-based model, now live, builds upon 14 years of Affirm's individualized, real-time underwriting with machine learning models.

The new model approves more eligible applications, resulting in 3.4% more completed purchases in its initial deployment, with those loans performing better than a control group under Affirm's previous models.

Credit scores traditionally compress a consumer's record into a single number, but the new model identifies patterns in credit accounts and how they change over time, offering better insights from the same data. Affirm's President, Libor Michalek, emphasized that the goal isn't to approve every transaction, but to make the right decision for each one, maintaining the same disciplined approach to underwriting.

Affirm has also built a proprietary algorithm for the model that ensures the same level of explainability as traditional machine learning models while maintaining speed for checkout use. The model is being well received, as customers are using Affirm for smaller, more frequent purchases, processing 53 million transactions in the latest quarter, a 41% year-over-year increase.