Affirm wants its data to tell a more complete customer story

Affirm is developing a new transformer-based underwriting model to extract more complete customer stories from existing data for its lending business. This model seeks to identify patterns in the timing and sequence of credit events that conventional models might overlook. The deployment of this model on U.S. checkouts enabled approvals for applicants previously declined by the older system. This change resulted in 3.4% more completed purchases than a control group at comparable risk levels and outperformed similar expansion efforts using previous machine-learning models. Furthermore, consumer engagement with Affirm has increased, with active consumers reaching 27.8 million and transactions per active consumer rising from 5.8 to 7.0 in fiscal 2026. The focus on the quality of the underwriting engine is driven by the increasing frequency and smaller purchase sizes consumers make with Affirm products.

The shift toward time-dimensional analysis in credit history moves the focus from compressing complex financial data into a single score to understanding the trajectory of financial behavior. Traditional models rely on compression, which risks obscuring crucial context—specifically, whether a customer is recovering from past difficulties or actively declining. Affirm's innovation targets this gap by valuing the sequence and timing of events, suggesting that temporal sequencing forms a more predictive signal than aggregate historical metrics alone. The successful outcome in increased completed purchases suggests that embedding this temporal view into lending decisions can yield tangible commercial advantages beyond simple risk assessment. This implies a potential revaluation of consumer data where the *narrative* of financial interaction is as valuable as the static score, forcing competitors to also develop richer temporal models if they wish to gain a competitive edge in evolving consumer financing landscapes. What assumptions about the stability or predictive power of sequence-based modeling are being made, and what are the implications for fairness when these nuanced patterns begin to dictate access to credit?