Affirm launches transformer model to expand real-time underwriting
- Affirm launches a transformer-based ML model for real-time US checkout underwriting
- Initial deployment yielded 3.4% more completed purchases vs a control group
- New loans performed better than comparable expansions under previous models
- System targets eligible consumers with limited credit histories or no FICO scores
- Proprietary algorithm ensures explainability and speed for real-time decisions

*this image is generated using AI for illustrative purposes only.
Affirm (NASDAQ: AFRM) has deployed a new transformer-based machine learning model for real-time underwriting at checkout in the U.S. The system leverages 14 years of transaction-level data to approve additional eligible consumers, including those with limited credit histories or no FICO scores.
The company reported that the initial deployment produced 3.4% more completed purchases compared to a control group. Affirm stated that these additional loans performed better than a comparable expansion under its previous machine learning models.
Model Architecture and Performance
The new model identifies patterns within and across credit accounts, analyzing how they change over time without requiring separate measures designed for each pattern in advance. This approach allows the system to find more signal in existing data beyond traditional credit-bureau measures such as balances, utilization, and payment history.
Libor Michalek, President of Affirm, noted that the architecture enables the company to find new information within data it already possesses. "Seeing a credit history more clearly means we can responsibly say yes to more people," Michalek said.
Real-Time Decisioning
Affirm built a proprietary algorithm to ensure the transformer model maintains the same level of explainability as traditional machine learning models while remaining fast enough for real-time use. The company emphasized that underwriting is central to its business model, which promises individual purchase decisions based on what a person can responsibly repay that day.
"The goal isn’t to approve every transaction, it’s to make the right decision for each one," Michalek added. "We don’t benefit from extending credit that can’t be repaid, which means saying yes to more people only works when we get even better at saying no."
What the Numbers Show
The disclosed performance metrics highlight a divergence between volume expansion and risk management. While the model drove a 3.4% increase in completed purchases by approving applications previously declined, the company asserts that loan performance improved relative to prior expansion attempts. This suggests the transformer architecture may identify creditworthiness signals in thin-file consumers that previous models missed, allowing for growth without a corresponding rise in default rates.
How might Affirm's improved underwriting accuracy for thin-file consumers impact its market share against traditional credit card issuers and other BNPL competitors?
What are the potential regulatory risks associated with using transformer-based AI models for credit decisions, particularly regarding algorithmic bias and explainability requirements?
Could the 3.4% increase in completed purchases signal a broader shift in consumer spending behavior among previously excluded demographics, and how might this affect long-term default rates?

































