Using AI for Credit Risk Analysis

Marc Stein shares his thoughts on developing novel underwriting techniques for lenders. Marc has applied advances in artificial intelligence derived from genomics and particle physics to provide lenders with non-linear, dynamic models of credit risk that radically outperform traditional approaches. He is a longtime entrepreneur and startup CTO, and is currently the CTO of Tua Financial and the CEO of underwrite.ai.

‍

Applying Nonlinear AI to Finance and Risk

In this talk, Mark Stein shares his perspective on how artificial intelligence has evolved and how nonlinear machine learning models can transform financial decision making. With experience in AI dating back to 1980, Stein has worked across speech recognition, natural language processing, expert systems, and multiple generations of neural networks, leading to today’s deep learning approaches.

Much of his career has focused on measuring risk in finance. That work eventually led him to cofound Underwrite.ai, where the goal is to apply advanced machine learning models to credit decisioning in a way that is fast, explainable, and compliant with regulatory requirements.

From Cancer Genomics to Credit Risk

Before founding Underwrite.ai, Stein explored a challenging problem in medicine: prostate cancer diagnosis. Human pathologists typically achieve only about 70 percent accuracy, making it one of the least reliable cancer diagnoses. By analyzing gene expression data across tens of thousands of genes using machine learning, Stein found it was possible to predict cancer presence with accuracy as high as 98 percent, even using samples of healthy tissue.

While the medical results were promising, the regulatory and financial barriers to commercialization were significant. That led Stein to ask a different question. Could the same nonlinear modeling techniques be applied to finance?

The answer turned out to be yes.

Why Traditional Credit Models Fall Short

Most consumer credit models rely on a small number of inputs, typically FICO score, debt to income ratio, trade line history, and inquiry count. These linear models work reasonably well for borrowers with FICO scores above 700, which is why banks tend to focus on that population.

The problem is that credit, like weather, is a nonlinear system. Linear models only work in the stable portion of the distribution. As soon as you move away from prime borrowers, prediction accuracy drops sharply. This is where nonlinear models become essential.

Underwrite.ai replaces traditional logistic regression with machine learning models that analyze thousands of raw credit bureau attributes rather than a handful of summarized scores. The value is not in the credit score itself, but in the underlying data.

Real World Results in Subprime Lending

Stein shares a case study involving a deep subprime lender in the United States. The lender issued short term, high interest loans and suffered a first payment default rate of over 38 percent. By replacing their traditional model with a machine learning driven approach that evaluated more than 2,000 attributes, the default rate dropped to 22 percent in the first month.

As the model continued to retrain dynamically using new payment data, default rates fell further, eventually reaching single digits. Just as importantly, the lender was able to identify the optimal balance between risk reduction and profitability. Over optimizing for low defaults reduced yield, so the model was tuned to lock in the most profitable risk level and run consistently over time.

Expanding Access to Credit

One of the most important insights Stein highlights is the difference between high risk and unknown risk. Many people lack credit scores not because they are unreliable, but because they are invisible to traditional systems.

By incorporating alternative data such as utility payments, rental history, bank account behavior, and cash flow patterns, lenders can accurately assess borrowers who would otherwise be excluded. In one audit, unscored borrowers actually outperformed certain high priced loan segments, demonstrating that lack of a credit score does not equal poor creditworthiness.

Automation, Explainability, and Compliance

Underwrite.ai focuses heavily on explainability, which is critical in regulated markets. While deep learning models can be powerful, they are often not legally acceptable for credit decisioning due to their lack of transparency. Instead, Stein emphasizes gradient boosting ensembles, which strike a balance between performance and interpretability.

This approach allows lenders to fully explain why a decision was made, generate compliant adverse action notices, and automate processes that previously required human underwriters. In some cases, loan decision timelines have been reduced from days to fractions of a second, enabling lenders to scale without increasing staff.

The Bigger Picture

Across finance, identity verification, fraud prevention, and even genomics, the pattern is the same. Complex problems cannot be solved effectively with simplistic models. Nonlinear machine learning allows organizations to work with reality as it is, not as simplified abstractions.

Stein closes with a broader message. As these tools become more accessible and automated, more people can apply them to meaningful problems. When used responsibly, AI has the potential not just to improve business outcomes, but to expand access, reduce bias, and create real societal value.

‍