AI Modernizes Credit Scoring

Underwrite.ai applies advances in artificial intelligence derived from genomics and particle physics to provide lenders with non-linear, dynamic models of credit risk which radically outperform traditional approaches. In this webinar, Marc Stein, Founder and CEO of Underwrite.ai, provides an overview of the creation of Underwrite.ai and the specific credit lending needs that are being met with H2O Driverless AI.

Presenters: Marc Stein, Founder and CEO, Underwrite.ai

Vinod Iyengar, VP Product Marketing and AI Transformation, H2O.ai

‍

Why Machine Learning Is Essential for Solving Complex Credit and Health Problems

In this session, Mark Stein, Founder and CEO of Underwrite.ai, explains why machine learning is critical to solving some of the most persistent and difficult problems in modern society. Two of the clearest examples are access to credit, particularly in underserved communities, and cancer prediction through genomics. While these problems appear very different on the surface, they share a common challenge. Both are inherently nonlinear systems that are still being addressed with linear tools.

The Limits of Traditional Credit Underwriting

Modern consumer credit underwriting is built on a model developed in the late 1950s by Fair and Isaac, now known as FICO. The system predicts future credit behavior by analyzing past repayment history and distills that information into a single score based on five primary inputs. Lenders then layer simple rules on top of that score, typically focusing on debt to income ratio, late payments, inquiry count, and the FICO score itself.

This approach works reasonably well for borrowers with FICO scores above 700, commonly referred to as prime borrowers. For this group, risk is predictable and portfolios behave in relatively stable ways. The problem is that the model degrades rapidly as scores fall below that threshold. In subprime lending, predictive accuracy becomes unreliable, and for consumers without a FICO score at all, the model fails entirely.

Today, roughly 25 percent of Americans have no FICO score. These thin file consumers include recent immigrants, gig economy workers, people paid in cash, and young adults new to the workforce. This group is growing, not shrinking. As a result, about half of the population has access to reasonably priced credit, a quarter has access only to extremely expensive products like payday loans, and another quarter has almost no access to credit at all.

Linear Models Applied to Nonlinear Problems

Stein explains that the root of this issue is not policy but mathematics. Credit risk outside the prime population is a nonlinear problem. Linear regression and rule based scorecards work only in the stable portion of the credit spectrum. As soon as more variables interact in complex ways, those tools break down.

Machine learning changes this by allowing models to analyze thousands of variables simultaneously and capture interactions that linear methods cannot. Credit bureaus already provide thousands of raw attributes, but most lenders use only a small summary of that data. Underwrite.ai focuses on using the underlying information directly rather than relying on compressed scores.

Time based behavior is a key example. Inquiry count alone is not very informative, but the rate at which inquiries occur is highly predictive. Six inquiries over six months tells a very different story than four inquiries in a single week. Late payments also carry different meanings depending on when they occurred. Traditional credit reports do not model these temporal patterns effectively, but machine learning can.

Modeling Thin File Consumers

To address thin file consumers, Underwrite.ai analyzed more than 16,000 attributes available even when a FICO score does not exist. Using information theory, the team identified which attributes had predictive value and reduced the set to about 600 meaningful signals. These were then fed into automated machine learning pipelines that handled feature engineering, model selection, and ensemble creation.

The results were striking. Roughly 50 percent of thin file consumers performed at or better than prime borrowers, with default rates well below five percent. This meant lenders could apply the same risk standards to consumers without FICO scores as they did to those with traditional credit histories. In practical terms, it allowed lenders to safely extend credit to most of the thin file population while excluding only the highest risk segment.

Explainability and Regulatory Compliance

Explainability is not optional in consumer lending. Regulations in the United States, Canada, and the European Union require lenders to explain why a loan was approved or denied. This is not just a legal obligation, but an ethical one. If a consumer is declined, they deserve to understand what factors led to that decision and what actions they could take to improve their chances in the future.

For this reason, Underwrite.ai does not rely on opaque deep learning models for credit decisions. Instead, it uses explainable techniques such as gradient boosting ensembles. These models allow every individual decision to be traced back to contributing variables, making it possible to generate compliant adverse action notices automatically.

The system can return a lending decision along with a full explanation in less than half a millisecond. This makes it possible to replace slow, manual underwriting processes with fully automated workflows without sacrificing transparency or compliance.

Reducing Bias Through Better Modeling

Human underwriting introduces bias, often unintentionally. Variables like zip code or name can act as proxies for protected characteristics such as race, gender, or religion. Underwrite.ai explicitly excludes variables that could serve as proxies for protected classes and reduces all inputs to numeric signals that comply with fair lending laws.

By abstracting decisions away from human intuition and focusing on measurable behavior, machine learning models can reduce bias rather than amplify it. This leads to fairer outcomes and more consistent decision making across large populations.

Scaling Credit Access Globally

Beyond North America, Underwrite.ai focuses heavily on emerging markets where access to formal credit is extremely limited. In countries like the Philippines, the vast majority of the population is unbanked and has no credit history at all. By combining machine learning driven underwriting with mobile technology and alternative data, lenders can deploy capital efficiently in markets that were previously inaccessible.

This ability to reduce latency, increase throughput, and maintain accuracy allows lenders to scale dramatically without expanding underwriting teams. In one case, a lender increased daily loan volume by more than ten times without increasing staff, simply by automating decisions that previously took days.

The Bigger Impact

Stein closes with a broader message. Nonlinear machine learning is not just about improving accuracy or reducing costs. It is about solving problems that were previously considered unsolvable. When applied responsibly, these tools can expand access to credit, reduce systemic bias, and create real economic opportunity in markets around the world.

The technology already exists. The challenge now is using it thoughtfully and at scale.