Chairman Foster, Ranking Member Hill, and members of the task force, thank you for the opportunity to appear before you to discuss the use of artificial intelligence in financial services.
My name is Douglas Merrill. I’m the CEO of ZestFinance, which I founded ten years ago with the mission to make fair and transparent credit available to everyone. Lenders use our software to increase loan approval rates, lower defaults, and make their lending fairer. Before ZestFinance, I was Chief Information Officer at Google. I have a Ph.D. in Artificial Intelligence from Princeton University.
The use of artificial intelligence in the financial industry is growing in areas like credit decisioning, marketing, and fraud detection. Today I will discuss a type of AI — machine learning (a.k.a ML) — that discovers relationships between many variables in a dataset to make better predictions. Because ML-powered credit scores substantially outperform traditional credit scores, companies will increasingly use machine learning to make more accurate decisions. For example, customers using our ML underwriting tools to predict creditworthiness have seen a 10% approval rate increase for credit card applications, a 15% approval rate increase for auto loans, and a 51% increase in approval rates for personal loans — each with no increase in defaults.
Overall, this is good news and it should be encouraged. Machine learning increases access to credit especially for low-income and minority borrowers. Regulators understand these benefits and, in our experience, want to facilitate, not hinder, the use of ML.
At the same time, ML can raise serious risks for institutions and consumers. ML models are opaque and inherently biased. Thus, lenders put themselves, consumers, and the safety and soundness of our financial system at risk if they do not appropriately validate and monitor ML models.
Getting this mix right—enjoying ML’s benefits while employing responsible safeguards—is very difficult. Specifically, ML models have a “black box” problem; lenders know only that an ML algorithm made a decision, not why it made a decision.
Without understanding why a model made a decision, bad outcomes will occur. For example, a used-car lender we work with had two seemingly benign signals in their model. One signal was that higher mileage cars tend to yield higher risk loans. Another was that borrowers from a particular state were slightly less risky than those from other states. Neither of these signals raises redlining or other compliance concerns. However, our ML tools noted that, taken together, these signals predicted a borrower to be African-American and more likely to be denied. Without visibility into how seemingly fair signals interact in a model to hide bias, lenders will make decisions which tend to adversely affect minority borrowers.
There are purported to be a variety of methods for understanding how ML models make decisions. Most don’t actually work. As explained in our White Paper and recent essay on a technique called SHAP, both of which I’ve submitted for the record, many explainability techniques are inconsistent, inaccurate, computationally expensive, or fail to spot discriminatory outcomes. At ZestFinance, we’ve developed explainability methods that render ML models truly transparent. As a result, we can assess disparities in outcomes and create less- discriminatory models. This means we can identify approval rate gaps in protected classes such as race, national origin and gender and then minimize or eliminate those gaps. In this way, ZestFinance’s tools decrease disparate impacts across protected groups and ensure that the use of machine learning-based underwriting mitigates, rather than exacerbates, bias in lending.
Congress could regulate the entirety of ML in finance to avoid bad outcomes, but it need not do so. Regulators have the authority necessary to balance the risks and benefits of ML underwriting. In 2011, the Federal Reserve, OCC, and FDIC published guidance on effective model risk management.1 ML was not commonly in use in 2011, so the guidance does not directly address best practices in ML model development, validation and monitoring. We recently produced a short FAQ, which we’ve also submitted for the record, that suggests updates to bring the guidance into the ML era. Congress should encourage regulators to set high standards for ML model development, validation and monitoring.
We stand upon the brink of a new age of credit. An age that is fairer and more inclusive, enabled by new technology — machine learning. However, “brink” can also imply the edge of a cliff; without rigorous standards for understanding why models work, ML will surely drive us over the edge. Every day that we wait to responsibly implement ML keeps tens of millions of Americans out of the credit market or poorly treated by it. Thank you for your time and attention.