Innovation in Lending

Lendit 2021: Ensuring Fairness In Lending With AI

Bruce Upbin

April 29, 2021

Our CEO Mike de Vere took part in a lively panel at Lendit Fintech this week called “Ensuring Fairness In Lending With AI.” Here’s a recap.

It’s pretty clear to most that AI has the potential to offer improved efficiency, enhanced performance, and cost reduction for financial institutions, as well as benefits to consumers and businesses. Yet, there’s still a lack of clarity around how algorithmic credit underwriting fits within the regulatory boundaries of existing fair lending laws or what kinds of regulations we need to ensure their safe and fair use.

The rising use of AI for high-stakes credit decisions brings more scrutiny from the Biden Administration and newly emboldened state banking regulators. A new federal interagency call for comment in late March signaled that regulators are putting AI higher on their agenda for regulatory reform. The panel addressed questions such as: How can all players better understand the appropriate governance and controls over AI? What are the top challenges in developing, adopting, and managing AI? And what kind of clarification would be helpful from Washington?

Panelists included Mike de Vere, CEO of Zest AI, Angela Ceresnie, CEO of Climb Credit, an innovative payments and finance platform for educational programs offering alternatives to four-year colleges; Melissa Koide, CEO of FinRegLab, a DC-based non-partisan non-profit research organization evaluating data and technology uses in financial services, Katie Neal, Advocacy & Outreach at Plaid, which provides the kinds of alternative credit data used by AI models to widen access to credit, and the moderator Kali Bracey, a fair lending expert, and partner at Jenner & Block.

Melissa Koide kicked off the conversation by calling for a more precise definition of the kind of AI we’re talking about in financial services and, especially, consumer lending. AI is a broad and nebulous category, and it manifests in so many different ways. “We could call my iPhone an AI,” said Koide. So, we shouldn’t get caught up in the conversations about regulation AI writ large when what we’re talking about is ensuring that complex machine learning algorithms used in underwriting are safe and fair to consumers and don’t exacerbate the inequalities that we are coming to terms with today.”

Credit access is a problem that affects millions of Americans through artificially suppressed scores due to generations of bias encoded in the data used to develop traditional credit scores. Angela Ceresnie talked about the new Zest-built credit algorithm that Climb will be using “for both approvals and pricing as a way to shift our reliance away from on FICO and towards a model built completely on credit-bureau data and factors.” The new model, she said, “will make smarter underwriting decisions and approve more people while continuing to reduce disparate impact.”

"The innovation we're doing here leverages credit bureau data to assess credit more fairly. It's a nice step in the direction of moving away from what we know is a currently broken system.”

Said Ceresnie, “We all know that FICO scores are inherently flawed, both in terms of its ability to predict performance, but also in the inherent bias that exists in the way it has been built over the years. The innovation we're doing here leverages credit bureau data to assess credit more fairly. It's a nice step in the direction of moving away from what we know is a currently broken system.”

Katie Neal from Plaid chimed in, noting that several studies are showing that the inclusion of rent or utility payments can significantly increase scores of traditionally subprime borrowers. Goldman Sachs did a study a few years ago and found that when consumers were allowed to report their rental data, they increased their credit scores by 40 points. And so it's data that can be used in a more predictive way and built into machine learning models to essentially allow consumers better access to credit that traditionally have sat outside of traditional credit score models.”

Koide pointed to the landmark research study that FinRegLab published in 2019 on the impact of cash flow data on credit access. If you haven’t seen it, the tl;dr is that yes, cash flow data is predictive of creditworthiness and allows lenders to extend credit to populations who otherwise might have been turned down. (She cites Petal as an example of a fintech credit card lender growing fast by using cash flow underwriting to serve hard-to-score borrowers. Seventy percent of its cardholders had no score before starting and could raise their score to an average of 678 within two months.) It also found that cash flow data does not cause or exacerbate disparate impact risk. The transaction data provides an independent risk assessment on top of what the traditional metrics offered. Koide added: “I know there are a lot of concerns out there about the bureaus and the scores that exist. I think we also need to pause and realize the U.S. has a pretty strong and remarkable credit system that has enabled millions across this country to access credit. On the other hand, there’s value in augmenting the data, and the research that we've done helps to show that.”

“Something that we think a lot about at Plaid,” said Neal, “is how we can modernize the regulatory system to account for these new technologies that regulators themselves have said are more predictive. That's going to be the crux of the next four years, especially when we think about improving access to credit to the 35 million credit invisibles.”

Kali Bracey then asked for examples of ML benefiting consumers today. “Have we seen this expansion of credit?” she asked. Zest’s de Vere said yes, every day. “When we look across all our clients, we see a roughly 20% increase in approvals and significant improvements in approvals for protected classes as well... There's a credit union that we're working with right now where they saw a 25% improvement in approvals for women. It's having a material impact when done carefully and right.”

The discussion then turned to how we make this shift happen. If ML underwriting and alternative data do achieve fairer outcomes for millions of consumers, then how, asked Ceresnie, “do we take the next step so that people and organizations can feel confident in using those innovations to do the right thing?”

“We've spent decades trying to take a step ahead, but the old math and old approaches are taking us around and around in circles, and we’re not making any real progress.

We need more empirical studies, said Koide. “The data that Petal has released is the exact type we need to foster innovation in this space and provide lenders and regulators with more certainty.”

“There is some skepticism around the extent to which this technology is penetrating, candidly,” she continued. “We have anecdotes, but there is more work to do to understand the kinds of credit products extended to underserved populations. Are the repayment activities being reported to the bureaus having credit building benefits for consumers? Are we seeing safe and affordable products extended to small businesses and the millions of sole proprietors who are thirsty for survival and getting the credit they need to do it?”

All the panelists agreed on the need for more certainty from regulators to build lender confidence in using alternative data and new analytics. Zest’s de Vere pointed out that some lenders have already gotten comfortable with the latest tools and tech. “We’re seeing ML models being put into production every week with our clients,” he said, citing lenders such as Freddie Mac that are actively using a Zest-invented approach called “adversarial de-biasing.” This technique automates the search for less discriminatory alternative models and gives lenders real choices around fairness for the first time. “We've spent decades trying to take a step ahead, but the old math and old approaches are taking us around and around in circles, and we’re not making any real progress. I'm very bullish, obviously, about AI, but I care about where we're at in the U.S., and we have an opportunity to leap forward in fairness,” said de Vere.

Thank you for subscribing!
Something went wrong while submitting the form.