February 26, 2021
Is credit good or bad? It depends, doesn't it? Public policy has always been schizophrenic about it. Decades of law and regulation have sought to make mortgage, consumer and small business credit more readily available to more people -- think of the fair lending laws, CRA and the government’s various loan guarantee programs. At the same time, too much credit hurts people, especially when it is priced highly to cover perceived risk, which, for tens of millions of people, it is. Concerns about easy, high-cost credit have brought us usury laws, and rules that limit payday lending and deceptive practices that push high-priced loans. We want the Goldilocks principle -- not too much, and not too little, but just right. Huge swaths of the population cannot get that healthy balance.
Many factors contribute to this problem, but possibly the most important -- and maybe the most fixable -- is that traditional risk scoring methods are not finely-tuned. As we’ve discussed in past shows, they work well for people with good traditional credit histories, but are limited in their ability to score people who don’t. This situation is a relic of the analog age, in which the information readily available to lenders was confined to a very narrow range. It’s a pre-digital problem. It’s about insufficient information.
My guests today have set out to change that. They are Mike de Vere, CEO of Zest AI, and Teddy Flo, the company's general counsel. Zest is seeking to revolutionize credit underwriting by leveraging more information and analyzing it through machine learning. They started out as a lender, but have pivoted to selling tools to banks and other lenders, to equip them to make better lending decisions.
In our conversation, Mike and Teddy lay out their critique of the status quo. They argue that today’s automated underwriting relies on “old math” that results in minority applicants being denied credit at far higher rates than the rest of the population. They describe the dramatic changes that occur when lenders move from analyzing 10 to 15 data points to more than 1,000.
Mike and Teddy also point out that the limitations of the old system hurt not only borrowers, but lenders too. Lenders do better when they can make more loans to people who can pay them back, and when they can price these loans attractively to compete for every creditworthy customer they can reach.
Today’s guests talk about what kinds of data they use. Significantly, they are very opposed to using behavioral and other factors that may correlate with risk but don’t logically cause it, as is advocated by other players in the AI credit space. They explain how machine learning techniques can assess risk factors in the complex interplay of data. They share some pretty eye-popping numbers on just how inaccurate some of the old models actually are, and argue that these should not be locked into use because they have been around forever and are considered “safe.” And they make a mathematical argument that if lenders would take a very small reduction in model predictiveness, they could reap massive gains in inclusion, without loss of credit quality.
They know, of course, that many people doubt their thesis -- are not ready to trust a “black box” algorithm that we humans cannot fully comprehend and evaluate. They talk about how they get past this objection, including how they prove that their model is more inclusive. And of course, they talk about how they address these same concerns held by regulators.They also have suggestions for clarifying regulatory policies where market uncertainty prevents adoption of new methods. They propose altering some rules that tend to lock in old math that is highly inaccurate and is highly exclusionary based on race. They would like to see regulators hold TechSprints that expose the industry to new techniques. And they urge that regulators proactively reward lenders that find less discriminatory models. US regulators have, of course, taken steps in that direction, including through a joint statement encouraging exploration of new models.
Crucially, we also talk about how to prevent the very real risk that AI will worsen, rather than reduce, racial and gender bias, either because it is trained on biased data or it teaches itself in ways that exacerbate these patterns. Zest uses “adversarial” de-biasing and other techniques to manage this danger. Mike and Teddy think that, contrary to popular assumptions, rigorous use of disparate impact fair lending standards will, over time, lead to mass adoption of these new techniques, simply because they reduce discriminatory effects.
To see research on using new kinds of data in underwriting, see the work of FinRegLab, which is looking rigorously at just these issues.
We know that technology is amoral and can do both good and harm. We know new credit underwriting could do a great deal of harm, if it is not designed rigorously and within clear, protective regulatory guidelines. Mike and Teddy are optimistic that this can be done, and that both lenders and borrowers will benefit.
More on Mike
Mike De Vere has 20 years of experience transforming organizations large and small, public and private, figuring out how to use technology to solve strategic business issues. Prior to Zest, Mike was managing director of Nielsen’s global insights business, converting it to a SaaS-based model. Before that, he was president and CEO of Harris Interactive and spent more than a decade in leadership roles at JD Power. Mike received his MBA from the University of Southern California.
More on Teddy
For more than a decade, Teddy Flo has served as a trusted legal advisor to consumer financial services companies, including international banks, credit card issuers, mortgage lenders, student loan servicers, and small-dollar lenders. Throughout his career, he has focused on providing practical legal advice that helps clients achieve their business objectives. He has litigated scores of cases, helped companies navigate government investigations, and advised his clients on key regulatory issues. As Zest’s General Counsel, he works closely with its legal, executive, engineering, data science, sales, and marketing teams to execute on Zest’s strategic decisions and ensure that the legal aspects of the company’s operations run smoothly, including compliance, litigation, contracting, and intellectual property. Teddy graduated summa cum laude from the University of Maryland with degrees in Finance and Economics and received his law degree with honors from The George Washington University Law School.
More for our Listeners
Pindar Wong of VeriFi will be my guest next week, in a show unlike anything we have done before — don’t miss it. We also have other great episodes in the queue including Rob Nichols of the American Bankers Association, and Sharmista Appaya and Ivo Jenik of the World Bank and CGAP, Paula Hunter and Lesly-Ann Vaughan of Mojaloop, as well as Douglas Arner of Hong Kong University.
2021 speaking events are well underway! I would love for you to join me at virtual South by Southwest with Cleve Mesidor in March — we will be talking about diversifying the tech world. For those of you in Asia — or who do well with time zone challenges — I’ll also be speaking in March at the Australian Regtech Association’s #ACCERATERegTech2021. Next week, I will be speaking at both the Chatham House Illicit Financial Flows conference and an event hosted by the National Policy Network of WOC in Blockchain. I’ll also be at the ACAMS International Conference on financial crime. And of course, I’ll be back this year at LendIt and Fintech South.
Last week, I was a guest on the BAI Podcast. I hope you have a chance to listen, as I discuss three financial regulatory predictions for 2021, and some ideas I have that, I hope, could help the financial realm and the country!
Stay informed by joining our mailing list