By John San Filippo
Finopotamus was onsite at Money20/20 USA in Las Vegas. Co-founder John San Filippo interviewed more than two dozen technology experts on a wide range of topics. The results of these interviews are presented in this series called “The Voices of Money20/20.”
According to Mike de Vere, CEO of Zest AI, community financial institutions are faced with a trillion-dollar opportunity if legacy loan decisioning models are abandoned in favor of artificial-intelligence (AI) based systems.
“Credit decisions have been made the same way since the 1950s using old maps,” de Vere told Finopotamus. “What we now know is by leveraging more data and better math and machine learning (ML), you can render a better decision.” How much better? According to de Vere, credit unions can increase approvals by 15% while keeping the risk level constant. He further claimed that credit unions can decrease loan losses by 30% while holding approvals constant.
“What's really becoming prevalent today is the interest around inclusion and lending,” he said. “Our approach and our software actually solve for that.”
Inclusion Is Key
“With the old approach to lending, it takes something like 15 signals of data to decide if I should lend you money. With our new approach, we're able to look at hundreds of variables, maybe even a thousand variables,” he said. “Imagine now we have a machine learning decisioning model, and we put a new AI model up next to it. Are there some variables that are a proxy for race? If there are, I'm going to tune them down.”
The result, according to de Vere, is a near colorblind machine learning model. “It's a method that Freddie Mac, for example, is using today. Some of the biggest banks in the U.S. are also leveraging this technology,” he said, “but credit unions of all sizes care about it, too, because they want to get to more yeses for more of their members.”
A Coin Toss?
“Most national credit scores do a good job of decisioning on super-prime and subprime, but most of us are somewhere in the middle of this continuum,” said de Vere. “When you look at study after study and compare a machine learning model to an industry score, the national scores come out slightly better than coin toss.” The net result is that many worthy borrowers are left out.
“They’re denying loans that they shouldn’t and, on the flip side, they're giving loans to people that they shouldn't,” declared de Vere. “That's where that trillion-dollar opportunity is: The meat of America isn't being served by the current credit system.” He said that AI-based lending systems not only lead to more approvals, but also denials for riskier loans that legacy models may have let pass.
Starting From Scratch
“When we started, the technology was costly to build. We created a machine learning model that has eight ML models assembled together with 10 different data sources,” said de Vere. “Our first model took 14 months to build. It was immense and expensive. That would tailor to tier-one banks.”
Fortunately, the cost of the technology keeps falling. “Over the last 12 to 16 months, we've really been focusing on automation to be able to bring the price down such that it’s accessible for credit unions that might have $150 million in assets,” said de Vere. This means that mid-sized credit unions can compete with big banks like never before.
Fear of the Unknown
According to de Vere, some credit unions get stuck in the “that’s the way we’ve always done it” rut.
“I like to have my coffee exactly the same way every day and it's totally human nature,” he said. “I visited a credit union out in Texas recently. The chief lending officer said, ‘I've had the same scorecard in place since 1985.’ Think of the things that have changed in our economy in 25 years and they're decisioning off of the same scorecard!” He added that any scorecard created before the pandemic is probably out of date.
“All these credit unions and banks are out there decisioning off of models that were created before the pandemic,” noted de Vere. “This economy has totally changed and these FIs are still using the old approach. There’s certainly a fear of change.”
What’s the remedy? “I find a good dose of ‘show me’ usually works,” he said. “During our sales process, we actually show them a machine learning model that's built for them and their members. There's no guesswork.”
Custom-Tailored Models
Finopotamus then asked whether Zest AI aggregates data from across its entire customer base to build a more accurate model. Surprisingly, de Vere said that’s not necessarily the best approach.
“That's not where the real power comes in,” he said. “For example, when we were working with Hawaii USA Federal Credit Union, we tailored a model specifically for the Hawaiian Islands for personal loans. It's a stair-step difference from a decision perspective.”
Does this mean that a member could get one decision from one Zest AI credit union client and a different decision from another? Probably not, according to de Vere.
“There are two big credit unions where I live in Thousand Oaks (Calif.),” said de Vere. “If there's overlap, a lot of the signal is going to be the same, so the decision will end up being the same. But because they deal with a different profile, if you look at the credit tiers, one shifting a little more up, one shifting slightly down, there could be a slight difference in decisions.”
The Compliance Factor
In closing, de Vere told Finopotamus that credit unions can’t overlook compliance.
“Building a powerful model is one thing. Ensuring that it's compliant when the National Credit Union Administration (NCUA) examiner shows up is something else,” said de Vere. “The real art is understanding how data is used, ensuring that it's fully explainable. Ensuring that it's fully documented is something that most credit unions don't know to ask for.”