© 2019 SourceMedia. All rights reserved.

The banks warming up to AI-based lending

Though alternative lenders like LendingClub, Kabbage and Upstart, and auto lenders like Ford Motor Credit use artificial intelligence every day in loan decisions, most traditional financial institutions have stayed away, continuing to rely on factors like credit scores and debt-to-income ratios for underwriting.

But there are signs that is starting to change.

Freddie Mac is partnering with ZestFinance on a pilot of AI-based mortgage underwriting software. Some large, global banks, meanwhile, are using an Experian sandbox to build explainable AI models as part of their credit decisions, according to Greg Wright, executive vice president and chief product officer of the company's consumer information services.

"Machine learning as an application of AI is becoming more prevalent in lending because of its ability to automatically improve performance," Wright said. "This can result in increased customer acquisitions, reduced risk, and overall process efficiency."

And a few banks are actually piloting the use of AI platforms to make lending decisions.

Among them are $23 billion-asset First National Bank of Omaha, which wanted to try AI-based lending software to see if it could expand access to credit.

“It’s about inclusion, it’s getting more access to customers at better rates, better products and hopefully a better financial situation for the customer,” said Marc Butterfield, senior vice president of digital and payment solutions.

chart tried and true.jpeg

TCF National Bank, a Detroit-based bank with $47 billion in assets, had similar goals for its test of AI in lending.

“No. 1 was the process itself,” said Rita Carroll, executive vice president of consumer lending. “In traditional underwriting, you have to upload a bunch of pay stubs, provide all sorts of proof of income, proof of other things. What a cumbersome process.”

The bank has used Upstart's AI software, which it says is "completely digital and seamless."

"Our customers are looking at it positively," Carroll said.

TCF also hopes AI will give it a competitive advantage by allowing it to offer better rates to consumers.

“If we were to underwrite like everybody else, with a traditional FICO and debt-to-income grid, it would come to a very similar outcome, with a very similar approval rate, and that doesn’t seem to be the case” with the Upstart software, Carroll said. “Our mission for TCF is to allow access whenever possible at a really good rate.”

The bias question

As bankers begin experimenting with AI, an overarching concern shared by regulators, policymakers, consumer advocates and lenders is that bias could be built into a lending algorithm that’s looking at many different data sets. For instance, if it includes college graduation data, the software could biased toward people who graduate from college and neglect groups that statistically have lower college attainment rates.

Butterfield had to overcome internal worries that the AI software might be biased. In part he does so by noting the "models in existence today are not perfect."

“I asked our chief credit officer to prove to me that FICO isn't biased,” he said. “We need to be open to trying different things, especially if the data looks promising.”

Carroll sees it similarly.

“All models are biased in one way or another,” she said. TCF applies the same discipline of continuous monitoring and continuous testing that it applies to in-house models.

Jackson Mueller, associate director of the Center for Financial Markets at the Milken Institute, can point to multiple studies that find benefits in AI, including a report in May that found algorithms used by fintech lenders discriminate 40% less than face-to-face lenders. Milken supported recent studies conducted by FinRegLab that found merit in the use of cash-flow data, a factor often used by algorithms but left out of traditional lending underwriting.

But Mueller also acknowledges that it's a controversial area, and each new study can have an outsize impact.

"It only takes one bad report to come out to overshadow all the positive ones that have been introduced,” he said. “We’ve seen that before.”

The regulators' concern

Many banks appear unwilling to test AI because they fear what regulators will think.

The banks moving forward with the technology agree it's important to keep examiners informed.

TCF has quarterly meetings with examiners from the Office of the Comptroller of the Currency in which it keeps them up to date on the AI lending project. It also communicates with the OCC’s Office of Innovation.

"In all sincerity they're trying to catch up to understand what exactly is going on,” Carroll said.

She’s also waiting for regulators to update model validation policies that they released in 2011.

“The language and the policies no longer apply,” she said.

Butterfield said First National Bank of Omaha has made a deliberate culture shift to be more proactive with regulators.

“We reached out to the OCC and wanted to have a meeting with them and Upstart, to tell them what we're doing,” he said. “We're having conversations with the examiners and the Office of Innovation.”

Both TCF and First National embed their normal credit policies, which include traditional credit metrics like FICO score and debt-to-income ratio, in the AI software. (The software can also ingest many other kinds of data, including education and employment information.)

One reason they’ve kept FICO scores in the new models is that they have to report borrowers' average FICO scores to regulators.

Jeff Keltner, general manager at Upstart Network, said the online lender tends to keep FICO scores in its models because it doesn’t want to throw out any data element that might have some value.

He also pointed out AI-based lending models’ fairness can be monitored and measured by looking at their results.

“You go, what’s the outcome of this? Was this decision fair? Are we treating people equitably? And you ask the same questions no matter what the underlying model is,” he said.

Another issue regulators bring up around the use of AI in lending is that lenders can't use a "black box" to make decisions. Decisions have to be clearly explained. Every AI underwriting software vendor says that it has built explainability into its platform, and that its systems can explain decisions more clearly than humans can.

Results

At TCF, the new, Upstart-enabled loan is a brand-new product. The bank has made thousands of loans so far.

“The customers love it,” Carroll said. “The branch employees love it.”

One challenge was deploying the loan hybrid: online and in the branches.

“A lot of our branch employees are used to being that helper to the customer and nothing was taken away from them,” Carroll said. “So it's a lot of change management. But overall you couldn't have a better experience.”

First National Bank is just completing its pilot, which it began in July. It has lent tens of millions of dollars so far, he said.

The bank’s net promoter score among these borrowers has been “off the charts,” Butterfield said.

Carroll said that at her bank, even when customers are declined, they still have positive comments about the experience.

“We're very comfortable with testing other potential pilots within the bank because the model tested out fine,” Butterfield said. “We're figuring out what's our overall risk appetite for Upstart-type loans, I call them.”

For reprint and licensing requests for this article, click here.