Alternative underwriting software provider ZestFinance today launched an AI-based tool it claims will dramatically reduce bias in lending.

Founder and CEO Douglas Merrill, a former Google CIO, told Bank Innovation that ZAML Fair, an add-on to the company’s existing ZAML (Zest Automated Machine Learning) platform, is a new algorithm that lenders can use to tune their credit scoring models for maximum fairness. This is achieved by automatically reducing the impact of discriminatory credit data, he said.

Merrill said lenders have long removed bias from their models by tossing out offending credit signals, which can leave a lot of performance on the table.

According to Merrill, one large mortgage lender has reduced the racial disparity in its approval rates by 10% using his Los Angeles-based firm’s new algorithm. If that result was replicated across the 7.4 million mortgages originated each year in the U.S., he claimed that would put 172,000 more minority borrowers into homes. “And that is just by using the technology to tune up traditional credit models,” he added.

Merrill said the reality is every algorithm is biased, leaving questions about what’s driving that bias and what impact it’s having on outcomes.

“Reducing that bias with legacy tools is a brute-force exercise,” he said. “Lenders may spot a variable that drives disparity, but their only choice is to take it out and suffer the cost or leave it in and defend its legitimate use.”

For example, Merrill said, traditional credit scoring is highly dependent on income as a variable or as a component of other variables, like debt-to-income. Income, he said, is often a close proxy for race, but lenders can’t remove it or they’ll hurt the accuracy of the model.

“In the ML world, no one variable contributes that much to bias,” he said. “There are hundreds of them, instead of a couple of dozen.”

He said the ZAML Fair tool takes things one step further by spotting the offending variables that are driving the disparity and using a “helper AI” to create new models that tune down the influence of those variables on the outcome. He said ZestFinance customers typically see a 15% increase in approval rates or 30% reduction in charge-offs, depending on how they want to run the models.

One subprime auto lender, Prestige Financial Services, doubled its lending volume in part by deploying a ZestFinance model, Merrill said. The company had worked with Prestige to identify more than 2,700 unique borrower characteristics—more than 100 times the number they had traditionally used to underwrite loans.

“Within months, their loan volume doubled with no additional risk, just by increasing the signals,” Merrill said.

ZestFinance relies on standard credit bureau data, application data, existing customer data, and alternative data sources such as rent, telecom, and utility bills.

“The secret sauce isn’t in any one or two exotic variables—it has to do with using more sources of data and watching how they interact,” Merrill said.

He said while a conventional model may ask, “Have you had a bankruptcy? YES/NO,” an ML model could look at how long it’s been since the last bankruptcy and what kind of bankruptcy it was. For example, he said, someone who had a Chapter 7 bankruptcy six years ago would be a lower risk than someone who filed four months ago.

ZestFinance doesn’t sell a one-size-fits-all platform or scoring engine, Merrill said, adding customers can customize their models according to their own credit policies and data sources.

“Some customers want us to do more of the work,” he said. “Others want to collaborate on model development and testing. Still, others just want us to teach them how to use our tools so they can build their own models.”

Failure to use fair lending practices can cost banks money, too.

The U.S. Office of the Comptroller of the Currency announced today it fined Citigroup $25 million for violating the Fair Housing Act by denying some borrowers preferential rates on the basis of their race, color, national origin or sex.

Citigroup had “certain control weaknesses related to its Relationship Loan Pricing (RLP) program designed to provide eligible mortgage loan customers either a credit to closing costs or an interest rate reduction,” the OCC said. As a result, some borrowers did not receive the RLP benefit for which they were eligible, the agency found.