It’s not just bank loan officers with racial biases who discriminate against black and Latino borrowers. Computer algorithms do, too.
That is the groundbreaking conclusion of University of California at Berkeley researchers who found that algorithmic credit scoring using big data is no better than humans at evening the playing field when it comes to determining home mortgage interest rates.
Both online and human lenders earn 11 to 17 percent higher profits off minority borrowers by charging African Americans and Latinos steeper rates, the study said. Black and Latino consumers pay 5.6 to 8.6 basis points higher interest on home purchase loans than their white or Asian counterparts with similar credit profiles — no matter whether they obtained their loans through a face-to-face process or online. The effect is smaller when it comes to refinancing, with black and Latino borrowers paying 3 basis points more.
The disparity results in African Americans and Latinos, together, paying up to a half a billion dollars more in mortgage interest each year, the study found.
“The move away from humans should remove malice forms of discrimination,” said Adair Morse, a finance professor at Berkeley’s Haas School of Business who co-authored the paper. “But we’re moving to an era where we’re using variables to statistically discriminate against people in lending.”
The findings are significant as more consumers shop for mortgages online. Nearly half of the 2,000 largest mortgage lenders offer complete online mortgage applications.
Morse and her colleagues — Nancy Wallace and Richard Stanton at Haas and Robert Bartlett at Berkeley Law — focused on 30-year, fixed-rate, single-family home loans issued between 2008 and 2015. They were able to link data on interest rates, loan terms, property location, income and credit scores with borrowers’ race for the first time. All the loans were guaranteed by the government-sponsored enterprises Fannie Mae and Freddie Mac, allowing researchers to remove credit risk as a factor in pricing differences.
“Even controlling for credit worthiness, we see discriminatory effects in the rates at which borrowers obtain mortgages,” Bartlett said.
Researchers said the racial disparities could result from algorithms that use machine learning and big data to charge higher interest rates to borrowers who may be less likely to shop around. For example, the algorithms may take into account a borrower’s neighborhood — noting who lives in banking deserts — or other characteristics such as their high school or college. The consumers least likely to comparison shop also happen to be black or Latino.
It’s legal to use statistical data to set prices that help maximize profits — in theory. The problem arises when the data correlates with race, independent of credit risk. Discriminating against minority borrowers — even unintentionally — is illegal unless it’s based on their creditworthiness, Bartlett said.
Homeownership and debt are key factors in racial wealth disparities.
Bartlett said banks that are increasingly using big data in determining approvals or lending rates should be subject to audit to ensure that their methods do not discriminate against minority borrowers who have the same credit scores as whites.
The researchers outlined a couple of silver linings in their study. Increased competition among lenders has resulted in less discrimination overall. And when it comes to determining whether to accept or reject a loan, online lenders do not discriminate against minorities — whereas their human counterparts are 4 percent more likely to reject Latino and African American borrowers.
If anything, the online lenders end up catering to those discriminated against by face-to-face lenders, the study found.
“Rejecting loans would be money left on the table for lenders,” Morse said.