Can an algorithm discriminate against minorities?

Apparently, it can and does, according to research by Professor Adair Morse of the Haas

School of Business at the University of California at Berkeley and three other Berkeley

professors.

In their working paper, “Consumer Lending in the Era of FinTech,” Morse, and her co-authors

Richard Stanton and Nancy Wallace of Haas and Robert Bartlett of UC Berkeley School of Law,

discovered that even so-called fintech (short for “financial technology”) lenders, driven by computer

algorithms, charge as big a premium to African American and Hispanic applicants as face-to-face lenders,

who would presumably be more driven by their prejudices.

MORTGAGE LENDING IS RAPIDLY MOVING ONLINE

But the algo-driven fintech lenders are moving in the right direction, the researchers found, and show

no greater propensity than human mortgage bankers to reject minority borrowers’ loan applications.

Mortgage lending is rapidly moving online, led by firms like Rocket Mortgage, an entirely digital division

of Quicken Loans, the nation’s largest mortgage lender. Some 45% of the 2,098 large U.S. mortgage

lenders the researchers examined offered completely online or app-based mortgage contracting.

“Nearly all of the big banks and most small lenders now act as fintech,” Adair and her colleagues write.

The U.S. residential mortgage market is dominated by the government-sponsored entities (GSEs),

Federal National Mortgage Association (Fannie Mae) and Federal Home Loan Mortgage Corporation

(Freddie Mac). Fannie, Freddie, and another GSE, Government National Mortgage Association (Ginnie

Mae), invest in or insure more than 90% of all U.S. mortgages.

Fannie and Freddie won’t buy or guarantee mortgages that don’t meet their lending standards,

which usually raises the cost of such “nonconforming” loans to borrowers. Those uniform standards

remove much of the subjectivity from lending decisions, which is where discrimination tends to creep in.

Quantitative standards, of course, are especially amenable to fintech companies and their algorithm-

driven decision-making.

PROFS USED MACHINE-LEARNING TECHNIQUES TO STUDY 8.7 MILLION LOAN APPLICATIONS

Since the financial crisis, the researchers write, Fannie and Freddie have “fully [determined] the price of

credit risk by their role as guarantors. In particular, the GSEs produce a predetermined grid pricing that

prices credit risk” based on loan-to-value ratios and credit scores.

That “grid pricing” enabled Morse and her colleagues to quantify the extent of housing discrimination.

They used machine-learning techniques to create a dataset that for the first time linked income,

ethnicity, and loan-to-value ratio, as well as such specific factors as coupon and loan amount, for five

million accepted loan applications and 3.7 million rejected loan applications between 2007 and 2015.

They looked at only 30-year fixed-rate mortgages on single-family homes securitized by the GSEs, whose

median loan amount was just over $100,000, and they zeroed in on borrowers who had FICO scores

between 630 and 770.

MORTGAGE ‘DISCRIMINATION’ ON 11% TO 17% OF LENDERS’ AVERAGE PROFIT PER LOAN

After stripping out other factors like differences in risk among mortgages issued to different groups, the

researchers found “discrimination…of 5.6 to 8.8 bps for purchase mortgages, or approximately 11-17%

of lenders’ average profit per loan.” And there was virtually no difference between traditional and

fintech lenders. “Interest rate discrimination is almost identical for fintech lenders (5.3 basis points extra paid by minorities) as for the overall set of lenders (5.6 basis points),” they wrote.

Their conclusion: Fintechs do not discriminate less than face-to-face lenders. “It must be the case that any loan officer discrimination that is removed by not seeing faces is captured by algorithms that better predict which borrowers can be captured at higher rates,” they write, suggesting that algorithms identify loans to

minority borrowers as potentially more profitable and so charge them slightly higher rates.

But the aggregate cost is high. Since “each extra basis point in mortgage interest charged due to

discrimination costs minority mortgage holders approximately $100 million per year,” that means

minority borrowers pay $500 million a year more in interest than white borrowers at fintech and

traditional lenders.

The good news is that the discrimination seems to be declining, which, the authors argue, “could be

due to competition from the platforms and/or the ease of shopping-around made possible by online

applications….” And of course, they found fintech lenders don’t discriminate in their decisions to accept

or reject new mortgage applications.

Morse, associate professor of finance, began her career as an entrepreneur in Poland. She studies a

wide range of topics, from entrepreneurship to household finance to tax policy. She got her BA at

Colgate University, two masters’ degrees at Purdue and a Ph.D. in finance from the Ross School of

Business at the University of Michigan. She teaches classes in New Venture Funding and Impact

Investing, and before Haas taught at the University of Chicago Booth School of Business.

 

Poets&Quants’ Professor of the Week Series

Columbia’s Khandelwal: Trump’s Trade War Hurts the Ones He Loves