The Racial Landscape of Fintech Mortgage Lending

ABSTRACT

Little is known about racial patterns in fintech mortgage lending, despite evidence of racial disparities in the broader mortgage market. This study leverages 2015–2017 Home Mortgage Disclosure Act data to assess disparities in lending outcomes between White and non-White applicants and between neighborhoods of varying racial composition in the United States’ 200 largest metropolitan areas at fintech and traditional lenders. Results of a series of binary logistic regression models suggest disparities in rates of loan approvals between White and similarly qualified non-White applicants are substantively small overall, but lower at fintech lenders relative to traditional lenders, most substantially for Latinos. Non-White applicants are more likely to receive subprime terms relative to similarly qualified White applicants at both lender types, and disparities in rates of subprime loan receipt between Black and similarly qualified White applicants are greater at fintech lenders than traditional lenders. Neighborhood racial composition has a mixed but substantively small impact on approval rates at both lender types. However, both lender types distribute subprime credit to non-White neighborhoods at significantly higher rates than to White neighborhoods. Findings suggest fintech lending contributes to racial and spatial disparities in subprime mortgage lending and warrants increased scrutiny from regulators.

KEYWORDS:

• fintech lending
• mortgages
• subprime lending
• racial disparities
• neighborhoods

Acknowledgments

I thank Lance Freeman, Malo Hutson, Magda Maaoui, Michael Snidal, and Gayatri Kawlra for their helpful comments on earlier drafts of this article. I also thank three anonymous reviewers for their insightful feedback and suggestions.

Disclosure Statement

No potential conflict of interest was reported by the author.

Notes

1. An autonomous machine learning algorithm, in contrast to a rules-based algorithm, has the capability to develop its own pattern of statistical analysis leading to a prediction. In the words of Mittelstadt et al. (2016, p.10), “Machine learning is defined by the capacity to define or modify decision-making rules autonomously.”

2. These percentage increase calculations are derived by comparing the predicted margins for fintech and traditional lenders found in Appendix Tables A1 and A2 regarding approvals and subprime lending across racial groups of borrowers, and in Appendix Tables A3 and A4 regarding lending across neighborhoods of varying racial composition. I share the percentage increase rather than the absolute increase in percentage points. For example, a hypothetical change from 10% to 11% would be presented as a 10% increase (the percentage increase) rather than a 1% increase (the absolute increase).

3. The HMDA was enacted in 1975 to assist in the enforcement of 1968’s Fair Housing Act. It requires that nearly all mortgage lenders submit their loan-level application data to the federal government. These data are available to the general public and facilitate assessments of patterns of discriminatory behavior among lenders.

4. A comparison of using a 1.5% threshold above the prime rate (as the new reporting rules require) and a 3% threshold above the Federal Treasury rate (as the old reporting rules required) suggests that, during my study period, the two methods would have produced a similar overall threshold for high-cost lending. However, to test whether a higher threshold would substantively alter the results, I conducted a sensitivity analysis with a threshold of 3% above the prime rate as my definition for high-cost loans. The results are directionally similar to those found in Table 3, with the most noteworthy difference being a more substantive disparity between White and similarly qualified Latino borrowers.

5. The 50% sample was drawn to avoid circularity issues. These out-of-sample data were used to estimate a logistic regression model predicting poor credit history, and the parameters of this model were, in turn, used to obtain predictors for poor credit in the remaining 50% of data. Of the 50% of data randomly sampled for the purposes of estimating this prediction, all observations (N = 1,506,374) that included a reported reason for denial or were approved were utilized in the logistic regression. The dependent variable was a dummy variable equal to 1 if an applicant was denied because of their credit history and equal to 0 otherwise. Independent variables included the loan amount requested, an applicant’s income, race, and sex, the presence of a coapplicant, whether a preapproval was requested, the loan-to-income ratio, a dummy variable indicating whether a monthly loan payment would exceed 30% of an applicant’s income, and a census tract’s percentage of homeownership and median income.

6. The Internet Archive’s Wayback Machine can be accessed at https://www.archive.org/web/; here, websites are intermittently archived to provide access to fully or mostly functional past site versions.

7. In Model 7’s regression estimates of loan approvals with all covariates interacted with fintech lending status, no borrower race or neighborhood racial composition variables change their direction relative to the results in Table 2, but the interaction between fintech lending and Asian borrowers and the impact of a neighborhood with more than 80% non-White residents on both traditional and fintech lenders lose their statistical significance. In Model 8’s regression estimates of subprime lending with all covariates interacted with fintech lending status, all neighborhood and borrower race variables remain in the same direction as the results in Table 3, but fintech’s interaction with Black applicants loses its statistical significance, and fintech’s interaction with neighborhoods with more than 80% non-White residents gains statistical significance. All of these shifts are substantively small.

8. Kuebler (2012) is a notable exception, but her study is also unable to incorporate credit scores.

Additional information

Notes on contributors

Tyler Haupert

Tyler Haupert is a doctoral candidate in urban planning at Columbia University’s Graduate School of Architecture, Planning and Preservation. His research focuses on the social, economic, technological, and regulatory mechanisms contributing to racial segregation and exclusion in the United States, with particular interests in mortgage lending, federal housing policy, public education, and criminal justice.