Are We Fair Yet? A Practical Analysis of Bias and Barriers to Fairness in Mortgage Lending

Abstract

The computer science community has made strides in identifying various ways bias can infiltrate a system, exposing bias in existing systems, and proposing mechanisms for mitigating bias. Unfortunately, much of this work has been limited by approaching bias like a bug in code and ignoring the socio-technical system surrounding any automated decision-making aid. This paper investigates observed bias in the U.S. mortgage lending system, where two underwriting algorithms are used to make approval decisions on 90% of mortgage applications. Recent work has exposed disparate rates of approval across racial lines by these supposedly “colorblind” underwriting algorithms. Using real loan application data, this paper explores the effectiveness of an algorithmic fairness intervention in mitigating the system’s disparate impact on minorities. While automated underwriting systems are in a unique position to have an industry-wide impact, this paper finds that the effectiveness of an algorithmic intervention, such as the one explored in this paper, is substantially limited by structural and legal barriers. Thus, this paper identifies bias as an inevitable byproduct of the socio-technical system rather than a result of flawed code operating in an otherwise fair context. Much of the work in algorithmic fairness assumes a just environment surrounding an algorithm and a fair legal structure supporting equitable processing of data. This paper, however, considers the societal and legal constraints in conversation with the technical challenges of implementing a fair system. Prior research has demonstrated the failure of existing laws to prevent discrimination in mortgage lending, resulting in biased outcomes. This paper expands on these findings to explore practical steps for mitigating the disparate impact of this system. This paper recommends technical bias mitigation but also regulatory and legal changes to enable and support such measures.