How Machine Learning Mitigates Racial Bias in the U.S. Housing Market

Abstract

I examine racial bias in the most popular home valuation algorithm and study the algorithm’s impact on racial bias in transaction prices. I find statistically significant but economically small racial bias in the algorithm. For example, while Black buyers overpay by 9.3% in prices relative to White buyers for similar homes, the algorithm only overvalues the same transactions by 1.1%. The algorithm inadvertently learns racial bias from patterns in historical transaction prices. The algorithmic racial bias is small because the algorithm is designed to be insensitive to transitory pricing factors related to behavioral biases, sellers’ liquidity conditions, and buyer or seller race. Exploiting the staggered rollout of the algorithm in a neighboring ZIP Code setting, I find that if the algorithmic valuation is available for all the homes in an area, it reduces the overpayment of Black buyers relative to White buyers by 4.8%. The results suggest that the application of slightly biased machine learning algorithms can mitigate social bias if they are less biased than humans.