Machine learning algorithm that prices properties’ value benefitted market

Housing tends to be a key part of household wealth, but despite its importance, it has been difficult to measure the value of a property. In a new article, researchers studied the impact of a popular machine learning pricing algorithm on the U.S. housing market. The study found that, overall, the algorithm benefitted the market.

The study, by researchers at Carnegie Mellon University, New York University, and the University of Toronto, is published in Marketing Science.

“Our work is among the first to demonstrate and explain the impact of a machine-generated property value prediction on the housing market, as well as its economic and social implications,” explains Param Vir Singh, Carnegie Bosch Professor of Business Technologies and Marketing at Carnegie Mellon’s Tepper School of Business, who coauthored the study.

Several online real-estate marketplaces have proprietary algorithms that use large amounts of data to estimate property values. The algorithms’ estimates are displayed on the marketplaces’ websites.

In this study, researchers analyzed the impact on housing market outcomes of Zestimate, a machine learning algorithm developed by Zillow, the most popular U.S. real estate marketplace company and one of the first to publish algorithm-generated property value estimates nationwide. Because Zestimate tends to be more accurate for rich than for poor neighborhoods, raising concerns that the tool may increase socioeconomic inequality, researchers also examined how and to what extent Zestimate’s effects differed across socioeconomic segments.

The study focused on the more than 4,000 properties listed in 140 neighborhoods in Pittsburgh, PA, between February and October 2019. Using data on Zestimate and housing sales, researchers built a structural model of the housing market in which sellers and buyers faced uncertainty about property values and Zestimate provided an unbiased signal of property value.

Overall, Zestimate benefitted both buyers and sellers in the housing market, leading on average to a 5.4 percent rise in buyers’ surplus and a 4.2 percent increase in sellers’ profits. This occurred primarily because its effect in reducing uncertainty allowed sellers to be more patient and set higher reservation prices to wait for buyers who truly valued the properties, which improved the quality of seller-buyer matches, say the authors.

In addition, Zestimate reduced socioeconomic inequality in housing: Both rich and poor neighborhoods benefitted from Zestimate, but poor neighborhoods benefitted more. The average seller profit increase by Zestimate in poor neighborhoods was nearly 5 percent, compared to 3.7 percent and 4 percent in midrange and rich neighborhoods; the average buyer surplus increase by Zestimate in poor neighborhoods was 8.3 percent, compared to 3.2 percent and 6 percent in midrange and rich neighborhoods. This occurred because poor neighborhoods face greater prior uncertainty and, therefore, would benefit more from new signals, according to the authors.

Among the study’s limitations, the authors note that they did not model a case where multiple buyers engaged in a bidding war, which usually boosts the selling price. In addition, their model considered buyers who looked at one property at a time, but buyers may look at multiple listings, leading to competition among sellers.

“The emergence of algorithms that predict the market value of properties has the potential to reduce uncertainty in a housing market,” suggests Yan Huang, Associate Professor of Business Technologies at Carnegie Mellon’s Tepper School of Business, who coauthored the study.

Coauthor Kannan Srinivasan, Professor of Management, Marketing, and Business Technology at Carnegie Mellon’s Tepper School of Business, notes, “When evaluating the disparate impact of algorithms, it is crucial to consider the status quo of different groups and the counterfactual outcomes when the algorithms are not there, rather than simply looking at the algorithm’s outputs.”