Human-Algorithm Interactions: Evidence from Zillow.com

Using bi-weekly snapshots of Zillow in three US cities, we document how home sellers and buyers interact with Zillow's Zestimate algorithm during the sales cycle of residential properties. We find that listing and selling outcomes respond significantly to Zestimate, and Zestimate is quickly updated for the focal and comparable houses after a listing or a transaction is completed. The user-Zestimate interactions have mixed implications: on the one hand, listing price depends more on Zestimate if the city does not mandate disclosure of sales information or if the neighborhood is more heterogeneous, suggesting that Zestimate provides valuable information when alternative information is more difficult to obtain; on the other hand, the post-listing update of Zestimate tracks listing price more closely in non-disclosure and heterogeneous neighborhoods, raising the concern that the feedback loop may propagate disturbances in the sales process. However, by leveraging COVID-19 pandemic as a natural experiment, we find no evidence that Zestimate propagates the initial shock from the March-2020 declaration of national emergency, probably because Zestimate has built-in guard rails and users tend to adjust their confidence in Zestimate according to observed market outcomes.