Why Does My Zestimate Fluctuate? Model Overfitting for Platform Ad Revenue

Machine Learning (ML) generated home value estimates, such as Zillow's Zestimate and Redfin Estimate, are gaining attention. They are usually marketed as free to use algorithmic prices designed to increase transparency in housing markets. While these estimates are shown to be reasonably accurate on average, they also tend to fluctuate significantly for a property over time. It is therefore puzzling where the fluctuations come from. Do they reflect real changes in property values, or are they simply unintended artifacts of the algorithms?In this paper, we demonstrate with an analytical model that one potential contributing factor of the large fluctuations is the platform's incentive to maximize its own payoffs. Drawing on the fact that these platforms usually serve or partner with real estate agents (e.g., Zillow) or are real estate agents themselves (e.g. Redfin), we show that the platforms have an incentive to manipulate ML estimates to induce excess market entry and increase sales volume. When directly manipulating ML estimates is not feasible, the platforms would adjust the level of model fitness in the algorithms to achieve the same goals. This leads to overfitting of the ML models and excessive fluctuations in the ML estimates. These results persist even if homeowners’ have full knowledge of platforms' choice of model fitness. Our results suggests that the seemingly harmless fluctuations may reflect the misaligned interests between platforms and homeowners, and may harm homeowners