Why does my Zestimate fluctuate? Platform design of Algorithmic Pricing Models

Abstract

Machine Learning (ML) algorithm-generated price estimates are increasingly common in home sales, used car sales, and short-term rentals. These algorithmic prices are informative to consumers but online platforms unilaterally control the underlying algorithm design. The revenue model for these platforms may not be fully aligned with the interests of consumers. Further, these price estimates tend to fluctuate significantly over time. It is therefore puzzling if the fluctuations reflect real changes in demand, or are simply artifacts of the platform’s opaque choice of algorithm design.

In this paper, we develop an analytical model grounded in the housing market. We show that the platform, relative to consumers (homeowners), prefers to induce excess market entry and sales volume. The platforms can achieve this objective by pricing excess features (compared to statistically optimal choice) that result in an over-fit Machine Learning model and excessive fluctuations in the algorithmic prices. The consumers (homeowners) are worse off under this platform’s optimal (over-fit) relative to the statistically optimal (best-fit) model choice. These results have implications for regulating algorithmic prices offered by online platforms.

Keywords: Platforms, Algorithmic Pricing, Bias-Variance tradeoff, Economics of AI, Housing Market