Does Machine Learning Amplify Pricing Errors in Housing Market? : Economics of ML Feedback Loops

Abstract

Numerous ML pricing models (Zillow’s Zestimate, Redfin Estimate) have been deployed to make house sale price predictions. They appears to be independent and unbiased signal to resolve pricing friction in the housing market. These ML models – learn from live sale prices and influence the same sales simultaneously. This creates a Feedback Loop where the ML model is confounded by its own previous version. We theoretically show how this Feedback Loop creates a self fulfilling prophecy where ML over estimates its own prediction accuracy and market participants over rely on ML predictions. We use data from Zillow’s Zestimate to establish necessary primitives for the theoretical Feedback Loop phenomenon. We also structurally estimate seller payoffs under current and counterfactual ML regimes. We show that ML pricing, instead of alleviating, may widen payoff disparity in favor of sellers with greatest ability to price. This happens because ML lowers pricing Disagreement but adds pricing Bias, with both effects amplified under strong Feedback and high capacity ML.