Automated valuation models (AVMs) are increasingly used to estimate home values and streamline mortgage lending. Yet these algorithmic tools can unintentionally reproduce disparities across homeowners. This study provides the first evidence of AVM valuation errors at the individual level. It examines whether AVMs systematically produce larger errors or undervaluation for certain groups of homeowners, even after accounting for property and neighborhood characteristics. The analysis offers an empirical framework for identifying algorithmic bias and advancing fairer, more transparent home valuation systems.

Why This Matters

AVMs are transforming how properties are valued, influencing lending, refinancing, and access to home equity. But when valuation errors differ across groups of homeowners, they can reinforce inequities in wealth building and financial opportunity. Policymakers, lenders, and regulators need to understand these risks to ensure that algorithmic innovations improve efficiency without deepening disparities in the housing market.

Key Takeaways

• AVMs tend to yield greater valuation errors for Black homeowners. In both Atlanta and Memphis, AVMs produced valuation errors 3.4 percentage points higher for Black homeowners than for white homeowners, even after controlling for property and neighborhood characteristics.
• Black-owned homes are systematically valued lower than comparable white-owned homes. On average, AVMs undervalued Black-owned properties by about 5 percent, limiting opportunities for equity building, refinancing, and credit access.
• Disparities in AVM valuations can stem from both data and algorithmic design. AVMs rely on historical sales and neighborhood data shaped by past segregation, and they are often optimized for the largest group of homeowners. This combination can unintentionally prioritize accuracy for majority populations.
• Policy action is needed to strengthen transparency and fairness. Solutions include improving data quality, publishing AVM performance metrics, and integrating fairness metrics during model training. Collaboration among regulators, industry, and researchers can help ensure that automation advances both accuracy and equity.

How We Did It

We used a three-step analytical approach to evaluate whether AVMs produce unequal valuation outcomes across homeowners. First, we applied a validated race imputation technique—fully Bayesian Improved Surname Geocoding (fBISG)—to estimate each homeowner’s likely racial or ethnic background using surname and geographic information. Second, we measured AVM valuation errors by comparing AVM estimates with actual sales prices, controlling for a range of property and neighborhood characteristics. Finally, we examined undervaluation patterns, assessing whether certain groups of homeowners experience systematically lower AVM estimates for comparable homes. This method allows us to identify equity and accuracy gaps in automated valuation systems even without access to self-reported demographic data.