Fair and equitable predictive modeling for property assessment

A mass-appraisal model can be made more accurate and still leave some owners systematically over-assessed; the two goals have to be optimized together.

Schematic. Attainable models sit above a frontier along which lower predictive error can only be bought with higher systematic inequity. No data is shown.
Status
Ongoing research
Period
2024 — present
Themes
Fair & reliable predictive modeling · Scalable optimization & computation

Research question

How can mass-appraisal models improve predictive performance without worsening systematic vertical inequity across properties?

Why it matters

Property assessments decide how a tax burden is divided. The model that produces them is a public instrument, so it has to be judged on more than aggregate accuracy: what matters is also whether its errors fall in a systematic pattern. When low-value properties are assessed high relative to their sale price and high-value properties low, the result is regressivity — a distributional outcome produced by a modeling choice.

The challenge

Vertical equity is measured by assessor-facing statistics such as the price-related differential (PRD), the price-related bias (PRB) and the vertical equity index (VEI). None of them is the loss a standard gradient-boosted model minimizes, and improving the loss does not reliably improve them. Treating equity as a post-hoc diagnostic therefore leaves the trade-off implicit, at exactly the point where it should be stated and chosen.

Approach

The public pipeline evaluates models under a realistic temporal protocol: time-split sales data with rolling-origin validation, so a model is always scored on sales it could not have seen. Against that protocol it compares baseline models with fairness-regularized ones on both predictive error and the assessor-facing equity measures.

The methodological direction is to move the equity criterion inside the optimization problem rather than around it. This takes two forms: penalizing the covariance between assessment ratio and value directly in the training objective, and combining models by convex stacking whose weights are chosen subject to vertical-equity constraints across validation folds. Both make the trade-off a decision variable instead of a diagnostic.

Research to practice

In July 2026 the Cook County Assessor’s Office merged support for the covariance-penalized LightGBM objective developed with the MIT research team into its open-source residential assessment model, where it is selectable as the mse_cov objective.

Status

Ongoing research at the MIT Operations Research Center. The public materials document the modeling and validation pipeline; no empirical findings are reported here.

All research