A fairer property algorithm can still drift with the housing market

In a new arXiv preprint, Meng Liu of Washington University in St. Louis examines this problem through Cook County’s property assessments. The study finds a substantial improvement after Chicago adopted a modernized valuation system in 2021, but no detectable reduction in price-based assessment inequality when the system first reached the northern and southern suburbs. The contrast illustrates what the paper calls fairness drift: disparities can change between updates even when assigned values do not.

What counts as fair assessment?

Property assessments help determine tax bills. A central fairness concern is whether inexpensive homes are assessed at a higher fraction of their market value than expensive homes. This pattern, called price regressivity, places a relatively heavier assessment burden on lower-priced properties.

Liu measures it using the ratio of assessed market value to sale price. If that ratio falls as sale prices rise, cheaper homes are relatively over-assessed. The study also examines whether equally priced homes receive different assessment ratios depending on their neighborhood’s Black and Hispanic population share. That is a neighborhood-level measure, not a direct measure of how individual owners are treated by race.

Cook County offers a useful comparison because it reassesses three geographic regions on a staggered, three-year cycle. The reformed system reached Chicago in 2021, the North suburbs in 2022 and the South and West suburbs in 2023.

The modernization combined gradient-boosted decision trees—models that combine many branching prediction rules—with richer property and geographic information, a documented data pipeline and revised post-modeling practices. It therefore changed more than the prediction algorithm alone.

Comparing changes, not just averages

The main Chicago analysis covers 176,567 home sales across 38 townships during 2018–2022, excluding northern suburban observations after that region adopted the reform.

Liu uses a difference-in-differences approach: comparing changes in Chicago with changes over the same period in regions not yet using the reformed system. Rather than asking whether average assessments rose or fell, the analysis asks whether their relationship with sale price and neighborhood composition changed.

The statistical design accounts for local quarterly shifts and differences in baseline relationships. Its causal interpretation depends on the assumption that, without the reform, these relationships would have evolved similarly in the treated and comparison areas.

Chicago improved; the suburbs differed

In Chicago, the estimated reduction in price regressivity was about half its pre-reform magnitude. It appeared immediately and persisted through the following two years. Mean absolute assessment error also fell, from 40.0% to 28.0%.

The evidence is consistent with the new system distinguishing properties that older, coarser valuation rules had treated too similarly. Finer geographic information became more important, and accuracy gains were especially pronounced away from middle-tier neighborhoods.

Among equally priced homes, the relationship between assessment ratios and neighborhood minority share also moved toward zero. This finding was less statistically decisive than the price result under the paper’s preferred uncertainty calculation.

The northern suburbs started with substantially more accurate assessments, leaving less room for improvement. Their mean absolute error declined only modestly, and the price gradient did not change detectably.

The southern suburbs presented a different challenge. Inexpensive neighborhoods had appreciated faster while assessments remained fixed. Their assessment-to-price ratios consequently fell, reducing part of the existing disparity before the reform arrived.

Then that unusually rapid appreciation slowed in the reassessment year. Fresh assessments were relatively high compared with subsequent sale prices in neighborhoods that had previously appreciated fastest. Overall accuracy did not improve. The paper interprets these patterns as evidence that market movements shaped both the baseline the reform encountered and the errors its new values displayed.

What the findings do—and do not—establish

This is a preprint, and several qualifications matter. Only three geographic rollout groups underpin the main comparison, making statistical uncertainty difficult to assess. The reform bundled model, data and process changes, so the study cannot assign Chicago’s gain to one component.

The proposed explanation for the South’s missing accuracy gain also draws on reconstructed models, not the assessor’s actual production models. Liu presents that exercise as a plausible explanation rather than an identified cause.

The practical implication is to monitor fairness between deployments, not simply at launch. The paper suggests using divergence between market segments to trigger reviews, but does not test whether this would outperform calendar-based audits. Its narrower finding is that an unchanged valuation system can become more or less equitable as its market moves.