Improving Property Tax Assessments
2025-10-26
Property taxes depend on OPA’s assessed values.
Assessments often deviate from actual sale prices.
Some areas are consistently over- or under-assessed.
Unequal assessments create unfair tax burdens.

Source: Reinvestment Fund, Examining the Accuracy, Uniformity & Equity of Philadelphia’s 2023 Real Estate Assessments (Apr 2024).
Fairness: Ensure residents pay taxes aligned with true property values.
Transparency: Build trust through objective, data-based methods.
Efficiency: Improve city revenue stability and policy planning.
Traditional assessments rely on outdated, manual approaches.
Machine learning models can better capture local market dynamics.
Philadelphia Property Sales (n= 24023,2023-2024)
Census ACS (Income, Education, Poverty)
OpenDataPhilly (Number and Distance: Crime, Park&Recreation, Transportation, Hospital,School )


M1: Basic home features (size, age) → simple but limited
M2: + Census data → adds community context
M3: + Spatial data → captures location effects
M4: + Interactions → reflects real neighborhood differences

> “Average Error($)” shows the average gap between predicted and actual prices (lower = more accurate).
Interpretation
Blue areas: Homes undervalued by the model → may face under-assessment
Red areas: Homes overvalued → may face over-assessment
Central & southern zones: show the largest mismatches — indicating uneven market patterns
> The areas, shown in deep red or blue, are likely where property assessments are least accurate, and where tax fairness may be at greatest risk.
> These “hard-to-predict” areas should be prioritized for review in future assessment updates.

Review Where the Model Shows the Largest Gaps → Our residual analysis pinpoints neighborhoods with the highest prediction errors — the same areas where assessments are likely least fair. Prioritize these zones for reassessment and data verification.
Use the Model as a Fairness Benchmark → Instead of replacing official assessments, use the model as a cross-check tool to flag properties with unusually high or low assessed-to-predicted ratios.
Institutionalize Annual Model Updates → Retrain the model each year using new sales and census data so assessments stay current with real market trends, preventing future inequities.
Data gaps: Some neighborhoods have limited or inconsistent sales data, and important local factors like school quality or public amenities are not fully represented.
Spatial variation: The model’s accuracy differs across regions — it performs very well in some areas but less so in others, suggesting that geography strongly influences results.
Spatial modeling: Next, we will apply models that explicitly account for spatial relationships and neighborhood effects to improve prediction accuracy.
Data enhancement: We plan to expand data coverage in underrepresented areas to reduce bias and strengthen fairness across communities.
Xiaoqing Chen
Zicheng Xiang
Lingxuan Gao
Zhiyuan Zhao
Fan Yang
Zhe Fang
Sixers Consulting 6ers@upenn.edu|www.6ers.com

Prepared for the City of Philadelphia — Office of Property Assessment (OPA)