Do County Characteristics Explain Home Values?
Behind every housing market is a place: its incomes, its people, and its housing stock. This study tests how much those characteristics tell us about county median home values, then separates differences between counties from changes within them.
The question
Can county characteristics explain differences in housing values, and do those relationships generalize across places and time?
The outcome is the Census ACS median value of owner-occupied housing, expressed in constant 2024 dollars. It measures county-level survey estimates. Transaction prices and individual-home appraisals answer different questions.
From data engineering to modeling
Census ACS characteristics, FEMA disaster declarations, NOAA storm events and NCEI climate records are cleaned and organized in DuckDB marts. The study audits keys, missingness, variable definitions and geographic changes before fitting models.
The main panel uses 9,420 county-period observations from three non-overlapping ACS windows: 2010–2014, 2015–2019 and 2020–2024. Its 57 candidate primary predictors exclude previous home values and price-derived ratios; joint model and predictor selection retains 30 predictors. Additional climate and affordability measures enter separate sensitivity analyses.
What the models found
The current published study compares Elastic Net, Extra Trees and histogram gradient boosting. Three county-disjoint folds within 2014 rank predictors and propose subsets. Six configurations compete on the same 57 unique subsets: 342 comparisons on 2019 validation data. The smallest predictor set within 1% of the best validation RMSLE is selected, with ties resolved by error, candidate and subset. The selected design is refitted on 2014 and 2019 before retrospective evaluation on 2024.
- Selected model: Extra Trees with 30 predictors, chosen using the 1% validation rule.
- Held-out 2024 performance: R² 0.798, with $37,860 mean absolute error across 3,139 counties.
- Baseline comparison: RMSLE is 62.6% lower than predicting the training-set median.
- Feature-group comparison: removing economic features raises RMSLE by 22.4%, the largest deterioration among the groups tested.
These predictions use characteristics from the same ACS period as the target. The 2024 outcomes had been inspected in earlier project results, so the revised comparison provides retrospective evidence. Household income, education, and other income measures lead the selected model’s permutation-importance ranking; correlated predictors can substitute for one another.
Secondary analyses
Geographic validation. Five county-disjoint outer folds repeat the complete selection procedure on development data. Extra Trees is selected in all five folds. Pooled out-of-fold R² is 0.8340, RMSLE is 0.1792, and mean absolute error is $25,254. These results measure transfer to excluded counties; neighboring counties can still share spatial similarities.
Within-county associations. A regression with county and period fixed effects uses the same 30 selected inputs across 3,122 geographically balanced counties, with median-imputed predictors. Its within R² is 0.2536, which has a different denominator from the cross-county benchmark. County-clustered intervals describe exploratory associations conditional on the selected inputs and imputation.
Climate and affordability measures. On a matched 1,594-county evaluation sample, additional climate inputs reduce RMSLE by about 0.9%. On the same 3,139 counties as the primary evaluation, affordability inputs increase RMSLE by about 3.0%. These comparisons hold the specification fixed and describe sensitivity to the added bundles and their coverage.
Local response curves. Varying a predictor near observed county values shows how the fitted model’s estimates respond within supported ranges. The curves describe model associations; real-world interventions require separate evidence.
Annual home-value forecasting is outside this study’s scope. Consecutive ACS five-year releases share four survey years, so predicting the next release targets an updated period estimate. This study focuses on relationships across non-overlapping periods and within counties.
What you can explore
- A searchable county map showing observed values or model errors using official 2024 boundaries.
- County profiles with model estimates, income, education, vacancy and period values.
- Model and predictor-subset comparisons, selected-model importance, geographic validation, fixed-effects estimates, and climate/affordability comparisons.
- Downloadable predictions, feature definitions, coverage audits and reproduction instructions.
The study estimates observational associations using same-period characteristics and outcomes. Geography changes are screened conservatively, and climate coverage is incomplete. Source marts assume zero when hazard events are absent. Reported intervals leave out ACS sampling margins of error, spatial dependence, and uncertainty from repeating the full model search. The 2024 ACS five-year estimates were released in January 2026, and their outcomes were previously inspected; interpretation is retrospective.
Explore the analysis
Follow the code and evidence in three executed notebooks on GitHub. Each documents the completed study and includes saved outputs; a compact evidence snapshot supports running the walkthroughs without the original database.
- Modeling dataset — Feature engineering, data audits and inclusion rules.
- Training and validation — Training-only preprocessing, joint model and predictor selection, and county-grouped folds.
- Evaluation and interpretation — Retrospective errors, feature importance, secondary analyses, and supported associations.
Reproducibility
Preprocessing is fitted inside training folds. Separate geographic validation holds out entire counties and repeats tuning inside each fold. The workflow exports its data audit, feature manifest, model selection, predictions and diagnostics; automated checks test temporal alignment, geographic separation and fixed-effects calculations.