What makes a place
valuable?
Behind every housing market is a place.
Its incomes. Its people. Its housing stock.
This study asks how much those characteristics can tell us.
Can county characteristics explain differences in home values across the US?
3 non-overlapping ACS periods
3 model families
of 2024 cross-county value variation
captured by the selected model¹
mean absolute prediction error
in constant 2024 dollars
counties in the held-out
2024 evaluation
¹ R² = 0.801.
Characteristics only.
No previous home values.
A national story.
Thousands of local realities.
The model knows the characteristics of a county—not its past home values. Explore where its estimates match reality, and where they fall short.
Select a county on the map or use the selector. Gray indicates no eligible target value. Boundaries are simplified; DC and small counties are easiest to find by name.
Source: Census ACS five-year estimates. Outcome: median value of owner-occupied housing, not sale price or the value of an individual home.
Prediction is a test.
Not an explanation by itself.
Three model families face the same held-out 2024 data. Each learns from earlier, non-overlapping periods. A simple median baseline sets the bar.
How far off are the estimates?
LOWER IS BETTERCounty characteristics carry a substantial signal.
The selected gradient boosting model reduces proportional-scale error by 63.5% relative to the training-median baseline.
It still misses important local variation. Predictions tend to understate values at the expensive end of the market.
Download model comparisonLearn initial relationships
Choose parameters and model
Evaluate the frozen selection
After selection, models are refitted on the first two periods. Predictors describe the same ACS period as their target. This is temporal transfer of an association—not an advance forecast.
The economy leads.
The rest adds context.
Remove one feature group and ask how much worse the selected model gets. Economic characteristics make the largest difference.
Increase in error when a group is removed
Relative increase in held-out RMSLE. Same counties and fixed model settings. Correlated groups can substitute for one another.
What changes within a county?
After controlling for persistent county differences and shared period effects, higher income and educational attainment are positively associated with home values; vacancy is negatively associated.
Changes within a county are harder to explain than differences between counties. These R² values describe different forms of variation.
See the six estimated associations
County and period fixed effects; 95% county-clustered intervals. Percentage predictors are measured in percentage points. Income uses log1p dollars. These are not causal effects.
History does most
of the heavy lifting.
Knowing a county’s previous home values creates a strong forecast. Adding its characteristics provides a smaller, measurable improvement.
lower RMSLE after adding characteristics
to the history-only model
Next-release prediction, 2023–2024
Mean absolute error in 2024 dollars. These annual ACS estimates share four survey years, which makes persistence unusually strong. Release vintages were not reconstructed; this is a retrospective experiment.
From raw records
to a defensible finding.
An end-to-end workflow, with explicit decisions about measurement, geography, leakage, and generalization.
- 01
Gather & organize
Census ACS, FEMA disaster declarations, NOAA storm events and NCEI climate records, joined by county and year in DuckDB marts.
- 02
Audit & align
Validate unique keys, missing values and variable definitions. Put dollar values in 2024 terms. Flag changing boundaries and document excluded counties.
- 03
Design the comparison
Use three non-overlapping ACS periods and remove price-derived inputs. Keep affordability and climate as separate sensitivity analyses.
- 04
Fit & validate
Compare Elastic Net, Extra Trees and gradient boosting. Fit preprocessing inside training folds; select on validation data before evaluating 2024.
- 05
Challenge & interpret
Hold out entire counties, remove feature groups, permute related features together, and estimate within-county associations with clustered uncertainty.
- 06
Make it reproducible
Export the panel, feature definitions, predictions and diagnostics. Preserve seeds and parameters, and test the assumptions that guard against leakage.
Useful evidence.
Honest boundaries.
A portfolio is stronger when the limits of its conclusions are visible.
Association is not causation
Income, education, housing stock and value influence one another and share unmeasured drivers. Predictive importance does not identify the effect of changing a county characteristic. County-level patterns do not describe individual households.
Home value is not sale price
The target is the ACS median value of owner-occupied homes. It is a survey estimate in constant 2024 dollars, not a transaction index or a property appraisal. The model does not incorporate every local market condition.
Time and geography require care
The main study compares non-overlapping five-year windows. Annual forecasts use overlapping estimates. The fixed-effects sample excludes incomplete histories and known geographic changes; no areal crosswalk is invented, and minor changes may remain.
Coverage and uncertainty are imperfect
Climate results cover a subset of counties. Event absences are assumed zero in the source marts. ACS margins of error and spatially clustered uncertainty are not modeled; neighboring counties may appear in different validation folds.
What was available, and when?
ACS end-year labels are not release dates. The 2020–2024 estimates were released January 29, 2026. Forecast experiments do not reconstruct real-time data or CPI vintages.