Charlotte, North Carolina
Build a robust predictive model that estimates home prices using spatial and non-spatial predictors, and evaluate residual autocorrelation patterns to understand where systematic bias remains.
Application of log-transforms to price data to address skewness. Integration of complex interaction terms and high-resolution distance-to-park metrics improved explanatory power across varied housing submarkets.
Global OLS regression implemented with rigorous 10-fold cross-validation. Performance assessment focused on RMSE reduction and predictive stability across central and peripheral neighborhoods.
Advanced residual analysis using global and local Moran's I revealed where the model failed to capture neighborhood-specific drivers of value, especially between urban core and suburban tracts.
This section assembles the core analytical outputs from the Charlotte housing model, from exploratory relationships to predictive fit and residual diagnostics. Together, these visualizations show both what the model explains well and where spatial bias persists.
This heat map shows strong positive relationships among square footage, building grade, and sale price. It also flags where predictors begin to overlap, which informed later variable selection and pruning decisions.
The scatter suggests a negative relationship between local crime incidence and home value. Even with substantial variation, higher crime counts generally align with lower-priced properties.
Log-transforming area produces one of the clearest positive associations in the dataset. This made square footage a core structural predictor in the housing price model.
After log transformation, predicted and observed values align more tightly. The improved fit suggests the transformation reduced skewness and stabilized variance across the sample.
The observed Moran's I sits away from the permutation baseline, indicating statistically meaningful residual clustering. In practice, that means spatial bias remains even after accounting for major housing attributes.
| Variable Category | Observational Narrative | Statistical Significance |
|---|---|---|
| Physical Attributes | Total home size combined with professional building grade classification emerged as the primary drivers of market valuation. | p < 0.001 |
| Socio-Economic | Elevated regional crime rates systematically depressed predicted property values across the metro area. | Strongly Negative |
| Environmental Proxies | Park proximity demonstrated a comparatively weak effect on valuation, challenging the assumption of a uniform green-space premium. | Weak Correlation |
| Systematic Bias | Critical discovery: the model consistently underpredicts central urban neighborhood values while overpredicting peripheral suburban properties. | High Magnitude |
This map shows the uneven geography of housing values across the metro area. Higher-value clusters concentrate in select central and southern neighborhoods, while lower values dominate more peripheral areas.
Crime intensity is spatially concentrated rather than evenly distributed. These concentrations overlap with some of the model's lower predicted-value zones, reinforcing crime as a meaningful neighborhood-level signal.
Residual clusters reveal where the model systematically underpredicts and overpredicts local values. Central neighborhoods tend to contain positive residual hotspots, while some outer areas show overprediction.
VIF testing revealed high redundancy between bedroom counts and total square footage, requiring variable pruning.
Exceptional high-value luxury properties in South Charlotte disproportionately skewed global regression coefficients.
Absence of high-granularity school district performance and transit proximity limits the model's localized predictive power.
Prioritize targeted infrastructure investment in central underpredicted zones to capture latent value.
Integrate spatial lag variables into official tax assessment workflows to reduce systematic valuation bias.
Deploy localized crime prevention strategies as a direct mechanism for property value stabilization.