Problem
Property assessments influence tax bills. A model that misses unevenly can shift tax burden toward lower-priced homes even when its average error looks acceptable.
Approach
I modeled log sale price with linear regression, engineered stable property features, and evaluated both accuracy and distributional error patterns.
Outcome
The model validated consistently, but diagnostics revealed a regressive pattern: cheaper homes were more likely to be overvalued than expensive homes.
Fairness Diagnostic
Share of homes overestimated by the model in the original notebook run. For assessment, higher overestimation means a higher likely tax burden relative to actual sale value.
Modeling Workflow
1. Clean
Remove invalid sale prices and handle missing values without dropping test parcels.
2. Engineer
Use log building value, log building square feet, and bathroom count features.
3. Validate
Check train, holdout, and cross-validation RMSE on log sale price.
4. Audit
Compare residual behavior across price tiers using RMSE, MAPE, and overestimation rate.
Original Process
The notebook workflow is preserved as a narrative process note covering EDA, feature engineering, validation, and fairness analysis.
What I Would Improve Next
Add neighborhood encodings, spatial validation, model cards, and monitoring dashboards for price-tier and community-level residual parity.
Skills Shown
Python, pandas, scikit-learn, feature engineering, model validation, fairness analysis, communication of statistical tradeoffs.
Run It
Use python tests/smoke_test.py for a local check, then
python scripts/run_model.py --train data/cook_county_train.csv
when the source data is available.