The question
Can nonlinear models improve residential price prediction?
I compared linear and tree-based approaches to understand whether nonlinear relationships produced a meaningful reduction in prediction error.
Boston University · Machine Learning
Housing Price Prediction and Model Validation
A comparative modeling project focused not only on predictive performance, but also on residual behavior, fairness risk, and whether the result was ready for real-world use.
My contribution
End-to-end model comparison
I prepared the data, explored feature relationships, trained competing models, evaluated cross-validation results, reviewed residuals, and translated the findings into production-risk recommendations.
Approach
From baseline to validation review
1. Establish a baseline
Used linear regression to establish an interpretable benchmark and reveal where simple assumptions were insufficient.
2. Compare nonlinear models
Evaluated Random Forest and Gradient Boosting to capture interactions and nonlinear patterns.
3. Inspect failure modes
Reviewed RMSE, MAE, R², residual patterns, high-value property errors, and possible location-proxy concerns.
Results
Best tested model: Random Forest
Validation resultRMSE approximately $290,919 · MAE approximately $184,569 · R² approximately 0.517
The model captured meaningful signal and outperformed the simpler baseline, but the error remained too high for high-stakes appraisal use without stronger location features, market-specific validation, fairness review, and ongoing monitoring.
What the project demonstrates
- Exploratory data analysis and feature preparation.
- Regression, Random Forest, and Gradient Boosting comparison.
- Cross-validation and business-focused metric interpretation.
- Residual analysis and model-risk thinking.
- Clear communication of limitations instead of overstating accuracy.
What I would improve next
- Add richer geographic and temporal features.
- Test robustness across market segments and price bands.
- Evaluate fairness and proxy-variable risk more formally.
- Create monitoring thresholds for drift and residual changes.
- Document intended use and non-use in a model card.
Technology