Abstract

Accurate property valuation is critical for real estate markets, financial institutions, and urban planning. Traditional appraisal methods are time-intensive and subjective, while complex machine learning models often lack interpretability. This study addresses these challenges by developing an advanced linear regression framework that balances predictive accuracy with model transparency through systematic feature engineering. Utilizing the King County House Sales dataset comprising 21,613 transactions from May 2014 to May 2015, we developed 40 engineered features including interaction terms, polynomial features, ratio calculations, and location-based composites. After outlier removal using the interquartile range method, our dataset consisted of 20,467 properties with 55 total features. The optimized linear regression model achieved a test R² of 0.7198 with a normalized root mean square error (NRMSE) of 0.20 (20% of mean property value) and mean absolute error of 82,626. Feature importance analysis revealed that basement-to-living ratio, above-to-living ratio, and geographic coordinates were the most influential predictors. Cross-validation demonstrated model stability with a mean R² of 0.7316 (±0.0101). This research demonstrates that strategic feature engineering can significantly enhance linear regression performance for real estate valuation, achieving an average prediction error within 20% of property values while providing a transparent and interpretable alternative to complex machine learning algorithms.

Graphical Abstract

Novelty Statement

This research introduces a comprehensive feature engineering framework that substantially enhances linear regression performance for residential property valuation.