Machine Learning Approaches to House Price Prediction

Abstract

Housing price prediction is of significant value to households, mortgage lenders, real estate developers, and policymakers. However, given that property values ​​are shaped by a complex interplay of factors, including hard-to-quantify intrinsic property attributes, socioeconomic conditions, and macroeconomic variables, traditional methods are not always effective. This paper examines mainstream machine learning approaches to housing price prediction and compares their theoretical foundations and practical applications, covering hedonic regression, random forest, XGBoost, and artificial neural networks. Previous studies applying these methods suggest that machine learning may be able to provide better predictions than traditional methods when non-linearities and interactions are important, but that no single model dominates in all markets. Tree-based ensembles are competitive at the property level in many applications but simpler regularized linear models can be competitive or better if the predictive signal is sparse or predominantly macroeconomic. The performance differences may also be driven by regional heterogeneity, feature engineering choices, validation design, and data quality. The paper concludes that successful house price prediction depends on a problem-tailored modeling approach that balances out-of-sample predictive performance with economic interpretability and sensible model diagnostics.