Comparative Study of Machine Learning Models for U.S. Housing Price Prediction

Abstract

This study compares machine learning models used to predict US house prices using a large-scale real estate dataset covering all US states, cities, and zip code regions. The goal is to evaluate prediction accuracy and identify effective preprocessing strategies for high-cardinality location features. After cleaning the data through missing value handling, deduplication, and outlier removal based on interquartile range (IQR), leakage-aware feature engineering was applied, including aggregating zip codes into ZIP3, grouping by the top K cities, and date decomposition, while excluding target derived variables. This study trains and evaluates three models-linear regression, random forest, and XGBoost-on a reserved test set using mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R²). The results show that XGBoost achieves the best performance, outperforming linear regression and random forest, and feature importance indicates that location indicators play a dominant role in prediction gain. The findings demonstrate that the XGBoost model outperforms linear regression and random forest models in predicting US house prices.