Comparing Linear Regression and Random Forest for Housing Price Prediction: Insights from the Boston Housing Dataset

Abstract

Housing price prediction is a critical task in real estate and economic analysis, providing valuable insights for stakeholders such as homebuyers, sellers, and policymakers. This study focuses on the Boston Housing dataset, a benchmark with 505 samples and 14 features, to predict the median value of owner-occupied homes (MEDV) using Linear Regression and Random Forest Regression. Exploratory data analysis reveals non-linear patterns, such as the right-skewed distribution of MEDV (skewness = 1.11) and strong correlations with features like LSTAT (-0.74) and RM (0.70). The dataset was standardized and split into 80-20 training and testing sets for model evaluation. Results show that Random Forest outperforms Linear Regression, achieving an MSE of 7.58 and R² of 0.864 compared to 19.38 and 0.652, respectively. Feature importance analysis highlights LSTAT and RM as key predictors, emphasizing socio-economic and structural influences. While Random Forest excels in capturing non-linear relationships, Linear Regression offers interpretability for policy insights. However, the dataset’s historical context and small size limit its applicability to modern markets, suggesting future research with larger, contemporary datasets and advanced models.