Comparative Evaluation of Linear Regression, Tree-Based Regression, and Neural Network Models for Structured Car Price Prediction

Abstract

Accurate car price prediction plays a critical role in modern digital marketplaces, dealership platforms, and automated financial systems. Traditional valuation methods are often subjective, whereas machine learning (ML) enables data-driven modeling of complex, non-linear relationships among vehicle attributes. This study evaluates the performance of three ML models—Linear Regression, Decision Tree Regressor, and Multilayer Perceptron (MLP)—on a publicly available dataset of 205 used car listings. After comprehensive preprocessing and an 80:20 train-test split using Scikit-learn with fixed randomization, model performance was assessed via RMSE, MAE, and R² metrics. The Decision Tree Regressor achieved the best results (R² = 0.886), outperforming both the linear and neural models. Additionally, a literature-based benchmark against ten recent studies shows that interpretable models like Decision Trees can rival more complex techniques such as XGBoost, CNNs, and stacked ensembles when applied to well-prepared tabular data. These findings highlight the practical value of simplicity and interpretability in real-world pricing systems.