Domain-Driven Feature Engineering and Optimized Stacking Ensemble Learning for Used Car Price Prediction

Abstract

Accurate used car price prediction is challenging because vehicle values are influenced by complex nonlinear relationships among depreciation, engine characteristics, and market perception that are not adequately represented by the original attributes alone. While recent studies have primarily focused on developing more sophisticated regression models, the contribution of domain-driven feature engineering has received comparatively limited attention. This study proposes a domain-driven feature engineering and optimized stacking ensemble learning framework for used car price prediction. The proposed feature engineering pipeline constructs interaction features, engine performance indicators, brand- and market-related representations, and clustering-based features to enrich the original predictor space. These engineered features are integrated with an optimized stacking ensemble consisting of CatBoost, LightGBM, and XGBoost as base learners and Ridge Regression as the meta-learner, where hyperparameters are optimized using Optuna. The framework is evaluated on the public Used Car Price Prediction benchmark dataset containing 4,009 vehicle records using both an independent hold-out test set and 5-fold cross-validation. Experimental results demonstrate that the proposed framework achieves an R² of 0.9112 on the hold-out test set and 0.9023 ± 0.0025 under cross-validation, outperforming the individual gradient boosting models. Ablation experiments further reveal that the domain-driven feature engineering contributes more substantially to prediction performance than model optimization alone, with interaction features providing the largest performance gain. SHAP analysis confirms that engineered interaction features dominate the prediction process, highlighting vehicle age–mileage interactions as the most influential factors affecting resale value. These findings demonstrate that constructing informative feature representations, together with optimized ensemble learning, provides an effective and interpretable solution for used car price prediction.