Accurate prediction of vehicle-level CO \(_2\) emissions is critical for environmental assessment and regulatory decision-making. This study proposes a leakage-safe stacked ensemble regression framework for predicting passenger vehicle CO \(_2\) emissions using technical specifications and fuel consumption data. The framework integrates four heterogeneous base learners Random Forest, XGBoost, Support Vector Regression, and Multi-Layer Perceptron whose out-of-fold predictions are combined through a Ridge Regression meta-learner to improve robustness and prevent data leakage. Domain-driven feature engineering, including efficiency-normalized fuel consumption ratios and engine-size-adjusted indicators, enhances both predictive accuracy and interpretability. Experimental results demonstrate that the proposed ensemble consistently outperforms individual models in terms of MAE, RMSE, and \(R^2\) . Temporal validation on an independent multi-year dataset (2015–2024) confirms stable performance under distributional shift. Feature importance analysis highlights engine size, fuel consumption patterns, and vehicle class as dominant factors influencing CO \(_2\) emissions.