Diagnosing systematic overestimation and feature dependencies in data-driven predictions of urban CO₂ emissions in China

Machine learning is increasingly used to simulate and predict complex urban phenomena, such as CO₂ emission patterns in cities. However, the effectiveness, biases, and uncertainty of these predictive models are not well understood. This study systematically evaluates 18 different machine learning algorithms using a rich dataset of county-level CO₂ emissions in China from 2001 to 2020, incorporating 41 diverse features spanning socioeconomic, environmental, and energy consumption factors. The analysis reveals a widespread issue: nearly all models, regardless of type, tend to systematically overestimate emissions in Chinese county-level cities across both space and time. By isolating the effects of feature selection and quantifying uncertainty through bootstrap resampling, the research demonstrates that aggressive dimensionality reduction often harms model performance, while retaining a high number of features is crucial for capturing the true complexity of urban emissions. Tree-based models, such as CatBoost and LightGBM, outperform others in accuracy and uncertainty quantification. These findings highlight the urgent need for the scientific and policy communities to prioritize diagnosing and correcting systematic biases in machine learning models, rather than focusing solely on improving accuracy, to ensure reliable guidance for urban carbon management and climate policy.