Nowcasting carbon emissions in a data-rich environment: a comparison of dynamic factor models and machine learning algorithms

Abstract This study investigates how the well-documented link between economic activity and carbon dioxide emissions, identified in previous research in energy and environmental economics, can be used to improve nowcasting carbon dioxide emissions in a data-rich environment. We compare classic and recent improvements in dynamic factor models with linear and nonlinear machine learning algorithms that have been shown to be effective in previous research. These machine learning algorithms are implemented in the mixed data sampling framework. The recent improvements in dynamic factor models include the use of nondifferenced data, which has increased prediction accuracy, especially during economically volatile periods. Additionally, there are structurally augmented dynamic factor models, which combine machine learning methods with dynamic factor models. For the structurally augmented models, we use machine learning algorithms to select the most important variables, which then augment the dynamic factor models. Our findings indicate that dynamic factor models outperform alternative approaches for carbon dioxide emission nowcasting. Specifically, models based on nondifferenced data demonstrate superior predictive ability with principal component extraction, whereas models using differenced data yield better results with Kalman filter extraction. These findings are essential for developing effective nowcasting models that enable timely emission assessments, which are critical for advancing ambitious climate policies. Our research, therefore, contributes to the ongoing discourse on the challenges of sustainable development by employing econometric models for the dynamic links between economic activities and environmental outcomes.