Carbon prices forecasting based on the singular spectrum analysis, feature selection, and deep learning: Toward a unified view

An accurate carbon price prediction is vital for governments to formulate emission reduction policies and for corporate managers to carry out sustainable management. Although many models have been proposed to meet this purpose, there is still a lack of comprehensive comparisons among the key competing models in a single study. This study conducted a comprehensive evaluation of six feature selection methods and six machine learning models for carbon price prediction, and then proposed a novel data-driven hybrid model that integrates singular spectrum analysis, random forest, and long short-term memory neural network. The results show that random forest and long short-term memory neural networks are the most competitive feature selection and prediction models, respectively. It is more reasonable to construct the model's input by comprehensively considering variables' short-, medium-, and long-term features. When the recombination value is approximately half the length of the embedding window, the singular spectrum analysis is most conducive to improving the prediction accuracy. The hybrid model proposed is always significantly superior to other benchmark models. Our work contributes to guiding carbon price modeling to help regulators and managers accurately grasp carbon price signals.