Forecasting carbon prices in China's pilot carbon market: A multi-source information approach with conditional generative adversarial networks
Zhigang Huang,Weilan Zhang
DOI: https://doi.org/10.1016/j.jenvman.2024.120967
IF: 8.7
2024-05-10
Journal of Environmental Management
Abstract:In recent years, the Chinese government has actively pursued the implementation of its 'dual-carbon' strategy, concurrently establishing a national carbon emissions trading market. Accurate carbon price forecasts have become essential for policymakers and investors involved in related initiatives. Nevertheless, influenced by the interaction of various information sources, carbon trading prices exhibit non-linear and non-stationary characteristics, posing challenges for accurate prediction. Current research, centered around deep learning models, predominantly emphasizes intricate network structures, optimisation algorithms, and data decomposition. However, these models face a developmental bottleneck in extracting carbon price features and efficiently leveraging multi-source information. Consequently, novel ideas and methodologies are imperative. This study focuses on the Hubei and Guangdong regional carbon markets as research subjects. It develops a prediction framework based on a generative adversarial network model to capture the time-series change characteristics of carbon trading prices and the condition matrix. First, a generator prediction model is used to obtain the input matrix features and extract the time series features through a complex network to predict the carbon price data at the next moment using a fully connected layer. Second, a discriminator is utilised to distinguish between the actual values and the predicted values. The generator and the discriminator undergo continuous iterative training and alternate optimisation. This process aims to bring the generated prediction distributions closer to the actual sample data, resulting in more accurate final predictions. The empirical results convincingly show that the proposed model achieves unparalleled forecasting precision in both markets. The proposed model demonstrates the lowest MAE (0.804 and 0.839), lowest MAPE (0.023 and 0.018), lowest RMSE (1.174 and 1.383), and highest R 2 (0.971 and 0.989) across both markets, indicating superior predictive accuracy. Additionally, the proposed model consistently outshines traditional forecasting approaches across one-step, five-step, and ten-step forecasts, affirming its robustness and universal applicability in modelling carbon trading price series. The findings suggest that this study can aid policymakers in optimizing the carbon pricing system. Furthermore, it offers a reference for policymakers to comprehensively leverage external factors, such as regulating traditional energy prices, leveraging international carbon market experiences, and monitoring economic dynamics. This comprehensive strategy can streamline the exploration and management of carbon price fluctuations, ultimately strengthening the carbon market's risk control system.
environmental sciences