Carbon Market Volatility and Its Response Mechanism Based on CNN-BiLSTM Model and Signal Decomposition

Abstract

The carbon market volatility series exhibits strong nonlinearity, non-stationarity, and noise interference, which limits the accuracy of both traditional econometric models and standalone deep learning approaches. To address this, a hybrid prediction framework is developed. Firstly, an improved bald eagle search algorithm is adopted to optimize the hyperparameters of variational mode decomposition (VMD). Through chaos mapping initialization, dynamic inertia weight adjustment, and Gaussian mutation mechanism, population diversity and global search capability are enhanced, and the optimal number of VMD and penalty factor combination are obtained. Secondly, VMD is utilized to decompose the volatility series into several intrinsic modal components. Finally, bidirectional long short-term memory neural networks, convolutional neural networks, and attention mechanisms are introduced to achieve high-precision prediction of volatility series. The results using the EU Emissions Trading System as an empirical object show that traditional generalized autoregressive conditional heteroskedasticity (GARCH) models have limited performance in predicting carbon futures volatility. The optimal GARCH model has an R² of 92.6779% and an average absolute percentage error of 9.8358%. The suggested model outperforms all traditional econometric models significantly in four evaluation indicators: The average absolute error is reduced by 71.19%, the root mean square error is reduced by 70.13%, the average absolute percentage error is reduced by 63.80%, and the R² is enhanced by 6.11%. The designed framework provides a scalable technical path for predicting carbon market volatility and has practical application value for pricing carbon financial derivatives and market risk management.