Carbon Market Price Forecasting Using a Bidirectional Temporal Convolution Exogenous-Enhanced Time-Series Model.

Carbon market prices are jointly shaped by policy interventions, energy market fluctuations, and macroeconomic dynamics, and thus exhibit pronounced nonlinearity, non-stationarity, localized abrupt changes, and time-varying uncertainty. From an information-theoretic perspective, carbon price forecasting can be viewed as the extraction and fusion of effective information from a complex market system driven by heterogeneous endogenous and exogenous signals. To address the challenges of accurately characterizing local high-frequency fluctuations in carbon price series, effectively modeling the interactions between endogenous and exogenous variables, and mitigating the structural noise introduced by conventional serial forecasting frameworks, this study proposes ConvTimeXer, a hybrid model combining bidirectional temporal convolution and TimeXer for carbon market price forecasting. Specifically, the model first employs front-end bidirectional temporal convolutions to extract local multi-scale fluctuation features from the endogenous carbon price series. It then leverages the global token and cross-attention mechanism in TimeXer to achieve dynamic decoupling and deep interaction between endogenous and exogenous variables. Finally, residual fusion of shallow and deep features is introduced to enhance the preservation of local details. Experimental results based on data from China's carbon market over the past three years demonstrate that the proposed framework delivers high predictive accuracy and strong robustness, effectively balancing responsiveness to local abrupt changes with global trend modeling. This study not only provides an effective approach for carbon price forecasting in complex and uncertain market environments, but also offers valuable insights into non-stationary time-series forecasting driven by multi-source heterogeneous information.