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

Area of Science:

• Environmental Economics
• Financial Forecasting
• Data Science

Background:

• Carbon market prices are influenced by policy, energy markets, and macroeconomics, leading to complex, non-stationary dynamics.
• Accurate carbon price forecasting is challenging due to nonlinearity, abrupt changes, and time-varying uncertainty.
• Existing forecasting models struggle with local fluctuations and interactions between endogenous and exogenous variables.

Purpose of the Study:

• To develop an effective hybrid model for carbon market price forecasting.
• To address challenges in modeling local high-frequency fluctuations and variable interactions.
• To improve the accuracy and robustness of carbon price predictions in complex market systems.

Main Methods:

• Proposed ConvTimeXer, a hybrid model combining bidirectional temporal convolution and TimeXer.
• Utilized bidirectional temporal convolutions to extract multi-scale features from carbon price series.
• Employed TimeXer's global token and cross-attention for dynamic interaction modeling between endogenous and exogenous variables.
• Implemented residual fusion to preserve local details in forecasting.

Main Results:

• ConvTimeXer demonstrated high predictive accuracy and robustness on China's carbon market data.
• The model effectively balances responsiveness to local abrupt changes with global trend modeling.
• Experimental results confirm the framework's superiority over conventional forecasting methods.

Conclusions:

• ConvTimeXer offers an effective approach for carbon price forecasting in uncertain environments.
• The study provides valuable insights into non-stationary time-series forecasting with heterogeneous information.
• The hybrid model enhances understanding of complex market dynamics and improves predictive capabilities.