Temporal Trend and Fluctuation Learning via Enhanced Attention Mamba for Carbon Price Interval Forecasting

Area of Science:

• Environmental Economics
• Data Science
• Financial Forecasting

Background:

• Carbon trading markets are complex and require accurate price forecasting for stability.
• Existing methods fail to capture nonlinear and non-stationary price dynamics.
• Long-term forecasting struggles with the inherent volatility of carbon prices.

Purpose of the Study:

• To propose a novel Temporal Trend and Fluctuation Learning (TTFL) model for interval-valued carbon price forecasting.
• To address the limitations of existing methods in capturing carbon price characteristics.
• To provide a practical tool for stakeholders in carbon markets.

Main Methods:

• Wavelet decomposition to separate price trend and fluctuation learning.
• Enhanced Mamba architecture with state space models for trend learning.
• Attention mechanism with Mamba for fluctuation learning.
• Interval-valued recovery loss function for training stability.

Main Results:

• TTFL model demonstrated superior prediction accuracy and robustness on real-world data.
• The model effectively captures long-term trends and short-term volatility.
• Comparative experiments confirmed TTFL's outperformance against baseline methods.

Conclusions:

• The TTFL model offers a novel forecasting paradigm integrating collaborative learning and selective state space modeling.
• The approach provides actionable insights for policymaking and investment strategies.
• TTFL enhances navigation of complex carbon market environments.