Sentiment‐Driven Forecasting of Carbon Prices: A Hybrid Neural Network Approach Based on BiGRU‐Inception‐Attention

Carbon trading forecasting faces challenges from strong nonlinearity, multiscale fluctuation, and the influence of market sentiment, which traditional models struggle to capture. To address these issues, this paper proposes a sentiment‐driven BiGRU‐Inception‐Attention hybrid neural network framework for carbon price forecasting. In the first step, we use the BERT‐BiLSTM model to construct a carbon news sentiment index. We then apply the XGBoost model to rank feature importance and select the eight most important features from 11 factors. Consequently, we use the BiGRU model to capture temporal dependencies of carbon prices, the Inception model to extract multiscale features, and the Attention mechanism to dynamically weight key features. Using data of China's Hubei carbon market from 2020 to 2024, the proposed approach significantly outperforms both traditional benchmarks and hybrid models without sentiment, achieving an MSE of 0.332 and R2 of 0.839. The results highlight the importance of sentiment in carbon price formation and provide a practical reference for improving carbon price forecasting and market design.