Multivariate events enhanced pre-trained large language model for carbon price forecasting

Carbon prices are affected by multiple factors, exhibiting high volatility and nonlinear complexity, making their accurate prediction still a challenging task. However, existing models are limited in scale and struggle to effectively extract complex features from high-dimensional data. Recently, pre-trained large language models have demonstrated significant performance advantages in a variety of tasks, but their application in the field of carbon price prediction remains limited. Therefore, this study proposes a multivariate event-enhanced pre-trained large language model, which is constructed based on pre-trained large language models and thus does not require large data for fine-tuning. Specifically, this study first adopts the least absolute shrinkage and selection operator method to identify the key variables that have the most significant impact on carbon price. Meanwhile, unstructured events information is quantified through text sentiment analysis techniques, thus forming a complete model input dataset. Subsequently, these datasets are sequentially processed through the input module, the frozen large language model and the output module to finally generate the prediction results. Carbon price datasets from Hubei and Shanghai carbon markets are used as study cases. The experimental results show that the proposed multivariate events enhanced pre-trained large language model exhibits significant advantages in both prediction accuracy and robustness compared with other existing artificial models.