A novel forecasting framework leveraging large language model and machine learning for methanol price

As an essential low-carbon clean energy source, methanol price fluctuations significantly impact global chemical industry chains and energy markets. This study develops a novel predictive model framework integrating a large language model with traditional machine learning techniques to enhance methanol price prediction accuracy. Utilizing the open-source LLaMA3 model, we develop a time series forecasting model (MPP-GPT) by comprehensively integrating 3.67 million time series data across 33 domains. We employ the Maximum Information Coefficient algorithm to identify and extract over 160,000 indicator data points highly correlated with Chinese methanol price from a non-public commercial database, enabling model parameter fine-tuning. To enhance short-term price fluctuation capture, we introduce a rolling-window-based Relevance Vector Machine model to optimize residual predictions of MPP-GPT forecast results. For model interpretability, we establish an indicator system analyzing macro and micro factors' impact on Chinese methanol price from both long-term and short-term perspectives. Empirical analysis demonstrates that the MPP-GPT + RVM hybrid model significantly outperforms existing models' accuracy and interpretability. The Mean Absolute Error, Root Mean Square Error, and Mean Absolute Percentage Error of the MPP-GPT + RVM model are 0.0996, 0.1083, and 3.96%, respectively. The conclusions provide valuable insights for methanol producers, coal suppliers, and policymakers while offering new perspectives for analyzing and forecasting energy product prices.