Carbon Price Forecasting with LLM-Based Refinement and Transfer-Learning

Haiqi Jiang, Ying Ding, Rui Chen, Chenyou Fan

Abstract: We propose a unified forecasting framework for accurately predicting carbon markets of EU Emission Trading Scheme (EU ETS) and Chinese Emission Allowance (CEA). Our framework utilizes a Time-Series Model (TSM) for initial prediction followed by applying a Large Language Model (LLM) to refine the forecasts. We prompt the LLM to refine the TSM forecasts by demonstrating an example pair of past TSM predictions and their corresponding true future prices to the LLM as a chain-of-thought. The in-context learning capacity of the LLM allows the LLM to rectify inaccurate predictions to reflect on TSM predictions and refine the forecasts. To further reduce the prompting delays and expenses involving LLMs, we innovate a post-finetuning approach to train a Gated Linear Unit (GLU) model to condense the LLM’s in-context learning capability. This enables direct fine-tuning of TSM outputs without the need for explicit prompting LLM during inference. Experimental results show that our method can refine the TSM prediction by 10% to 40% in various zones, as well as enhance transfer learning by 10% to 21% through the inclusion of market context of the source zone when predicting the target zone. Remarkably, our GLU model achieves comparable, and in some cases superior, performance compared to LLM prompting. It effectively combines the short-term forecasting capability of classical Time Series Models with the long-term trend prediction ability typically associated with the LLMs.