An Interpretable Mid-Term Carbon Price Prediction Method Based on Multi-Markets Temporal Fusion Transformer

Abstract

Accurate carbon price prediction can assist enterprises in enhancing their cost management and planning strategies. Plenty of models have been used to predict short-term carbon price both for national and pilot carbon markets. However, the interpretability of these models is weak, and the connections among different carbon markets are commonly overlooked. Additionally, in China, the mid-term trends of carbon price are also very important, yet there is scant research in this area. To address these drawbacks, this study proposes an interpretable mid-term carbon price prediction method. This study utilizes the Temporal Fusion Transformer to develop one multivariate prediction model that integrates data from multiple pilot carbon markets in China to predict their carbon price for the upcoming month. Additionally, a known future variable is incorporated into the model to achieve more accurate predictions. Utilizing data from the Guangdong, Hubei, and Shenzhen carbon markets in China, we demonstrate that the proposed method indeed improves prediction accuracy. To enhance the model's interpretability, this study analyzes the distance between the attention weight and its average value. Anomalous fluctuations in carbon price at the beginning of 2022 are identified, attributed to shifts in market sentiment driven by policy changes and events in both domestic and international carbon markets. The interpretability of the model is also enhanced by calculating the importance for all variables using the variable selection network. Our findings reveal that the three most critical variables for carbon price prediction are the past value of carbon price, domestic oil price, and carbon price of the European Union Allowance. This study innovatively applies explainable algorithms to mid-term carbon price prediction and employs one model that learns information from multiple markets, thereby enhancing prediction accuracy. This study can also serve as a reference for carbon price prediction in the national carbon markets of China and other countries.

Keywords: Carbon price, deep learning, Interpretability, Multivariate prediction