Improving Carbon Price Prediction by Decomposing Single Influencing Factor with XGBoost

Carbon credits are devised as an effective mechanism to limit global carbon dioxide emissions in the world. These credits permits the stakeholders to release a specified quantiy of carbon dioxide or other greenhouse gases. Thus, accurate carbon price forecasting is essential for stakeholder decision-making and policy makers' reference. However, the relatively small sample size of carbon trading price data has significantly hindered the improvement of the prediction models, in particularly with the deep learning algorithm. In contrast, we observe that by further incorporating the domain knowledge of carbon credits, we can enhance both the performance and efficiency with traditional models. In this context, a carbon price prediction model combining adaptive noise complete set Empirical Mode decomposition (CEEMDAN) and XGBoost model is proposed in this work. We first leverage the CEEMDAN algorithm to decompose the raw carbon trading price data into multiple sub-sequences with low volatility and low randomness, and then uses these decomposed data as input for prediction model training. We have conducted extensive experiments, specifically using real-world carbon trading market data, to verify the universality and superiority of the proposed model. The results demonstrate state-of-the-art performance in comparison with other works. We anticipate this work will contribute to the fields of carbon credit engineering and prediction.