TL;DR
This paper introduces a novel hybrid model, Lasso-COA-XGBoost-CEEMDAN, for dynamic carbon price prediction. By combining adaptive feature selection, an optimized gradient boosting model, and multi-frequency decomposition of residuals, the model significantly improves prediction accuracy and stability, offering a more robust tool for carbon market analysis and policy-making.
What Problem Does This Paper Address?
The global push towards net-zero emissions necessitates robust carbon markets and accurate carbon price forecasting. China, in particular, is rapidly developing its carbon market, but current price prediction models often suffer from low accuracy and instability. This is partly due to the inherent complexity and non-stationarity of carbon price data, which is influenced by a multitude of factors. Existing methods often struggle to capture these dynamics effectively, leading to unreliable forecasts that can hinder market efficiency and emission reduction efforts. This paper aims to address these limitations by developing a more sophisticated and accurate prediction model.
Methods
The proposed Lasso-COA-XGBoost-CEEMDAN model is a multi-stage approach:
- Feature Selection (Adaptive Lasso): The model begins by employing an adaptive Lasso method to identify the most influential factors from a set of 15 primary indicators. This step is crucial for reducing model complexity and mitigating multicollinearity.
- Optimized Prediction (COA-XGBoost): The selected features are then fed into an XGBoost model. The parameters of XGBoost are optimized using the Cheetah Optimization Algorithm (COA), a nature-inspired meta-heuristic algorithm known for its efficiency and ability to handle complex optimization problems. This COA-XGBoost model serves as the initial predictor.
- Residual Decomposition and Correction (CEEMDAN): The residuals from the COA-XGBoost model are further processed using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) method. CEEMDAN decomposes the residuals into multiple intrinsic mode functions (IMFs), representing different frequency components.
- Multi-Frequency Prediction: Appropriate predictive models (COA-XGBoost for high-frequency components and SVR for low-frequency components) are then applied to these decomposed residuals. This allows for tailored prediction of different signal characteristics.
- Reconstruction: Finally, the predictions from the decomposed residuals are combined with the initial COA-XGBoost prediction to produce the final, more accurate, carbon price forecast. A rolling time window is also incorporated to ensure real-time adaptation to market changes.
Datasets
The study utilizes daily carbon price data from:
- Hubei Province carbon market
- Guangzhou carbon market
- EU carbon market (EUA futures settlement price)
These datasets span from January 1, 2018, to March 31, 2024, for Hubei and Guangzhou, and January 1, 2018, to March 31, 2024, for the EU market. The data is sourced from the CHOICE Financial Terminal database.
Key Insights and Ideas
- Decomposition is Key for Residuals: Simply predicting residuals with a single model is insufficient. Decomposing residuals into high- and low-frequency components allows for specialized modeling, significantly improving error correction.
- COA Enhances XGBoost: Optimizing XGBoost with the COA algorithm provides a substantial boost in prediction accuracy compared to using XGBoost alone. This highlights the value of meta-heuristic optimization for complex models.
- Feature Selection Matters: Identifying and using only the most relevant influencing factors (via Lasso) not only simplifies the model but also improves its predictive power and avoids issues like multicollinearity.
- Multi-Frequency Modeling Adapts Better: Different frequency components of the carbon price signal have distinct characteristics. Using tailored models for high-frequency (volatile) and low-frequency (smoother) residuals leads to more accurate overall predictions.
- Baidu Index as a Proxy for Sentiment: Incorporating emerging media indicators like the Baidu index provides valuable insights into market sentiment and expectations, which are often overlooked but crucial for accurate forecasting.
Key Results
The proposed Lasso-COA-XGBoost-CEEMDAN model demonstrated superior performance across various metrics:
- Hubei Province: Compared to the XGBoost model, the hybrid model achieved improvements of 99.9987% in RMSE, 99.9039% in MAE, and 99.9960% in MAPE. The R² was improved by 0.2004%.
- Guangzhou Province: Similar significant improvements were observed.
- EU Carbon Market: The model achieved an MSE of 1.2653e–07, MAE of 0.0003, MAPE of 0.000006, and R² of 0.9999, outperforming other benchmark models.
- Crude Oil Price Forecasting: The model also showed excellent predictive performance in forecasting Brent crude oil prices, with MSE = 0.1553, MAE = 0.3544, MAPE = 0.0058, and R² = 0.9997.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Model Complexity: The COA algorithm, while effective, is complex to configure, requiring significant time and computational resources. Future work should focus on model streamlining and automating parameter settings to enhance convenience and efficiency.
- Single-Step Forecast: The current study performs only single-step forecasts, neglecting a comprehensive analysis of future trends. Future research should explore multi-step forecasting.
- Time-Varying Factors: The time-varying characteristics of influencing factors were not fully considered. Future work should investigate how these dynamics impact carbon prices.
- Quadratic Decomposition: Exploring quadratic decomposition in residual processing could further reveal deep nonlinear relationships and improve accuracy.
- Broader Applications: The developed method could be extended to other related fields, such as energy market prediction and climate change risk management, to verify its generalizability.
Summary updated