TL;DR
This paper introduces a novel carbon price forecasting model that leverages Lasso regression to intelligently select and weight multiple individual forecasting models. By doing so, it aims to overcome the limitations of arbitrarily choosing models and improve prediction accuracy for the volatile carbon market.
What Problem Does This Paper Address?
Accurately predicting carbon prices is crucial for effective carbon market operation, informing decision-makers, and driving emission reduction efforts. However, carbon prices are notoriously nonlinear and unstable, making them difficult to forecast. Existing "combination forecasting" methods often arbitrarily select a few models, which may not be optimal for all datasets due to the complex nature of carbon price data. This approach can lead to redundant information and suboptimal predictions. The challenge, therefore, is to systematically select the most relevant forecasting models and determine their optimal weights to create a robust and accurate combined forecast.
Methods
The core innovation of this paper lies in its two-stage approach:
- Model Screening with Lasso Regression: Instead of using all available forecasting models, the authors employ Lasso regression. Lasso is a statistical technique that performs feature selection by shrinking the coefficients of less important predictors to zero. In this context, it effectively "filters out" underperforming or redundant single forecasting models, identifying the most impactful ones for the combination. This addresses the issue of multicollinearity among the predictions of individual models.
- Optimal Combination Forecasting with IOWA: After selecting the most relevant models using Lasso, the paper utilizes the Induced Ordered Weighted Averaging (IOWA) operator for combination forecasting. This method goes beyond simple averaging by dynamically reordering and weighting the predictions based on their historical accuracy. This allows the model to adapt to changing market conditions and assign higher importance to models that have performed better recently.
The study screened a total of 11 single forecasting models, encompassing both traditional statistical methods (like ARIMA) and artificial intelligence models (like LSTM, GRNN, BPNN).
Datasets
The study uses two real-world carbon price datasets from Shenzhen, China, a significant pilot city for carbon trading:
- SZA-2017: Data from July 31, 2017, to March 9, 2022 (427 data points).
- SZA-2019: Data from October 29, 2019, to March 7, 2022 (315 data points).
For both datasets, the data was split into training and testing sets. The training sets were used to determine model parameters and regression coefficients, while the test sets were used to evaluate prediction accuracy.
Key Insights and Ideas
- Lasso as a Smart Filter: The application of Lasso regression for selecting forecasting models is a key novelty. It demonstrates that not all models contribute equally, and Lasso can effectively prune the less useful ones, leading to a more efficient and accurate combination.
- Dynamic Weighting is Crucial: Simple averaging of forecasts is suboptimal. The IOWA operator's ability to dynamically adjust weights based on past performance is essential for capturing the evolving dynamics of the carbon market.
- AI Models Aren't Always Superior: While AI models generally performed well, the study found that traditional econometric models can sometimes outperform them on specific metrics or datasets, highlighting the value of a hybrid approach.
- Heterogeneity of Model Performance: The Lasso screening results showed that the optimal set of models and their assigned weights varied significantly between the SZA-2017 and SZA-2019 datasets, underscoring the importance of data-specific model selection.
Key Results
The proposed Lasso-based combination forecasting model demonstrated superior performance compared to benchmark models across various evaluation metrics:
- Prediction Accuracy: The model achieved significantly lower error indicators (RMSE, SSE, MAE, MAPE) and a better U1 statistic on both SZA-2017 and SZA-2019 datasets compared to individual models and a simple average combination model. For instance, on SZA-2017, the proposed model's RMSE was 0.2011, significantly lower than many individual models.
- Generalization Ability: The model showed strong performance in terms of Index of Agreement (IA), Success Ratio (SR), and Accuracy, indicating good generalization capabilities. For SZA-2017, IA, SR, and Accuracy were 0.9674, 0.8621, and 0.9758, respectively.
- DM Test Validation: Diebold-Mariano (DM) tests confirmed the statistical significance of the proposed model's superiority over most benchmark models, including both classical econometric and artificial intelligence approaches.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Historical Data Reliance: The current model relies solely on historical carbon price data. Future work could explore incorporating unstructured data (e.g., news, policy announcements) to enhance prediction accuracy.
- Limited Decomposition/Integration: While the model combines multiple forecasting methods, it doesn't explicitly incorporate advanced decomposition or integration techniques that could further refine the forecasting process.
- Generalizability: The authors suggest applying the developed model to other domains like business cycles, economic growth, and stock price forecasting to test its broader applicability.
- Model Selection and Weighting: Further research could explore more sophisticated methods for model selection and dynamic weight adjustment within the combination framework.
Summary updated