摘要
摘要
碳排放权是一种新型经济环境资源，对我国企业的生产、管理和能耗均存在
影响，而碳排放权价格的高频波动是碳交易市场的关键问题。建立碳排放权价格
的预测模型可以为广大投资者提供合理科学的碳交易决策工具，引导我国企业交
易者更好地参与碳交易市场，促进碳交易市场的高效发展。
目前我国的碳交易市场尚不完善，碳排放权的价格具有高频波动、非线性、
非平稳等特征，因此，使用传统的计量建模方法来拟合碳排放权价格的曲线难度
颇高，且不全面。而模态分解方法能够深入探究不同频率下碳排放权价格的规律，
从而降低噪音，更好地掌握市场价格波动的内在特征。因此，本文采用变分模态
分解（Variational Mode Decomposition，VMD）方法，将上海试点市场的碳排放
权（SHEA）现货价格分解为不同频率的多层模态分量，并选取前 N层进行重构，
以减少噪声序列，并利用群智优化算法对 VMD的两项参数进行优化。考虑到现
有研究均采用单一的广义自回归条件异方差模型（ Generalized Autoregressive
Conditional Heteroskedasticity，GARCH）模型或深度学习方法预测碳价格序列，
预测精度不高，而混合深度学习模型能更好地拟合价格的线性和非线性特征。因
此，本文采用混合深度学习方法，结合计量经济学模型和深度学习算法的优点，
对 SHEA碳排放权价格进行滚动预测。实证结果表明，动态的指数条件异方差-
门控循环单元结构模型（ Exponential GARCH - Gate Recurrent Unit，EGARCH-
GRU）模型在测试集上的 5个损失函数整体小于其他模型，说明动态
EGARCH-GRU模型对 SHEA碳排放权价格的预测效果最好，而变分模态分解对
碳价降噪的效果不明显，预测精度并未显著提高。同时，基于最优预测模型的结
果，生成择时信号，构建碳排放权购买策略，利用测试集数据回测，并构建相应
的随机购买策略进行策略有效性评价，结果表明，在 1000个随机策略中，基于
最优预测模型的碳排放权购买策略优于 99.6%的随机策略，对碳交易企业具有一
定的启发作用。
关键词：碳金融；VMD；GARCH族；深度学习；购买策略
I

Abstract
Abstract
Volatile prices of carbon emissions can affect the production and operation of the
enterprises and the development of carbon trading market. Meanwhile, the price of
carbon emission rights is the key issue of carbon trading market. Price prediction
models of carbon emission rights can offer scientific decisions for investors, and
guide the investors to better use of carbon trading market, promoting the development
of carbon market.
Due to the carbon price in Shanghai pilot market is unstable, multi-frequency,
nonlinear and other irregular characteristics, it is difficult to use traditional
econometric models to comprehensively describe the characteristics of carbon price
fluctuation. However, carbon price decomposition can deeply explore various inherent
pattern of carbon price in different frequencies as well as better grasp the rule of
carbon price fluctuation. Therefore, this paper uses Variational Mode Decomposition
(VMD) method to decompose the spot price of Shanghai carbon emission permit
(SHEA) into multi-layer modal components with different frequencies and takes the
first N layers for reconstruction, so as to reduce the noise sequence. The two
parameters of VMD are optimized by using swarm intelligence optimization
algorithm. Considering that existing researches have used single GARCH model or
deep learning method to predict the carbon price series, and the prediction accuracy is
not satisfing, while the hybrid deep learning model can better fit the nonlinear
characteristics of the price. Therefore, this paper uses the hybrid deep learning method,
combining the advantages of econometric model and deep learning algorithm, to
predict the carbon price series of Shanghai pilot market. The empirical results show
that the five loss functions of the dynamic EGARCH-GRU model on the test set are
all better than those of other models. It indicates that the dynamic EGARCH-GRU
model has the best prediction performance for the carbon emission price in Shanghai
pilot market, while the effect of variational mode decomposition on the carbon price
noise reduction is not obvious for the prediction accuracy. At the same time, based on
the results of the optimal prediction model, this paper generates a timing signal, and
designs a kind of purchasing strategy of carbon emission rights, which uses the
prediction data back to the test and constructs other strategies to compare. The results
II

Abstract
show that, based on the optimal prediction model of carbon emission rights, the
purchasing strategy is better than the 99.6% of the random strategy. In conclusion, this
paper has certain enlightenment for carbon trading enterprises.
Key words: Carbon Finance; VMD; GARCH Models; Deep Learning; Purchasing
Strategy
III
。。。以下略