Forecasting Carbon Price In China Based On Mixed-frequency Unstructured Data

With the unrestrained consumption of fossil energy in human production and life,excessive emission of greenhouse gases leads to environmental problems such as greenhouse effect,extreme weather and ecosystem disorder.In order to deal with the climate problem,the world has started carbon emission trading system one after another to achieve the goal of energy saving and emission reduction.China has also started to build carbon trading markets on a pilot basis in 2011,and after nearly ten years of exploration launched the national online trading of carbon emission rights,becoming the largest carbon market in the world.However,the development of the domestic carbon market has not yet been perfected.The price of carbon emission rights is affected by energy,macroeconomic,policy,climate and other factors,and has great ups and downs.Then,the effect of emission reduction is limited,which is not beneficial to the achievement of carbon peak and carbon neutrality goals.Thus the research of carbon trading price provides reference for the government to issue corresponding policies and enterprises to plan emission reduction work.It is of great significance to the benign development of carbon market and the mitigation of greenhouse effect.In this paper,four representative carbon trading pilot markets in Shenzhen,Beijing,Guangdong and Hubei and the national carbon trading market are taken as the research object.And the paper organizes the system of influencing factors of carbon trading price involving five aspects,such as similar substitution,energy industry,macroeconomy,social intervention,and climate environment,from perspectives of homogeneous and mixed frequency,structured and unstructured,domestic and foreign.In the aspect of social intervention,unstructured indexes such as sentiment index and Baidu search index were introduced to reflect the state of policy public opinion.After that,LASSO model was used to screen variables and further obtain the set of indicators that are more closely related to each carbon market.In the prediction,considering the limitation of the econometric model to identify nonlinear information,this paper combines the dynamic factor model of mixed-frequency with XGBoost and LSTM models,which are better at capturing nonlinear information.The combination models can make full use of mixing information and predict timely and accurately.Finally,the RMSE between the combined model and competing models such as ARIMA and homogeneous models was tested by DM from four perspectives,including data type,data frequency,single model and combined model,to compare model prediction effects.The empirical results of this paper show that:(1)In the operation of the market,the carbon emission trading price is vulnerable to external influences.In terms of the development stage of the carbon market,macro-economy,climate environment and social intervention factors occupy the main position of influencing factors in the initial stage of the carbon market.But when the market develops to the mature stage,similar substitution and energy industry factors have prominent influence on markets and gradually occupy a dominant position.From the perspective of the categories of influencing factors,the direction and degree of influence of indicators in different markets are not completely consistent.Among them,similar alternative products,influential coal and crude oil markets at home and abroad,macroeconomic factors represented by domestic and foreign stock indexes,exchange rates and industrial indicators,Sentiment reflecting policy opinions and climate and environmental factors such as temperature play a more important role.(2)In the empirical prediction,the prediction performance of the MF-DFM-LSTM model based on mixed-frequency unstructured data is significantly better than other models,and has better stability and generalization ability.After the introduction of mixed-frequency unstructured data,the predictive effectiveness of each model in the carbon market improved significantly,especially in some immature,volatile and irregular carbon markets.Compared to other models,MF-DFM-LSTM model captures and utilizes a large amount of information that other models fail to identify.It has the smallest RMSE,MAE and MAPE evaluation indexes in the five carbon markets,and significantly outperforms other models in terms of synchronization,accuracy,sensitivity and volatility.In this paper,the impact of multiple dimensions outside the market is fully considered when examining the price of carbon emission trading.It innovatively introduces the mixed-frequency unstructured data,and combines the dynamic factor model of mixed-frequency with the machine learning method for empirical prediction.The research methods and conclusions can provide some references for the relevant academic research in the field of carbon market,and also assist the government and enterprises in making relevant decisions,so as to promote the development of China’s carbon trading market.