Improvement of carbon price prediction with social factors and mixed-frequency unstructured data

Carbon emissions trading right is one of core financial instruments for climate risk management, and carbon trading price prediction can be an effective tool. China's carbon emissions trading prices fluctuate greatly on account of many exogenous factors. However, (i) scholars just consider one or several factors, and few take the comprehensive influencing factors into consideration in study; (ii) very little attention has been paid to employ the unstructured data to research the impact of policies on the carbon market; (iii) This sample frequency approach not only result in factors information loss but cannot depict non-stationary, non-parameters, complexed nonlinear and multi-frequency features of the time series of variables in carbon market. The introduction of social factor data, effective processing of high-dimensional data, and prediction methods improvement are still in great need. To address these gaps, with the examples of carbon emission rights trading prices in China's markets, this paper proposes a mixed frequency method with unstructured data for carbon price prediction: (1) First, we intrude the social policy of sentiment index and Baidu search index as influencing factors, and effectively deal with such unstructured data. The results show that they act as important factors in four carbon trading markets. (2) Second, by screening key indicators and combining the mixed frequency dynamic factor model (MF-DFM), we explore the overall influence of the typical external factors on carbon prices and forecast the volatility of carbon prices. (3) Third, deep neural network (DNN) models are added to predict diligently the residual, and the effect of deep learning model and method on improving the forecasting accuracy is discussed. Our study has further improved the factor system that affects carbon trading prices and proposed more predictive methods, which can have certain reference value for carbon emission trading management.