关键词: 组合预测, 碳价格, 混合分解集成, 非结构性数据, 主成分分析

Abstract: Carbon trading, as an effective mechanism for controlling carbon emissions, has generated a growing body of interest towards its trading price. Forecasting carbon trading price accurately is of vital significance. Not only can it assist investors in making informed decisions, but also it can aid governments in formulating scientifical carbon trading policies. However, the majority of the previous research on carbon trading price prediction generally builds prediction models based on historical data, and thus the forecasting results often exhibit significant hysteresis. Meanwhile, due to the complexity of carbon trading systems, time series of carbon trading price are endowed with characteristics such as non-linearity and high noise. Consequently, the promoting effect of a single decomposition integration method on forecasting accuracy is relatively limited. In sum, to enhance the forecast accuracy of the carbon trading price, it is of vital necessity to combine alternative source of information as well as various decomposition integration methods to construct the forecast model.
This research posits an unstructured data driven carbon price combined forecast model based on mixed decomposition and integration to forecast carbon trading price. First, relevant carbon trading unstructured data are obtained through Baidu Index and six corresponding main components are extracted by using principal component analysis (PCA). Second, empirical modal decomposition (EMD), variable modal decomposition (VMD) and wavelet transform (WT) are carried out respectively on the gained components and carbon trading price historical data, whose high frequency sequences, low frequency sequences and trend items are obtained after reconstruction. Then, based on the adaptive method, the self-regression integral sliding average model (ARIMA), Holt exponential smoothing method and artificial neural network (ANN) are selected to predict the high frequency sequence, low frequency sequence and trend item in combination with non-structural information. Furthermore, the predicted values from various forecasting methods such as ARIMA are integrated based on method of BP neural network. Meanwhile, the obtained forecasting results of the high frequency sequences, low frequency sequences and trend items are summed up to acquire the forecasting results of the carbon trading price decomposed by one of these three decomposition methods such as EMD. Eventually, the predicted values under these three decomposition methods are hierarchically integrated based on method of BP neural network again, and the final prediction results are obtained.
In an effort to test the predicting accuracy of the aforementioned model, we conduct an empirical analysis based on carbon trading price data from the carbon trading market in Hubei Province, China. Moreover, to further confirm the effectiveness of the proposed model, we compare its forecasting accuracy with seven other predicting models. Meanwhile, the prediction accuracy of these models is measured using five error evaluation metrics, including MAE, SSE, MSE, MSPE, and MAPE. Eventually, the result demonstrates that the values of the error indicators of our model are considerably smaller than those of the other models, indicating that the adoption of unstructured data, hybrid decomposition-integration method, as well as the combined forecast method can markedly improve the prediction accuracy of carbon trading price.
In summary, the proposed unstructured data driven carbon price combine forecast model based on mixed decomposition and integration provides with a new direction for constructing predicting models of carbon trading price as well as offers references for similar research. Based on this model, future research can further incorporate other sources of information such as structured data concerning carbon trading when predicting carbon trading price.

Key words: combined forecast, carbon price, mixed decomposition-integration, unstructured data, PCA