Stock index forecasting based on multivariate empirical mode decomposition and temporal convolutional networks

Modeling and forecasting of stock indexes is an important and challenging work in the field of financial research, which is of great significance for investors to reduce risk on investment and improve investment returns. However, the basic data of the stock index includes five indicators, namely, opening price, highest price, lowest price, closing price and trading volume (COHLV), each of which contains some information related to the future trend. And these indicators are affected by different factors such as politics, economy, psychology and so on, so they have nonlinear, high noise and other complex characteristics. These reasons make the existing methods cannot effectively improve the accuracy of stock index forecasting. In order to solve this problem, we propose a hybrid stock index forecasting model named MEMD-TCN, which is based on multivariate empirical mode decomposition (MEMD) and temporal convolutional networks (TCN). The novelty of the proposed model focuses on decomposition of multivariate time series and using only fully convolutional layers to forecast stock index. The proposed model first uses MEMD to decompose the data of COHLV to obtain subsequences of different fluctuation frequencies of each time series, then inputs subsequences of the same frequency into TCN to predict the subsequences of closing price in the next period. Ultimately, the forecasted values for the closing price are obtained by reconstructing the prediction results of all the subsequences of closing price. To evaluate the performance of the proposed MEMD-TCN model, stock index of several countries that can reflect the overall changes of the market are used. The experimental results verify the notable effectiveness and necessity of the decomposition of multivariate time series. Meanwhile, the results demonstrate the significant superiority of the proposed MEMD-TCN model on accuracy and stability.