A Numerical-Based Attention Method for Stock Market Prediction With Dual Information

Abstract

Stock market prediction is of great importance for financial analysis. Traditionally, many studies only use the news or numerical data for the stock market prediction. In the recent years, in order to explore their complementary, some studies have been conducted to equally treat dual sources of information. However, numerical data often play a much more important role compared with the news. In addition, the existing simple combination cannot exploit their complementarity. In this paper, we propose a numerical-based attention (NBA) method for dual sources stock market prediction. Our major contributions are summarized as follows. First, we propose an attention-based method to effectively exploit the complementarity between news and numerical data in predicting the stock prices. The stock trend information hidden in the news is transformed into the importance distribution of numerical data. Consequently, the news is encoded to guide the selection of numerical data. Our method can effectively filter the noise and make full use of the trend information in news. Then, in order to evaluate our NBA model, we collect news corpus and numerical data to build three datasets from two sources: the China Security Index 300 (CSI300) and the Standard & Poor’s 500 (S&P500). Extensive experiments are conducted, showing that our NBA is superior to previous models in dual sources stock price prediction.