Research On Stock Price Movement Prediction Via Stock Market Macro Information

Stock movement prediction is to predict the price trend of stocks in the future,which has fundamental applications in quantitative trading and investment decision making.The dissemination of information in the stock market has a huge impact on stock prices.However,due to the complexity of the stock market structure and the diversity of information,stock price changes show strong volatility,which makes stock trend forecasting extremely difficult.Most of the existing research treat stocks independent of each other,or uses graphs to model complex high-order relationships in the stock market,lacking the considerations of hierarchical and dynamic interaction between stocks and stocks,stocks and industries,and industries and industries.Therefore,in response to the existing problem,we model the complex relationship in the stock market and conduct a stock price movement prediction study.First,we design a network architecture which can capture market macro trend information and propose a dynamic macro memory network(DMMN).The model can process and analyze the financial time series data of multiple stocks at the same time,and dynamically extract market macro information from them.When iteratively processing the features of each time step,the long-term dependence of each stock and the whole market is simultaneously captured with the help of the gate unit.Finally,the macro information of the market is dynamically integrated into the micro representation of each stock,and the stock price movement predication is made by combining the micro information of each stock and the macro information of the whole market.Because DMMN does not consider the hierarchy and complexity of the market when capturing market macro information,we propose a Macro Hierarchical Memory Network(MHMN)to address this kind of problem.Considering that the stock market is a very complex system with various of industries,and the industries will also affect each other,we follow the "stock-industry-market" hierarchy when extracting macro market information.First,we extract all industries’ information at each time step,then generate market representations for each stock according to different weights from all industries’ information under the consideration of different industries have different importance for different stocks.The market representation containing market information is then incorporated into the micro representation of each stock.This model not only takes into account the interaction between industries and industries,industries and stocks,but also enhances the model’s ability to perceive the local environment of the market.Finally,we collected the CSI300 data set on a public data interface and conducted a lot of experiments on it.In addition to using the F1 value to evaluate the classification effect,considering its practical application value,we also constructed a set of investment portfolios on the test set according to the prediction results of the model,and analyzed the actual profitability of the model by evaluating the Sharpe ratio of the portfolios.The experimental results show that in the classification task of stock price movement prediction,our proposed DMMN model has good prediction performance,and the Sharpe ratio also reflects the higher practicality of DMMN.In addition,compared with DMMN,the f1 value and Sharpe ratio of our proposed MHMN are further improved by 2.96% and19.56%,respectively.