关键词: 股票趋势预测, 宏观记忆网络, 动态依赖, 层次宏观信息, 门控单元

Abstract: Abstract: The complex structure of the stock market and the variety of information in it makes it extremely difficult to predict stock movement. A majority of existing research treats stocks independent of each other, or model the complex high-order relations in the stock market through graph, which lacks hierarchical and dynamic considerations of the interaction among stocks, industries and markets. The Dynamic Macro Memory Network (DMMN) which can forcast multiple stocks movement simultaneously was proposed to address this problem. The market macro-environmental information was modeled by the hierarchies of "stock-industry-market", and long-term dependencies on time series of it were captured. Then the market macro-environmental information was integrated into stock micro-characteristic information dynamically to enhance the ability of each stock to perceive the macro state of the market and capture the interdependence among stocks, industries, and markets indirectly. Experimental results on the collected CSI300 dataset show that, compared with ALSTM, GCN-LSTM, CNN and other models, DMMN achieves better results in F1-score and Sharpe Ratio, and improved by 4.88% and 31.93% in these two indicators respectively compared with ALSTM, which is the best model among all comparison methods. This shows that DMMN has higher prediction performance and better practicability.

Key words: stock movement prediction, macro memory network, dynamic dependence, hierarchical macro information, gate unit