Research On Stock Market Investment Prediction Based On Hypergraph Attention Network

The stock market is an important part of the national economy.With the rapid development of economy,quantitative investment is gradually paid more and more attention by investors.Quantitative investment uses quantitative methods and computer programs to analyze data and issue trading orders based on rules in order to obtain considerable and stable returns.For the stock market,quantitative investment mainly includes timing,stock selection,asset allocation,risk control and so on.Benefiting from the powerful ability of deep neural networks,especially recurrent neural networks capabilities of the long-term dependence of time-series data,the performance of stock market investment prediction methods based on deep learning has been significantly improved compared with those based on traditional technical indicators.However,it usually models only temporal correlation,treating stocks as independent of each other.With the development of the stock market,there are extensive connections between stocks,which contains valuable signals that are conducive to trend prediction,and these relationships tend to be group-wise rather than pair-wise.On the other hand,the return on investment based on the prediction model depends largely on its accuracy.However,due to the highly nonlinear,random and chaotic nature of the stock market,it is not easy to accurately predict stock prices or trends.In addition,how to convert price prediction into trading signals is still a problem to be solved.Therefore,this paper studies the stock market investment prediction based on the hypergraph attention network,mainly aiming at quantitative timing and stock selection strategies.The former solves when to buy or sell stocks,that is,predicts the future ups and downs of stocks,and the latter solves which stocks are worth paying attention to or holding,namely the allocation of stock portfolio.At the same time,the two parts of work are accompanied by risk control management,and the main work done is as follows:(1)Predicting the future price trend of stocks is a challenging quantitative timing problem.Traditional stock prediction methods are mostly based on time-series model,ignoring the complex group-wise relationship between stocks.To this end,this paper proposes a temporal-relational collaborative modeling framework for end-to-end stock trend prediction.First,a recurrent neural network combined with the attention mechanism is used to capture the temporal dynamics of stocks.Then,different from existing studies relying on the pairwise correlations between stocks,a hypergraph structure is introduced to jointly describe the group-wise relationship of industry-belonging and fund-holding.Based on this,a novel hypergraph tri-attention network is proposed to augment the hypergraph convolution with a hierarchical organization of intra-hyperedge,inter-hyperedge,and inter-hypergraph attention modules.In this manner,the model can adaptively distinguish the importance of nodes,hyperedges,and hypergraphs during the information propagation among stocks,so that the potential synergies between stock movements can be fully exploited.Extensive experiments on real-world data demonstrate the effectiveness of our approach,and investment simulation back-testing show that our approach can achieve a more desirable risk-adjusted return.(2)Stock portfolio selection is a basic financial planning task,namely quantitative stock selection problem,which can dynamically reallocate assets to selected stocks to achieve goals such as maximum profits and minimum risks.This paper proposes a hypergraph-based reinforcement learning method for stock portfolio selection,in which the basic problem is to learn a policy function to generate the optimal portfolio vector under the current environment and guide the appropriate trading action.The policy network first captures the historical time-series pattern of stocks,and then,unlike the previous work of implicitly modeling the pairwise correlation of stock assets,we propose a hypergraph attention module,which utilizes the hypergraph structure to explicitly model the group-wise industry-belonging relationships among stocks.Experiments show that this method has significant advantages in portfolio selection compared with existing methods based on online learning and reinforcement learning.