Multiple stocks recommendation: a spatio-temporal hypergraph learning approach

Stocks do not exist in isolation in the financial markets, but there are complex relationships among them, which leads to the fluctuations in the stock prices have the properties of synchronism, systematic linkage and conductivity. From the relevancy-based perspective, a group of stocks with distinct degrees of correlation constitute the spatial structure of hypergraph, which combines with temporal information together to provide the important basis for the analysis of financial market behaviors. However, how to effectively learn the intrinsic relevancies of stocks is still an open problem. In this article, we proposed the Multiple Stock Recommendation system based on a novel Spatio-Temporal Hypergraph Learning framework (MSR-STHL). Firstly, inspired by the existing works, LSTM-attention module is applied to learn the temporal features of stocks, where Hawkes process is involved to enhance the attention mechanism in the long-term time scale. Secondly, by means of the prior knowledge and data-driven methods, the relevancy-based spatial structures among stocks are modeled from several aspects, where data-driven way can provide the potential relations among stocks and make up the disadvantage of untimely change of prior knowledge. Thirdly, graph attention networks and hypergraph convolution operations are used to achieve the fusion learning of multiple graphs/hypergraphs. Finally, the recommended stocks on the return prediction are provided. The proposed model is evaluated on three stock market datasets, i.e. NASDAQ, NYSE, and China A-share, respectively. In comparison with state-of-the-art methods, the proposed model can outperform the existing methods and the validity is confirmed.