Predicting Price-Limit-Hitting Stocks with Hierarchical Graph Neural Network

Abstract

In most stock markets, stock prices are not allowed to rise above a daily limit (called price limit). In order to make overnight profits, some investors buy price-limit-hitting stocks with their price limits and sell them in the next trading day. But it is high risk because the price-limit-hitting stocks might close with a lower price than their price limits. Fortunately, we found that the price-limit-hitting stocks will rise in the next trading day with high ratio if they close with their price limits. Therefore, it is of meaningful to predict whether a price-limit-hitting stock will close with its price limit (Type I) or not (Type II). In this paper, we propose a Hierarchical Graph Neural Network (HGNN) for predicting price-limit-hitting stock classification once a stock touches its price limit. In HGNN framework, we construct stock relation graph, and fuse stock information hierarchically from different views including node view, relation view and graph view, which takes historical sequence feature and stock relation into consideration. Extensive experimental results show that our method achieves high classification accuracy, and high return ratio on two real datasets.