HGNN: Hierarchical graph neural network for predicting the classification of price-limit-hitting stocks

Abstract

In some stock markets, stock prices are not allowed to rise above a daily limit to restrain the surge of price (called price limit). When the price limit occurs, investors tend to chase the continuing upward momentum for profit-making. However, For the stocks that hit daily price limit, we observe whether they close at daily price limit will lead to the opposite price trends of the next trading day. Therefore, this work aims to predict whether a stock that hits its daily price limit will also close at the same price level (i.e., Type I or Type II). The occurrence of price limit is driven by different levels of market state. For example, it can result from macro-economic changes of the whole market, or it can be traced to some industry-specific factors. A challenging task is to learn a better stock representation with less uncertainty by comprehensively considering the hierarchical property of market state. Accordingly, we design a novel hierarchical architecture, called Hierarchical Graph Neural Network (HGNN), to investigate the market state at hierarchical view for stock type prediction. In HGNN, we construct the stock relation graph and merge stock information hierarchically extracted from multiple views of market state, including node view, relation view and graph view, which takes both historical sequence pattern and stock relation into consideration. Our key innovation is the introduction of hierarchical structure makes the predictive model able to more comprehensively infer the hierarchical property of market state. Further, it also provides the deeper insight for the actual investment practice. To validate the effectiveness of our method, we conduct back-testing on the two-year historical data of more than 2500 main-board stocks in two China stock markets, SSE and SZSE. To support further study of the stock type prediction task, we have published two long-range stock datasets (Datasets are available at https://drive.google.com/file/d/1TXiAyqt3rHveuzdGT6YtswU1e-tBSFUe/view?usp=sharing). Extensive experiments show that our method outperforms the state-of-the-art solutions including ALSTM, GCN and GAT with the improvements of at least 3.54% on average in accuracy. In addition, the average return ratio of SSE and SZSE has improved by 18.57% and 8.75%, respectively.