Integrated GCN-LSTM stock prices movement prediction based on knowledge-incorporated graphs construction

Stock prices movement prediction has been a longstanding research topic. Many studies have introduced several kinds of external information like relations of stocks, combined with internal information of trading characteristics to promote forecasting. Different from previous cases, this article proposes a reasonable assumption that major fluctuations of stock prices are mainly triggered by high-volume transactions which usually occur on a group of stocks that share some common features (e.g., stocks in the same industry, region, concept or yield similar volatility), and further develops an integrated GCN-LSTM method to achieve more precise predictions from the perspective of modelling capital flows. First, we construct four kinds of graphs incorporating various relational knowledge (edge) and utilize graph convolutional network (GCN) to extract stock (node) embeddings in multiple time-periods. Then, the obtained temporal sequences of stock embeddings are put into long short-term memory recurrent neural network (LSTM) to discriminate the moving direction of prices. Extensive experiments on major Chinese stock indexes have demonstrated the effectiveness of our model with best accuracy of 57.81% acquired, which is much better than baselines. Moreover, experimental results of GCN-LSTM under different graphs and various node embedding dimensions have been compared and analyzed, indicating the selection of key parameters to achieve optimal performances. Our research findings provide an improved model to forecast stock prices movement directions with a reliable theoretical interpretation, and in depth exhibit insights for further applications of graph neural networks and graph data in business analytics, quantitative finance, and risk management decision-makings.