A Stock Prediction Model based on LightGCN

Stock prediction aims to predict future prices and trends of stocks in order to help investors make good investment decisions and get more profits. An advanced stock prediction model is Relational Stock Ranking (RSR), which has a great improvement in the cumulative investment return ratio (IRR) than previous studies. However, in RSR model, the relation between stocks is learned with the one-layer graph learning method, which can't comprehensively describe the dependency between different stocks. In this paper, we propose a new model for stock prediction, named L-RSR, which can effectively predict the stock ranking and help investors select stocks with higher return ratios. L- RSR contains a graph learning technique based on LightGCN with multiple layers, leading to a better extraction of the relation features. We conduct a series of experiments with the data from National Association of Securities Dealers Automated Quotations (NASDAQ) market. Experimental results show that L-RSR outperforms RSR with more than 25% improvements in IRR.