StockRanker: A Novelty Three-Stage Ranking Model Based on Deep Learning for Stock Selection

Abstract

The use of deep learning to identify stocks that will yield higher returns in the future and purchase them to achieve returns greater than the market average is an attractive proposition. However, in recent years, many studies have revealed two major challenges facing this task, including how to effectively extract features from historical stock price data that can be used for stock prediction and how to rank future stock returns using these features. To address these challenges, we propose StockRanker, an innovative three-stage ranking model for stock selection. In the first stage, we use autoencoder to extract features embedded in the historical stock price data through unsupervised learning. In the second stage, we construct a hypergraph that describes the relationships between stocks based on industry and market capitalization data and use hypergraph neural networks (HGNN) to enhance the features obtained in the first stage. In the third stage, we use a listwise ranking method to rank future stock returns based on the stock features obtained earlier. We conducted extensive experiments on real Chinese stock data, and the results showed that our model significantly outperformed baseline models in terms of investment returns and ranking performance.