A framework for stock selection via concept-oriented attention representation in hypergraph neural network

Abstract

Stock selection is an integral aspect of quantitative trading. In contemporary stock markets, where traders can easily exchange information, stock prices tend to exhibit high correlations due to their common characteristics. In recent years, several methods have been proposed to reveal the relationships among stocks using the attributes of the companies to which the stocks belongs. However, these studies do not delve deeply into the concepts behind stocks, as they frequently assume stationarity and ignore the impact weights of stock-to-concept associations. Our study addresses these limitations by introducing a novel hypergraph-based framework for stock selection. The distinctive contribution of this framework is its representation of concepts using hypergraph attention. Stock relationships were initially classified into predefined and hidden concepts. The acquisition of predefined relationships from wikidata was optimized, and hidden concepts were extracted to compile a new similarity attention hypergraph network framework for stock selection. The experimental results on stock data from two markets over a seven-year period demonstrate the superiority of our framework in terms of investment returns compared with baseline methods.