Scholarlink

<p>This study examines the application of Graph Neural Networks (GNNs) for monthly U.S. stock return prediction, comparing conventional approaches with a novel method that constructs networks from firm-level characteristic similarity. The research tests whether these similarity-based networks enhance GNN performance relative to networks built on historical return correlations. Findings indicate that the Cosine-Similarity-based Graph Attention Network (CS_GAT) significantly outperforms both traditional benchmark methods and Pearson-correlation-based models over the out-of-sample period 1999–2022. Additionally, we show that cosine-similarity-based models deliver superior risk-adjusted returns in a long machine-learning portfolio. Our evidence indicates that these gains are driven by the pairwise-similarity models’ emphasis on non-momentum characteristics and their ability to capture fundamentals distinct from correlation-based counterparts. The implementation is available on <a href="https://github.com/Baaijens/GNNs-Using-Similarity-In-Firm-Level-Characteristics">https://github.com/Baaijens/GNNs-Using-Similarity-In-Firm-Level-Characteristics</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Graph learning on financial networks from pairwise similarity of firm-level characteristics

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月25日

查看全文
  • Tein Baaijens,
  • Çiçek Güven,
  • Gonzalo Nápoles

摘要

This study examines the application of Graph Neural Networks (GNNs) for monthly U.S. stock return prediction, comparing conventional approaches with a novel method that constructs networks from firm-level characteristic similarity. The research tests whether these similarity-based networks enhance GNN performance relative to networks built on historical return correlations. Findings indicate that the Cosine-Similarity-based Graph Attention Network (CS_GAT) significantly outperforms both traditional benchmark methods and Pearson-correlation-based models over the out-of-sample period 1999–2022. Additionally, we show that cosine-similarity-based models deliver superior risk-adjusted returns in a long machine-learning portfolio. Our evidence indicates that these gains are driven by the pairwise-similarity models’ emphasis on non-momentum characteristics and their ability to capture fundamentals distinct from correlation-based counterparts. The implementation is available on https://github.com/Baaijens/GNNs-Using-Similarity-In-Firm-Level-Characteristics.