Harnessing firm similarities with graph neural networks for superior stock predictions

Tein Baaijens; Çiçek Güven; Gonzalo Nápoles

Harnessing firm similarities with graph neural networks for superior stock predictions

Research output: Contribution to conference › Paper › Scientific › peer-review

Abstract

This study explores the enhancement of Graph Neural Networks
(GNNs) for stock return prediction through a novel approach of constructing
graph networks based on firm-level characteristic similarities. Traditional
methods often fail to capture dynamic relationships between firms. By using
measures such as Cosine Similarity and Euclidean Distance to create graph
edges, this research empirically demonstrates a significant improvement
in predictive accuracy, especially when using the Cosine Similarity-based
Graph Attention Network. The results indicate that this approach not only
outperforms conventional models but also provides a deeper understanding
of stock return dynamics by effectively integrating firm characteristics into
the graph structure. The implementation is available on GitHub.

Original language	English
Publication status	Published - 11 Apr 2025
Event	13 th International Conference on Complex Networks & Their Applications - istanbul, Turkey Duration: 10 Dec 2024 → 12 Dec 2024

Conference

Conference	13 th International Conference on Complex Networks & Their Applications
Abbreviated title	COMPLEX NETWORKS 2024
Country/Territory	Turkey
City	istanbul
Period	10/12/24 → 12/12/24

Keywords

Financial Networks
Graph Neural Networks
Stock Return Prediction
Firm Characteristics

Cite this

@conference{f5c1f648cd07474fb3d94aa95c7e1629,

title = "Harnessing firm similarities with graph neural networks for superior stock predictions",

abstract = "This study explores the enhancement of Graph Neural Networks (GNNs) for stock return prediction through a novel approach of constructing graph networks based on firm-level characteristic similarities. Traditional methods often fail to capture dynamic relationships between firms. By using measures such as Cosine Similarity and Euclidean Distance to create graph edges, this research empirically demonstrates a significant improvement in predictive accuracy, especially when using the Cosine Similarity-based Graph Attention Network. The results indicate that this approach not only outperforms conventional models but also provides a deeper understanding of stock return dynamics by effectively integrating firm characteristics into the graph structure. The implementation is available on GitHub.",

keywords = "Financial Networks, Graph Neural Networks, Stock Return Prediction, Firm Characteristics",

author = "Tein Baaijens and {\c C}i{\c c}ek G{\"u}ven and Gonzalo N{\'a}poles",

year = "2025",

month = apr,

day = "11",

language = "English",

note = "13 th International Conference on Complex Networks \& Their Applications, COMPLEX NETWORKS 2024 ; Conference date: 10-12-2024 Through 12-12-2024",

}

TY - CONF

T1 - Harnessing firm similarities with graph neural networks for superior stock predictions

AU - Baaijens, Tein

AU - Güven, Çiçek

AU - Nápoles, Gonzalo

PY - 2025/4/11

Y1 - 2025/4/11

N2 - This study explores the enhancement of Graph Neural Networks (GNNs) for stock return prediction through a novel approach of constructing graph networks based on firm-level characteristic similarities. Traditional methods often fail to capture dynamic relationships between firms. By using measures such as Cosine Similarity and Euclidean Distance to create graph edges, this research empirically demonstrates a significant improvement in predictive accuracy, especially when using the Cosine Similarity-based Graph Attention Network. The results indicate that this approach not only outperforms conventional models but also provides a deeper understanding of stock return dynamics by effectively integrating firm characteristics into the graph structure. The implementation is available on GitHub.

AB - This study explores the enhancement of Graph Neural Networks (GNNs) for stock return prediction through a novel approach of constructing graph networks based on firm-level characteristic similarities. Traditional methods often fail to capture dynamic relationships between firms. By using measures such as Cosine Similarity and Euclidean Distance to create graph edges, this research empirically demonstrates a significant improvement in predictive accuracy, especially when using the Cosine Similarity-based Graph Attention Network. The results indicate that this approach not only outperforms conventional models but also provides a deeper understanding of stock return dynamics by effectively integrating firm characteristics into the graph structure. The implementation is available on GitHub.

KW - Financial Networks

KW - Graph Neural Networks

KW - Stock Return Prediction

KW - Firm Characteristics

UR - https://link.springer.com/chapter/10.1007/978-3-031-82427-2_19

M3 - Paper

T2 - 13 th International Conference on Complex Networks & Their Applications

Y2 - 10 December 2024 through 12 December 2024

ER -

Harnessing firm similarities with graph neural networks for superior stock predictions

Abstract

Conference

Keywords

Other files and links

Fingerprint

Cite this