Document Type : Research Paper
Authors
1
Department of Business Management, Faculty of Management, Islamic Azad University, Firoozkooh, Iran.
2
Department of Management, Financial Management, Azad University, Tehran, Iran.
3
Department of Business Management, Faculty of Management, Kermanshah Branch, Islamic Azad University, Kermanshah, Iran.
4
Management Faculty, Islamic Azad University, Branch of Urmia, Urmia, Iran.
Abstract
Retail investors play a pivotal role in shaping financial market trends, yet accurately forecasting
their behavior remains a complex challenge. Traditional models often fall short in capturing the temporal
dynamics and evolving relationships inherent in investor behavior. In this paper, we introduce a novel frame-
work based on Dynamic Graph Neural Networks (Dynamic GNNs) to predict retail investor actions with high
accuracy and interpretability.
Our approach constructs evolving graph representations of interactions between investors and assets over
time, integrating both psychometric attributes (e.g., risk tolerance, decision-making tendencies) and sentiment
signals derived from news and social media analysis. This fusion enables a comprehensive view of investor
behavior in changing market contexts. We evaluate our model on a large-scale dataset of real-world retail
investor transactions from brokerage platforms and compare its performance against a variety of benchmarks,
including static GNNs, traditional machine learning models (XGBoost, Random Forest), and dynamic base-
lines (e.g., RNNs, Temporal Graph Networks). Experimental results demonstrate that our Dynamic GNN
framework achieves 12% higher accuracy, 15% improvement in precision, and 10% better recall over existing
static GNN methods. Furthermore, it outperforms traditional dynamic methods by 8% in accuracy, thanks to its ability to capture fine-grained temporal patterns and incorporate rich investor-level features. However, scalability challenges arise when processing very large graphs, necessitating efficient sampling strategies.
This research contributes to the advancement of behavioral finance by offering a robust, scalable, and
interpretable method for modeling investor behavior. The proposed framework can support applications in algorithmic trading, risk management, and personalized financial advising, helping financial institutions better understand and serve retail investors
Keywords
Main Subjects
)