Symmetry-Aware Graph Neural Approaches for Data-Efficient Return Prediction in International Financial Market Indices

Tae Kyoung Lee, Insu Choi, Woo Chang KimSymmetry 2025 · DOI

TL;DR

This paper introduces novel Graph Neural Network (GNN) approaches, specifically GCN and GAT, to predict international financial index returns. By modeling inter-market relationships as graphs, these GNNs effectively capture complex, symmetric dependencies, leading to superior performance, especially during volatile market conditions and financial crises, compared to traditional machine learning models.

What Problem Does This Paper Address?

The increasing interconnectedness of global financial markets makes it challenging to accurately forecast market movements. Traditional forecasting methods often treat financial indices in isolation, failing to account for the dynamic and often symmetric relationships (bidirectional spillover effects) that exist between them. This limitation becomes particularly pronounced during periods of market instability, where these interdependencies intensify. Existing research has also faced challenges with robust validation and demonstrating the effectiveness of GNNs in a global, multi-market context. This paper aims to bridge this gap by developing and rigorously evaluating GNN-based models for international financial index return prediction, focusing on their ability to capture these complex, symmetric relationships and their performance across different market regimes.

Methods

The core of this research lies in the application and adaptation of Graph Neural Networks (GNNs) for financial forecasting. Specifically, the authors employ:

  • Graph Convolutional Networks (GCNs): These models aggregate information from neighboring nodes in a graph to update a node's representation. They are effective in capturing persistent dependency structures within relatively stable markets.
  • Graph Attention Networks (GATs): GATs introduce an attention mechanism, allowing the model to assign different weights to neighbors, enabling it to focus on the most relevant inter-market connections. This makes GATs particularly adept at handling dynamic changes and emphasizing salient relationships, especially during volatile periods.

The graph structures themselves are constructed using a correlation-based network formation technique. Specifically, edges between nodes (representing financial indices) are established if their Pearson correlation coefficient exceeds a threshold of |ρ| ≥ 0.3. This thresholding aims to capture meaningful relationships while filtering out noise. The models are trained and evaluated using a rigorous 30-fold time-series cross-validation framework to ensure robustness and avoid overfitting.

These GNN approaches are benchmarked against traditional machine learning models: Random Forest (RF), XGBoost, Multi-Layer Perceptron (MLP), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM).

Datasets

The study utilizes daily closing prices and trading volume data for major global stock indices. The initial dataset comprised 41 indices, which were filtered down to 25 based on data quality. Ultimately, 15 representative indices were selected based on their cumulative trading volume and market capitalization. The data spans from January 1, 2004, to December 31, 2024. The closing prices were transformed into daily log return series for stationarity.

Key Insights and Ideas

  • Symmetry is Key: The paper strongly emphasizes that the symmetric nature of inter-market relationships is crucial for accurate financial forecasting, especially during crises. GNNs, by their graph-based structure, are inherently suited to capture these bidirectional influences.
  • GNNs Shine in Volatility: The most significant advantage of GNNs (both GCN and GAT) over traditional models is observed during periods of market instability, such as financial crises (2008 Global Financial Crisis, European Debt Crisis) and recovery phases (post-GFC, post-COVID-19).
  • Architecture Matters: GATs, with their attention mechanism, tend to outperform GCNs during highly volatile and crisis periods, suggesting their adaptability to rapidly changing market dynamics. GCNs, on the other hand, show more stable performance in less volatile markets.
  • Data Efficiency: The GNN approach relies solely on intrinsic market data (returns), reducing the need for external indicators that can introduce delays and biases.
  • Rigorous Validation: The use of 30-fold time-series cross-validation provides a robust framework for evaluating model performance, addressing a known limitation in prior GNN finance research.

Key Results

The study found that GNN models (GAT and GCN) consistently produced lower Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) values compared to traditional ML models, particularly during crisis and recovery periods. For instance, during the 2008 Global Financial Crisis, GAT significantly outperformed MLP and XGBoost. Similarly, during the COVID-19 crisis, GAT showed clear advantages in RMSE and MAE. While the performance gap narrowed during stable market periods, GNNs still demonstrated competitive results. The statistical significance of these improvements was rigorously tested using independent t-tests across 1% , 5%, and 10% significance levels.

Limitations and Future Work

The authors acknowledge several limitations:

  • Computational Complexity: GNNs can be computationally intensive, especially with increasing network size. Their advantage diminishes in stable markets where simpler models are often sufficient.
  • Graph Construction: The current graph construction relies on linear correlation, which might not capture more complex, nonlinear, or causal dependencies.
  • Interpretability: While GATs offer some transparency via attention weights, deep graph models generally lack intuitive explanations for their predictions, which is crucial for practical adoption.
  • Centralized Training: The current framework uses centralized training, which may face scalability limitations in real-time, multi-agent environments.

Future work suggested includes:

  • Developing advanced graph construction techniques (e.g., mutual information, Granger causality, transfer entropy) to better represent market dynamics.
  • Integrating explainable AI (XAI) techniques to enhance model interpretability.
  • Exploring distributed and quantized optimization techniques for more scalable and efficient training, particularly for real-time applications.
  • Investigating log-quantized GNN training for decentralized financial forecasting systems.

Summary updated

BibTeX
@article{lee2025symmetry,
  author = {Tae Kyoung Lee and Insu Choi and Woo Chang Kim},
  title = {{Symmetry-Aware Graph Neural Approaches for Data-Efficient Return Prediction in International Financial Market Indices}},
  journal = {Symmetry},
  year = {2025},
  doi = {10.3390/sym17091372},
}
Figure