[Paper Review] Dynamic graph neural networks for enhanced volatility prediction in financial markets

Pulikandala Nithish Kumar, Nneka Umeorah+1 more|arXiv (Cornell University)|Oct 22, 2024

Stock Market Forecasting MethodsDecision Sciences3 citations

TL;DR

This paper proposes a Temporal Graph Attention Network (Temporal GAT) that models global financial markets as dynamic graphs to improve volatility prediction. By integrating volatility spillover indices and attention mechanisms, it captures non-linear, time-varying dependencies between indices, significantly outperforming GARCH and other ML models in short- to mid-term forecasts with lower MAPE and MSE.

ABSTRACT

Volatility forecasting is essential for risk management and decision-making in financial markets. Traditional models like Generalized Autoregressive Conditional Heteroskedasticity (GARCH) effectively capture volatility clustering but often fail to model complex, non-linear interdependencies between multiple indices. This paper proposes a novel approach using Graph Neural Networks (GNNs) to represent global financial markets as dynamic graphs. The Temporal Graph Attention Network (Temporal GAT) combines Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs) to capture the temporal and structural dynamics of volatility spillovers. By utilizing correlation-based and volatility spillover indices, the Temporal GAT constructs directed graphs that enhance the accuracy of volatility predictions. Empirical results from a 15-year study of eight major global indices show that the Temporal GAT outperforms traditional GARCH models and other machine learning methods, particularly in short- to mid-term forecasts. The sensitivity and scenario-based analysis over a range of parameters and hyperparameters further demonstrate the significance of the proposed technique. Hence, this work highlights the potential of GNNs in modeling complex market behaviors, providing valuable insights for financial analysts and investors.

Motivation & Objective

• To address the limitations of traditional GARCH models in capturing complex, non-linear interdependencies between global financial indices.
• To model financial markets as dynamic graphs where indices are nodes and their volatility relationships are weighted, directed edges.
• To improve short- and mid-term volatility forecasting accuracy by leveraging temporal and structural dynamics through a novel GNN architecture.
• To evaluate the impact of different graph construction methods—correlation-based vs. volatility spillover indices—on model performance.
• To assess model robustness under varying market conditions, including periods of high volatility such as the COVID-19 pandemic.

Proposed method

• Constructs dynamic, directed graphs using volatility spillover indices and correlation matrices to represent interdependencies between 8 major global indices.
• Employs a Temporal Graph Attention Network (Temporal GAT) that combines Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs) to model both local and global node relationships over time.
• Uses attention mechanisms to dynamically weight the importance of neighboring indices in volatility prediction, adapting to changing market conditions.
• Trains the model on 15 years of daily closing prices from major global indices, with volatility estimated using high-frequency realized measures.
• Applies hyperparameter tuning and cross-validation to optimize model performance across different forecast horizons (1, 5, 10, 22 days).
• Conducts leave-one-out sensitivity analysis to evaluate the contribution of individual indices to overall predictive accuracy.

Experimental results

Research questions

• RQ1 Can a dynamic graph neural network architecture outperform traditional GARCH models in forecasting short- and mid-term volatility?

Key findings

• The Temporal GAT model achieved significantly lower Mean Absolute Percentage Error (MAPE) than GARCH, GNN, and other ML baselines, particularly in short- to mid-term forecasts (1–10 day horizons).
• The model maintained robust performance across all horizons, with MAPE stabilizing between 13% and 30% for long-term forecasts (5, 10, 22 days), though some indices like GDAXI and FCHI showed higher error during volatile periods.
• Volatility spillover indices outperformed correlation-based graph construction, capturing shock propagation more effectively and improving predictive accuracy.
• The S&P 500 and DAX were identified as key drivers of market volatility, as their removal drastically reduced model performance, while indices like HANG SENG added noise due to weaker spillover effects.
• During the COVID-19 pandemic, the Temporal GAT maintained superior accuracy compared to traditional models, demonstrating resilience under extreme market stress.
• The model exhibited higher prediction errors during high-volatility regimes due to sudden shocks and increased uncertainty, indicating that performance is sensitive to market regime shifts.