Title: Artificial Intelligence Across Disciplines: Research, Innovation, and Intelligent Solutions
Editors: Dr. Subita Bhagat, Dr. A. Balamurugan, Dr. P. Krishna Kumar, and Mrs. S. Nandhini Devi
ISBN: 978-81-69857-83-3
Chapter: 12
DOI: https://doi.org/10.59646/815/12
Author: S. Bharathi
Abstract
Modern financial architectures are characterized by high-frequency interdependencies, non-linear spillover dynamics, and structural topological shifts that render traditional econometric models inadequate during periods of systemic market distress. This chapter develops a unified temporal graph deep learning framework for predictive modeling of systemic risk and macroeconomic tail vulnerabilities across global capital markets. By combining dynamic spatial-temporal graph attention networks (ST-GAT) with extreme value theory (EVT) and multi-task learning objectives, our architecture simultaneously models bilateral institutional counterparty exposures, common asset-holding overlaps, and macroeconomic state vectors. Evaluated on a cross-market panel encompassing 120 global systemically important financial institutions and macroeconomic sovereign indicators over a multi-decade span, the framework generates multi-horizon conditional value-at-risk (CoVaR) and structural default cascade forecasts. The model achieves superior out-of-sample tail calibration and mitigates systemic contagion estimation errors compared to classical parametric baselines, providing regulators with an auditable computational engine for macroprudential stress testing and real-time systemic surveillance.
Keywords: Systemic Risk Prediction, Graph Attention Networks, Extreme Value Theory, Conditional Value-at-Risk, Macroeconomic Stress Testing, Financial Contagion