Adversarial Robustness and Stress Testing for Stock Prediction: Defending Graph-Based Models Against Market Manipulations and Extreme Events

Abstract

The deployment of deep learning models in stock prediction systems creates new vulnerabilities that adversaries can exploit for profit or market disruption. Beyond conventional machine learning threats, stock prediction models face adversarial risks specific to financial markets—sophisticated actors can intentionally manipulate prices, create fictitious trading patterns, and exploit model sensitivities to generate profitable signals or cause systematic losses. The Stock State Space Graph model introduced by Lu, Hu, and Zhang achieves state-of-the-art prediction accuracy through graph-based relational reasoning, yet like all neural network models, it is potentially susceptible to adversarial perturbations and extreme market conditions that fall outside the training distribution. This paper proposes a comprehensive framework for adversarial robustness and stress testing of graph-based stock prediction models. Our approach, Robust Stock State Space Graph (R-S3G), integrates adversarial training with worst-case optimization to defend against price manipulation attacks, extreme value theory-based stress testing to evaluate model behavior under market crises, and robustness verification techniques to bound model behavior under adversarial conditions. Through extensive experiments on benchmark datasets and simulated attack scenarios, we demonstrate that R-S3G substantially improves robustness to adversarial perturbations while maintaining competitive prediction accuracy under normal market conditions. Our work contributes to the growing field of secure and trustworthy AI in finance, providing principled methods for defending prediction systems against adversarial manipulation and extreme events.