Generative AI for Stress Testing: Scenario Fabrication and Model Risk Governance in Capital Markets

Abstract

This review examines the emerging application of Generative Artificial Intelligence (GenAI) in fabricating synthetic scenarios for trading stress testing and model risk management. The central question is whether large language models (LLMs) can generate coherent, data-consistent “what-if” macroeconomic and market shock narratives that propagate across linked asset classes to support pre-trade risk assessment and algorithmic robustness testing. Existing research and practice are synthesized around three pillars: (i) prompting LLMs with macroeconomic triggers to produce narrative-driven stress scenarios; (ii) calibrating generated outputs against historical covariance structures and factor dependencies to ensure internal consistency and statistical plausibility; and (iii) integrating calibrated scenarios into pricing engines and execution simulators to evaluate portfolio profit-and-loss (PnL) distributions and tail-risk behavior.

The review highlights key methodological considerations, including transparency, reproducibility, explainability, and governance in the generation and validation of synthetic stress libraries. Contributions include a structured framework for embedding GenAI-driven scenario fabrication within CCAR-style regulatory regimes and internal model validation workflows, as well as for pre-deployment sign-off of trading algorithms. By aligning generative scenario fabrication with established stress-testing practices, the paper argues that GenAI can enhance the breadth and responsiveness of stress testing in capital markets, provided that governance, auditability, and model risk safeguards are rigorously maintained.