Lightweight AI-Driven Stress Testing for Small and Medium Financial Institutions: A Variational Autoencoder Approach with Extreme Value Theory for Macroeconomic Scenario Generation

Abstract

Small and medium financial institutions remain critically underserved by existing stress testing methodologies, which demand computational resources and specialized expertise far exceeding their operational capacity. This paper introduces a novel lightweight AI-driven framework that uniquely integrates Variational Autoencoders (VAE) with Extreme Value Theory (EVT) to generate highly realistic macroeconomic stress scenarios for rigorous capital adequacy assessment. The proposed "CCAR-Lite" methodology decisively overcomes the computational and expertise barriers that have historically excluded smaller institutions from conducting CCAR-style stress tests. By combining advanced dimensionality reduction with principled tail risk modeling, the framework captures complex nonlinear dependencies across multiple asset classes while achieving computational efficiency suitable for standard desktop hardware. Extensive experimental validation on 408 months of macroeconomic data (1990–2023) demonstrates that the framework generates stress scenarios with a plausibility score of 0.89 and tail realism of 0.87, statistically indistinguishable from historical crisis distributions (Kolmogorov-Smirnov p > 0.05). A practical use case applying the framework to a hypothetical community bank’s commercial real estate portfolio confirms its immediate applicability for real-world capital planning. This research delivers a transformative, open-source solution for democratizing advanced stress testing across the financial sector, directly strengthening systemic financial stability and regulatory compliance nationwide.