Financial Stress Tests – Generative Machine Learning

Stress testing is a simulation exercise to determine ability of a financial institution to deal with an adverse economic environment. European union bank stress test performed by EBA, DFAST (Dodd Frank) and CCAR performed for US Bank Holding Companies (BHC’s) are few of the major regular stress tests. Typically an economic scenario is a combination of macro (GDP, unemployement) and micro (stock prices, bond prices) risk factors; while the losses or impact on a financial institution is measured using capital ratio, leverage ratio etc. This poses a unique challenge for the regulator to design a realistic economic scenario with an adverse impact on financial institutions. Our research draws inspiration from AI (Artificial Intelligence) methods employed at creating life-like images to instead learn to create a realistic but unseen economic environments. More precisely an environment that is very likely to highlight potential weaknesses of the banks without the knowledge of exact assets and portfolios held by the bank.

Federal reserve governor Daniel Tarullo said that: ”we certainly do not want them (banks) to construct their portfolios in an effort to game the model”. We realizing that banks will employ even more sophsticated AI that allows them to take high levels of risk while performing well (“gaming”) on such tests. We attempt to find optimal strategy for regulators to still achieve their desired goals faced with this adverserial AI vs AI game against a more resourceful entity.