A Method to Find Diverse and Manageable Sets of Plausible Yet Severe Financial Scenarios

Abstract

We introduce a new practical data-intensive method to generate/discover consistent finite representative collections of plausible yet severe macroprudential, microprudential, book-specific, and individual obligor/instrument scenarios. These scenarios are conditioned on current information (including current macroeconomic, index, industry and instrument/obligor-specific information), and can be conditioned on partial future scenario specifications as well (to accommodate regulatory stress testing requirements, for example, the CCAR requirements for banks, the projections of economists, or senior management). Our method is scalable, is designed to work with limited training data, can incorporate the fat-tailed and mutually dependent behavior that is characteristic of many financial quantities, and can reflect model misspecification risk.