Using Generative AI for Robust Regulatory Stress Testing

Periodic stress testing of systemically significant financial institutions was introduced in the aftermath of the Great Financial Crisis (GFC) of 2007�2008 as part of Dodd-Frank Wall Street Reform and Consumer Protection Act of 2010 in the United States. Since its inception, the act has mandated regulatory agencies to conduct periodic stress testing to ensure systemically important financial institutions have the requisite capital to continue functioning as viable businesses during times of economic stress without jeopardizing the stability of the financial system. The tests envision a variety of hypothetical economic scenarios unfolding over the ensuing quarters and are designed to test the adequacy of capital reserves. Because they are hypothetical, they must be comprehensive enough to encompass a range of possible economically challenging scenarios and be realistic enough to capture the evolving correlations between macroeconomic variables that are expected to unfold in those scenarios. Manual design of these scenarios using historical data, exclusively or primarily, is hamstrung by the inherent limitations of historical experience, which may be inadequate to model unforeseen economic scenarios. To further compound the problem, correlations between macroeconomic variables may change and evolve in markedly different manner during those periods of economic malaise and a manual design of testing scenarios is likely to overlook those aspects of macroeconomic variable evolution. Generative artificial intelligence has the potential to confront these challenges by providing an automated tool for generating a range of economic scenarios, encompassing stressed economic scenarios that have not been witnessed in the past. The generated scenarios hew to the patterns of evolution of macroeconomic variables consistent with correlations observed in similar scenarios from the past. This work presents variational autoencoders coupled with LSTM (long-short-memory model)-based recurrent neural networks to generate new economic scenarios and presents a qualitative assessment of their consistency. It showcases the ability of variational auto-encoders and deep neural networks to generate novel and yet realistically evolving economic scenarios to enhance the robustness of stress testing framework. With the advent of new financial products such as crypto currencies and ever-evolving technologies such as blockchain, it is imperative for financial regulators to use automated tools for scenario generation to assure market participants, investors, and public about the continuing relevance of stress testing as reliable indicators of financial wellbeing of systemically important financial institutions