Deep learning for financial banking stress test analytics

Since the recent financial crisis of late 2008, several global regulatory authorities have collaboratively mandated stress-testing exercises. These exercises evaluate the potential capital shortfalls & systemic impacts on large banks in hypothetical adverse economic scenarios, which try to simulate the macro-economic conditions similar to recent crisis'. The ability to relate dynamic economic conditions with banking performance profiles to identify meaningful relationships could provide significant insights for bank capital & loss projections. In this dissertation, the practical challenges that face bank stress-test analytics are examined and approached using advanced analytical techniques. Initially, (1) through a rigorous examination of an economic condition estimator (ECE), which learns joint approximation representations among exogenous factors by analyzing the complex non-linear relational combinations among the real-world economic indicators using a multi-modal conditioned variational auto-encoder (MCVAE). Experimentation on real-world economic conditions from the U.S. regulatory stress test exercise (CCAR) over the last three decades demonstrates the model's effectiveness. Additionally, (2) a focused study on bank capital & loss prediction (BCLP) methodology that can incorporate economic conditions as an estimated variable while also considering dynamic variability of potential crisis profiles that better provide a robust prediction of capital & loss. Demonstrations through experiments show that the BCLP model outperforms baseline & state-of-the-art methods from literature when evaluated on a sample of 1000 U.S. bank holding companies' historical consolidated financial statements (FR-9YC) from the past three decades. Both the ECE & BCLP model frameworks together form the Integrated Multi-modal Bank Stress Test Predictor (IMBSTP) framework to provide a data-driven end to end bank stress testing analytical tool. Lastly, (3) a preliminary overview of the Transferable Knowledge for the Bank Capital Components (TKBCC) model framework is discussed. The framework assumes that banks inherently share hidden intrinsic qualities and leverages inductive transfer learning techniques to improve bank capital-components predictions for domain tasks with limited training data. The performance of preliminary experiments on the proposed model framework through consolidated financial statements from the China Stock Market Accounting Research Database (CSMAR), and the Wharton Research and Data Service's (WRDS) repositories from the last two decades demonstrate the utility of the TKBCC model framework.Ph.D.Includes bibliographical reference