Essays on the application of machine learning in finance

Abstract

This dissertation examines how financial institutions can evaluate and govern artificial intelligence (AI) models when predictive gains may introduce additional model risk. Three essays combine empirical analysis with governance research. The first compares econometric and machine-learning volatility forecasts across market regimes. Although Autoformer improves point accuracy and xLSTM captures volatility dynamics, GARCH often yields fewer Value-at-Risk exceedances, revealing a trade-off between forecasting precision and risk adequacy. The second identifies twelve weaknesses in established model risk management (MRM) frameworks and proposes targeted enhancements across the model lifecycle. The third develops a framework that maps AI-specific risk drivers to materiality and complexity for inherent model risk tiering. Together, the essays show that responsible AI adoption requires assessing not only predictive performance but also governance effectiveness and model risk.