Abstract
This thesis offers a multifaceted exploration into the interaction between financial regulations, advanced Machine Learning, and Financial Risk Management, aiming to shed light on the intricate workings of financial markets and the quest for greater inclusivity within them.The exploration embarks with a critical examination of the second Markets in Financial Instruments Directive (MiFID II) mandated the unbundling of payments for research and trading. This research explores whether the impact of MiFID II differs between large and small firms in terms of analyst coverage and stock liquidity. Focusing on the UK stock markets we find a significant drop in analyst
coverage on the Main Market, which leads to a deterioration in market liquidity. In contrast, the requirement of AIM firms to retain a Nominated Adviser (NOMAD), who often provides research coverage, has mitigated the impact of MiFID II.
Further, the thesis progresses to discuss the transformative potential of machine learning in reshaping mortgage lending paradigms and enhancing financial inclusion. Using a unique database from one of the UK's largest mortgage providers, we investigate how machine learning enhances risk assessment
for borrowers, and in particular, improves financial inclusion for low-income borrowers. In a sampled, non-biased equilibrium credit market, I find that traditional models disproportionately disfavour low-income
borrowers in mortgage approval. However, the use of machine learning significantly reduces default rates and increases financial inclusion. These improvements are primarily attributed to machine learning's ability to effectively capture soft information and nonlinear relationships between features and credit risk, which often go undetected by conventional methods. Consequently, this
approach allows for more efficient monitoring of mortgage borrowers and reduces information asymmetry, thereby potentially alleviating financial exclusion.
Moving forward, the research focuses on the domains of VaR backtesting and machine learning forecasting techniques. This research pioneers the evaluation of SVR-GARCH models in VaR forecasting, offering a nuanced understanding of the efficiency and superiority of SVR-GARCH in volatility forecasting compared to the traditional econometric models, such as GARCH, and newer methodologies, such as neural networks. The study underscores the heightened precision of SVRGARCH models in VaR predictions, reflecting advancements in risk management and capital optimization. A comprehensive series of VaR backtests affirm the model's superiority in managing risk, ensuring efficient capital allocation, and offering high-quality forecasts.
| Date of Award | 27 Mar 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Ru Xie (Supervisor) & David Newton (Supervisor) |