Risk is a perennial problem in the financial market because, despite the best efforts of forecasting techniques, we cannot foresee the future with absolute certainty. Numerous quantitative measures of risk are used in the literature, such as Value at Risk (VaR) and Expected Shortfall (ES). These measurements, however, typicaly only account for static and frequency-based uncertainty, making it difficult for them to keep up with the financial markets' rapid changes. These limitations impede the continued development of a universal, practical framework for managing risks in financial time series analysis in the big data era. Bayesian non-parametric (BNP) models promise to loosen the assumption (such as normality, linearity, stationarity) without sacrificing interpretability and increasing computational cost in order to solve these limitations as a probabilistic strategy.
The project's objectives are to theoretically and empirically examine financial market risk and further explore the idea of applying BNP to this risk. Risk measurement, risk modelling and forecasting, and the evaluation of risk models are some of the subjects that make up our project.
Knowing how to quantify risk is important before making any predictions about market risk. In order to achieve this, we explicitly investigated the criteria that a risk measurement must satisfy, namely Translation Invariance, Subadditivity, Positive Homogeneity, and Monotonicity, before theoretically examining the statistic and stochastic representation for the risk measurement. Even though VaR was the major emphasis of this seed fund project, once stochastic analysis tools are introduced, we can take a dynamic representation of risk into consideration.
We thoroughly investigated the various approaches for risk prediction both theoretically and empirically, in contrast to previous studies that either only compared the data-driven methods, such as neural networks and quantile regression, or only investigated the traditional model-driven approaches for market risk prediction, such as GARCH, EGARCH, and CAViaR. Additionally, a clear road map for the relationships between various models and how they are upgraded and evolved was extensively illustrated with a discussion of their benefits and cons.
We chose some of the most representative and State-Of-The-Art models, such as GARCH, EGARCH, LSTM, and so on, to run our experiment on the appropriate configuration of the datasets in order to provide an adequate comparison between various methodologies. In addition to predicting VaR for the test dataset, various assessment techniques (such as the Kupiec-test and exception rate) have also been used to confirm the statistical superiority of one model over others. Our tested model spans a wide range, from the ARCH family to cutting-edge neural network models.
Furthermore, we created a helpful framework (a Python package) that enables academics to contrast various market risk models, including both traditional and data-driven risk models. In order for us to create a new BNP framework to control financial market risk, this effort is crucial.
Overall, our initial research on financial market risk has prepared us to consider the next stage, which involves using BNP tools to control financial market risk. During the project's duration, I visited my collaborator, Dr Bo Wang, at the University of Leicester, and I presented a research seminar on financial market risk from a Bayesian non-parametric perspective. We settled on a topic, 'functional data analysis for financial market risk modelling with Gaussian process quantile regression', after a lengthy discussion. After carefully developing the concept, Bo and I will submit an EPSRC proposal jointly.