Interpretable and Scalable DataDriven Models in Financial Time Series Analysis
In big data age, financial time series are increasingly brought to the attentions of researchers in business management and financial econometrics. With the development of data science, classical statistical time series models are being replaced by a variety of data-driven models. Although these data-driven models empowered by black-box machine learning techniques can usually accommodate the data scalability issues, they are being widely criticised for its lack of interpretability and transparency.
In financial market, risk is an eternal topic because we cannot know tomorrow for 100% sure no matter how advanced the forecasting techniques are. In the literature, there are many quantitative measurements of risk, e.g., Standard Deviation (SD), Value at Risk (VaR), and Expected Shortfall (ES). However, these measures usually capture the static and frequency-based uncertainty only, which means that they can hardly catch up the fast-changing financial markets. These drawbacks restrict the further development of a general, effective risk control framework in financial time series analysis in this big data age. To address these drawbacks, as a probabilistic approach, Bayesian non-parametric (BNP) model can relax the assumption (e.g., normality, linearity, stationarity) without losing interpretability and without substantial loss of computational efficiency.
Therefore, the whole project aims to apply Bayesian non-parametrics (BNP) to establish a scalable and transparent framework for financial time series analysis in big data age and provide new insights into the development of interpretable data-driven market risk management.