Repositorio Académico Institucional

Tail risk forecasting holds great significance in various contexts, including financial risk management, derivative pricing, hedging, and portfolio construction. Accurate risk forecasts are of paramount importance from both policy and regulatory perspectives, as well as for the internal risk control of financial institutions. Risk managers and regulators have been challenged to implement new internal models to capture tail risk and market illiquidity. The main objective of this thesis work is to propose new multivariate dynamic models for extreme financial risk forecasting. The proposed specifications consider both dimensions of systemic risk (time and cross-sectional) and are theoretically grounded in extreme value theory. Thus, are framed within the systemic tail risk models. The first chapter of this thesis provides a comprehensive review of the existing literature on univariate specifications based on directly modeling the dynamics of observations that exceed a high threshold, which are considered extreme events. This review identifies the characteristics of the best-performing models for tail risk forecasting, as well as those specifications that, due to their lower dimensionality and simplicity in updating the parameters, are easier to extend to multivariate contexts. Based on these findings, multivariate models have been proposed. The second chapter contributes to the systemic tail risk literature by presenting novel models that utilize observable factors of comovement in the commodities market to tackle the challenge of high dimensionality. These models are designed to capture multivariate extreme dependence and provide a valuable financial risk management tool for regulators, buyers, and sellers of commodities, particularly in the energy sector, where the most significant extreme risk spillovers are identified. The third chapter presents a new multivariate dynamic model designed to predict extreme financial risks, which considers data from both extreme and non-extreme events to refine its parameters using a vector autoregressive approach. Furthermore, the model can be improved by incorporating realized volatility measures. The crucial takeaway from this chapter is that incorporating a wider range of information (cross-excitation effects and realized volatility measures) will lead to a better fit and more accurate risk prediction.