A1661
Title: Accurate Bayesian inference for tail risk extrapolation in time series
Authors: David Carl - Bocconi University (Italy)
Simone Padoan - Bocconi University (Italy)
Stefano Rizzelli - University of Padova (Italy) [presenting]
Abstract: A Bayesian approach is developed for inference on extreme values, focusing on high-threshold exceedances, and explicitly addressing two common sources of model misspecification: the choice of the peaks distribution and the presence of temporal dependence not captured by the working likelihood. In practice, peaks-over-threshold analyses typically rely on modelling assumptions (e.g., a Generalised Pareto model, justified by extreme-value theory) and independence across exceedances, both of which are typically violated. The proposed methodology is designed to remain robust to misspecification of the peaks distribution and to residual temporal dependence, delivering reliable inference for key tail quantities even when these assumptions fail. Posterior consistency, asymptotic normality, and the asymptotic validity of credible regions are established, thereby providing theoretical support for the methodology. Simulation studies illustrate the impact of such misspecification on standard methods and demonstrate the improved accuracy of the proposed approach, while applications to real data highlight its practical value.
