Upcoming Talks
Upcoming Talks
October 29, 15:00 UTC time = (11:00 New York) = (16:00 Paris) = (23:00 Shanghai)
Extreme hourly precipitation is becoming more severe, yet monitoring its associated risk over time remains challenging. Recent studies borrow from lower-frequency precipitation analyses, which usually rely on regression techniques based on Extreme Value Theory (EVT). While this approach is very popular and useful for assessing covariate effects, it may not be suited to monitoring purposes. This talk moves to a dynamic framework in which the parameters of EVT distributions evolve according to autoregressive-type “score-driven” dynamics. Score-driven dynamics are suitable for filtering the different layers of serial dependence arising in high-frequency data, they are parsimonious and completely determined by past information (making inference tractable), and enjoy information-theoretic optimality properties even under model misspecification. We develop a fixed-effect score-driven model for the Generalized Pareto distribution, designed for panel data. The model “borrows strength” across locations to improve marginal estimation, while preserving cross-sectional heterogeneity. A study of extreme hourly precipitation in the US Midwest region shows that the model is able to detect trends in estimated extreme quantiles and represents a valid alternative to well-known non-linear EVT regression strategies.
Ths joint work with Debbie Dupuis and Luca T