B0554
Title: Partially-interpretable neural networks for extreme quantile regression
Authors: Jordan Richards - King Abdullah University of Science and Technology (Saudi Arabia) [presenting]
Raphael Huser - King Abdullah University of Science and Technology (Saudi Arabia)
Emanuele Bevacqua - Helmholtz Centre for Environmental Research (Germany)
Jakob Zscheischler - Helmholtz Centre for Environmental Research (Germany)
Abstract: Quantile regression is a powerful tool for modelling environmental data which exhibits spatio-temporal non-stationarity in its marginal behaviour. If our interest lies in quantifying risk associated with particularly extreme events, we may want to estimate conditional quantiles that lie outside the range of observable data; it is practical to describe characteristics of the data using a parametric extreme value model with its parameters represented as functions of predictors. Classical approaches rely on linear, or additive, functions and such models suffer in either their predictive capabilities or computational efficiency in high dimensions. Neural networks can capture complex structures in data and scale well to high dimensions, but statisticians may choose to forego their use as a result of their ``black box'' nature; However, they facilitate highly accurate prediction, statistical inference with neural networks is difficult due to their abundance of estimable parameters. We propose a framework for performing extreme quantile regression using partially-interpretable neural networks, which combine semi-/parametric methods with deep learning and facilitate both high predictive accuracy and statistical inference. We use our approach to estimate extreme quantiles for a relatively high-dimensional dataset and gain insights into the drivers of extreme wildfires occurring within, and around, the Mediterranean Basin.