Climate and weather extremes pose some of the most consequential risks associated with a changing climate, yet their rarity, complexity, and strong dependence on environmental conditions make them particularly challenging to model. Deep neural networks offer powerful tools for learning such relationships from high-dimensional data, but their limited interpretability can be problematic when models are used to understand changing hazards, inform adaptation, or support high-stakes decisions.
Jordan Richards will introduce deep extremal regression models: neural-network-based approaches designed specifically to learn features of conditional tails, including extreme quantiles and tail indices. By embedding asymptotically justified extreme-value models within neural networks, these methods provide a form of grey-box modelling—retaining the ability of deep learning to capture complex, high-dimensional relationships while preserving interpretable descriptions of how covariates influence extreme-event risk.
Jordan will illustrate this framework through models for extreme quantile regression, return level estimation, and generative compound extremes. Applications to European wildfire extremes, UK metocean storms, and Alpine runoff demonstrate how explainable deep learning can help not only predict extreme events, but also identify the environmental drivers of changing climate risk.
About Jordan Richards
Jordan is a Lecturer in Statistics in the School of Mathematics at the University of Edinburgh. He received his PhD in Statistics and Operational Research in 2021 through the STOR-i Centre for Doctoral Training, supervised by Jon Tawn and Jenny Wadsworth at Lancaster University and Simon Brown at the UK Met Office. From 2021 to 2024, he was a postdoctoral researcher in the XSTAT group at KAUST, led by Raphaël Huser, where his research focused on developing sparse models for spatio-temporal extremes.
He works at the intersection of extreme value theory, spatial statistics, and probabilistic AI, with a particular emphasis on natural hazards and extreme climate risk. He is a co-organiser of the One World Extremes seminar, the Spatio-Temporal Statistics and Data Science seminar, and is one of the founding members of the Glasgow-Lancaster-Edinburgh-Newcastle Extremes Network (GLEN). He also sits on committees for the Young Statisticians and Environmental Statistics Sections of the Royal Statistical Society, and co-leads the TIES working group on High-Performance Statistical Computing (HPSC) for Environmental Sciences.