Statistics Seminar: Sebastian Engelke (University of Geneva)

These methods often fail when interest is in extrapolation, that is, prediction in areas of the predictor space with few or no training observations.

Extreme value theory provides the mathematical foundation for extrapolation beyond the range of the training data. In this talk we present recent methodology that combines this extrapolation theory with flexible machine learning methods to tackle the out-of-distribution generalization problem. Moreover, we design a statistical test to decide whether a new predictor point requires extrapolation or not.

We highlight the extrapolation problem using AI-driven weather forecasting models such as GraphCast and Pangu-Weather. While these models outperform classical physical models in terms of overall mean squared error, they fall short when it comes to accurately forecasting extreme events.