MathSEED Lecture: Risk modeling in the face of rare events and distribution shifts | Nicola Gnecco

Abstract:;
In risk modeling, the events and distributions we wish to reason about are often not adequately represented in the available data. This gives rise to two challenges. (i) The first is the prediction of rare events, that is, events that are under-represented or absent from the available data. This problem has traditionally been addressed through extreme value theory. (ii) The second is prediction under distribution shift, where the target distribution differs from the one that generated the training data, due to interventions or changes in the environment. This challenge has traditionally been addressed through causal inference and, more recently, distribution generalization. This talk focuses on the setting where both challenges (i) and (ii) are present, that is, predicting rare events under distribution shifts. In the first part, we discuss a framework that uses causal assumptions and tail invariance properties of multivariate regular variation to study extreme interventions and causal effects in the tails of a distribution. In the second part, we consider situations in which no assumptions about the nature of the distribution shift or its tail behavior are available. Using a dynamical system as a test bed, we show that predictive disagreement across a model ensemble can detect when a test observation is both extreme and out-of-distribution.

Date and Venue:
Tuesday, 01 December 2026, 15:45
Mathematics Building 20.30
Room 2.059

More Information