Date: Friday, 29 May 2026, at 12.30pm (CET)

Dual format: ESSEC Paris La Défense (CNIT), Room TBA
and via Zoom, please click here

This seminar presents a new approach for simulating multivariate extreme events, motivated by applications such as large-scale rainfall, simultaneous stock market shocks, and widespread transportation failures. The method is grounded in Extreme Value Theory (EVT) and assumes that, after transformation to a standard scale, the underlying distribution is multivariate regularly varying. The proposed framework combines a Wasserstein-Aitchison Generative Adversarial Network (WA-GAN) with joint modeling of univariate marginal tails to generate realistic extreme observations on the original scale. Central to the approach is the angular measure, which captures tail dependence through distributions on the unit simplex. By transforming these angles into Aitchison coordinates, the method enables Wasserstein-GAN techniques to operate in a linear space. The methodology is illustrated using both simulated data under different tail dependence structures and a financial dataset from the Kenneth French Data Library. Results demonstrate strong performance relative to existing approaches, particularly in capturing complex tail dependence and generating accurate extreme-event scenarios. This is a joint work with S. Lhaut and J. Segers.

Kind regards,
Pierre Alquier, Marie Kratz, Roberto Reno, and Riada Djebbar (Singapore Actuarial Society - ERM)