A0209
Title: Neural classification of asymptotic (in)dependence
Authors: Troy Wixson - University of Massachusetts Amherst (United States) [presenting]
Daniel Cooley - Colorado State University (United States)
Abstract: Studies in extremes often aim to extrapolate into the tail beyond the range of the data; For example, to assess the risk of the combined effect of extreme precipitation and storm surge. Extrapolation under the wrong dependence regime can have large negative effects, and thus classification is a necessary early choice in the modeling of multivariate extremes. Inference about the dependence regime is complicated as the regimes are defined asymptotically. A series of experiments is performed to determine whether a finite sample has enough information for a convolutional neural network to reliably distinguish between these regimes in the bivariate case. A new classification tool is developed for practitioners, which is called \texttt{nnadic}, as it is a neural network for asymptotic dependence/independence classification. This tool accurately classifies over 97\% of test datasets and is robust to a wide range of sample sizes. The datasets that are unable to be correctly classified tend to either be nearly exactly independent or exhibit near-perfect dependence, which are boundary cases for both the ADep and AInd models used for training. These experiments highlight that ADep and AInd models do not so much differ in the strength of tail dependence they can capture (as both regimes can range from independence to complete dependence), but they instead differ in whether the dependence completely decays in the limit, irrespective of the path of that decay.
