Estimation of extreme quantiles from heavy-tailed distributions with neural networks (Q6089222)

The authors introduce an extrapolation principle for estimating extreme quantiles in the non-conditional heavy-tailed distribution scenario with emphasis on bias corrected estimators. A neural-network estimator is constructed and its associated approximation error is presented alongwith an extrapolation principle for estimating conditional extreme quantiles and two conditional neural-network estimators with their approximation properties. The finite sample properties of the estimators are illustrated on simulated data in the unconditional case and are compared to seven bias reduced extreme-value competitors on six heavy-tailed distributions. The conditional neural-network estimators are tested on real rainfall data. This article should spur more research on the use of neural networks in extreme quanitle estimation.