WG Risk :“Reinforced Quantile Regression via Extreme Value Theory”

The Working Group on Risk - CREAR, with the support of the ESSEC IDO dpt/Ceressec, Institut des actuaires, Labex MME-DII (CY), and the group Risques AEF - SFdS, has the pleasure to invite you to the seminar by :

Prof. Yi HE, University of Amsterdam, School of Economics, Netherlands

Dr. Yi He is a Data Scientist and Associate Professor at the Amsterdam School of Economics, University of Amsterdam, in the Netherlands. He is also a research fellow at the Tinbergen Institute. Yi received his Ph.D. from Tilburg University in 2016. Prior to joining the University of Amsterdam, he was a tenured Assistant Professor in the Department of Econometrics and Business Statistics at Monash University in Australia from 2016 to 2019. His research has been published in top journals in statistics and econometrics. His research interests include extreme value statistics, high-dimensional econometrics, and financial econometrics. In extreme value statistics, he has worked on multivariate quantile inference and has recently extended his work to the analysis of heterogeneous data. In high-dimensional econometrics, his work focuses on optimal prediction and testing in non-sparse models using random matrix theory. In risk management, Yi He has developed data-driven methods based on empirical likelihood and bootstrapping to quantify the statistical uncertainty of risk measures such as value-at-risk and related systemic risk metrics.

“Reinforced Quantile Regression via Extreme Value Theory”

Extreme-order regression quantiles suffer from inconsistency and non-normal asymptotic distributions due to data sparsity in the tails. Building on extreme value theory, we propose a novel regression estimator called reinforced quantile regression, which explicitly extrapolates the score functions of quantile regression by incorporating power-law behavior for heavy-tailed outcomes or exponential transformations for light-tailed outcomes. We establish the asymptotic normality of this estimator, which converges at a rate strictly faster than that of classical quantile regression at extreme quantile levels. Furthermore, we have proved the asymptotic validity of bootstrap inference using random weights. Simulation results show that our estimator outperforms existing methods, offering more accurate estimation and narrower, near-exact bootstrap confidence intervals. Applying the proposed method to a dataset on Chinese twins, we find significant positive marginal predictive effects of education on upper-income percentiles. Unlike classical approaches, which show diminishing or even negative effects in the tails, our method yields stable and significantly positive estimates even at high quantile levels.

Dual format:

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