【Metrics Seminar】Uniform Confidence Bands for Generalized Random Forests
研討會日期 : 2023-09-05
時間 : 14:30
主講人 : Professor Wolfgang Karl Härdle
地點 : Conference Room B110
演講者簡介 : Professor Wolfgang Karl Härdle received his PhD in Mathematics from Heidelberg University in 1982. He is currently a Professor of Statistics at Humboldt-Universität zu Berlin. His research expertise includes Digital Finance, Financial Econometrics, Smart Data Analytics, Ultra High Dimensional Modelling, Deep Learning, Resampling Techniques, and AI in Digital Economy.
演講摘要 : Athey et al. (2019) provide an extension of random forests in their paper, called Generalised Random Forests (GRF) and show that an estimate obtained from GRF at the point is asymptotically normal. In this paper, we would like to check that this estimate is uniformly converging to the true function. More precisely, as the number of observations gets large and therefore also a uniform confidence band becomes smaller and smaller. Given a data set at hand, we find a critical value such for the uniform confidence bands using multiplier bootstrap, since it is well known that the standard approach via extreme value theory have a very slow asymptotic. We use simulations to verify that our method gives reliable results for the estimation and coverage even in small samples. As a real life application, we extend the example
provided in Athey et al. (2019) and construct uniform confidence to characterize the uncertainty in conditional local average treatment effects.