现代产业经济学系列讲座第298期

时    间：2026年06月04日14:30-16:00

地    点： 复旦管院史带楼304室

题    目：Fast Online Learning and Inference on Semiparametric Index Models

主讲人：Qingsong Yao, Assistant Professor, Louisiana State University

主持人：俞学文 副教授

摘要: Economic environments increasingly involve sequential information arrival and continuously evolving decision-making processes. This paper develops an online estimation and inference framework for semiparametric index models, extending conventional offline semiparametric analysis to streaming-data environments. We propose a two-phase online learning procedure. In the first phase, globally consistent warm starts are constructed for both finite- and infinite-dimensional components. In the second phase, local online refinements produce rate-optimal estimators through sequential updating. Policy-relevant objects, including average marginal effects, can be updated simultaneously along the same learning path. For monotone index models, we further develop model-specific warm-start procedures.

The proposed framework processes each observation only once, making it suitable for modern data environments with storage, privacy, and computational constraints. Given a sequence of observations, the algorithm generates a trajectory of updated estimators for structural parameters and policy functionals. These trajectories naturally induce online inference procedures for the corresponding targets, bypassing the difficult variance construction typically associated with semiparametric inference involving nonparametric components.

Our analysis further shows that the proposed two-phase structure extends beyond monotone index models and provides a general online implementation strategy for sieve-based semiparametric estimators. We apply the method to the trade data studied by Helpman et. al. (2008). Monte Carlo experiments demonstrate favorable finite-sample performance and substantial computational gains relative to conventional offline estimation.

个人简介：Yao is currently an assistant professor in the Department of Economics at Louisiana State University. He received his Ph.D. in Economics from Boston College in 2024. His research interests lie in semiparametric and nonparametric estimation and machine learning, and his work has appeared in the Journal of Econometrics and the Journal of Business & Economic Statistics, among others.