Session: Technical Session III

Date & Time: Saturday, May 23 — 9:30 AM–11:15 AM

Talk Title

High-Dimensional Nonparametric Local Projections

Abstract

This paper develops a flexible two-step high-dimensional nonparametric local projections estimator (HDNLP) by constructing Neyman-orthogonal pseudo-outcomes whose expectation conditional only on low-dimensional policy variables identifies the average structural function. In the first step, we construct pseudo-outcomes with high-dimensional nuisance parameters using flexible machine learning tools and cross-fitting. In the second step, we estimate the conditional expectation using local linear estimation, allowing for flexible inference on potentially nonlinear impulse responses. We establish asymptotic properties of the HDNLP estimator at standard nonparametric rates and demonstrate superior finite-sample performance relative to existing methods in simulations and in an application on the nonlinear effects of financial conditions on the economy. (Joint work with Jiming Huang, Massimiliano Marcellino and Tommaso Tornese.)