Projection-based inference with particle swarm optimization

Abstract

This paper introduces Particle Swarm Optimization [PSO] to econometrics with focus on projection-based test inversion. Econometricians have developed such methods to enable a robust analysis of imperfectly identified models. Despite important theoretical breakthroughs, computational and numerical tool kits have not followed suit. This paper compares stochastic solvers including PSO on speed and accuracy for the problem. Empirically, the paper analyzes a three-equation New Keynesian model for the U.S.. Results are confirmed via a synthetic sample with relevant and weak instruments. In contrast to PSO, we find that popular solvers may converge to local optima suggesting misleading decisions on the nature of the New Keynesian Phillips Curve, determinacy of monetary policies, and the persistence of the Taylor rule. Results confirm that far more attention needs to be paid to numerical precision as test inversion duly gains popularity in applied econometrics.