A Constrained Dynamic Nelson-Siegel Model for Monetary Policy Analysis

Abstract

The Dynamic Nelson-Siegel (DNS) model implies that the instantaneous bond yield is a linear combination of yield curve’s level and slope factors. However, this constraint is not used in practice because it induces a singularity in the state covariance matrix. We show that this problem can be resolved using Bayesian methods. The key idea is to view the state equation as a prior distribution over missing data to obtain a hyperplane truncated multivariate normal conditional posterior distribution for the latent factors. This distribution can then be reparameterized as a conditional multivariate normal distribution given the constraint. Samples from this distribution can be obtained in a direct and computationally efficient manner, thus bypassing the Kalman filter recursions. The empirical significance of the resulting Yield-Macro Constrained DNS (YM-CDNS) model is demonstrated through both a reduced form analysis of the US Treasury yield curve, and a structural analysis of functional conventional and unconventional monetary policy shocks on the yield curve and the broader macroeconomy.