Slow Expectation–Maximization Convergence in Low-Noise Dynamic Factor Models

This paper addresses the slow convergence of the expectation–maximization (EM) algorithm in the estimation of low-noise dynamic factor models, commonly used in macroeconomic nowcasting and forecasting. We show analytically and in Monte Carlo simulations how the EM algorithm stagnates in a low-noise environment, leading to inaccurate estimates of factor loadings and factor realizations. An adaptive version of EM considerably speeds up convergence, producing substantial improvements in estimation accuracy. An empirical nowcasting exercise of euro area GDP growth shows gains in root mean squared forecast error up to 34% by using the adaptive EM relative to the standard algorithm.