A0369
Title: Expectation-maximization for non-Gaussian dynamic factor models: A novel approach for macroeconomic nowcasting
Authors: Domenic Franjic - University of Hohenheim (Germany) [presenting]
Abstract: Macroeconomic data often exhibit heavy tails due to extreme events such as financial crises or the Covid-19 pandemic, where traditional Gaussian models struggle to capture outliers effectively. To tackle these issues, an efficient estimation approach for dynamic factor models (DFM) with heavy-tailed errors is presented. The method employs an expectation-maximization (EM) algorithm that extends conventional approaches by incorporating the multivariate Student's t-distribution. The estimation algorithm applies bootstrap filtering to estimate the latent states. An augmented Lagrangian method is used to estimate the factor loadings under regularization constraints that induce shrinkage and enforce exact zero loadings in the factor loading matrix. Hyperparameter tuning is efficiently integrated into the EM framework. The method also handles mixed-frequency data and missing observations directly. By modeling heavy tails, inducing shrinkage, and improving robustness, the method reduces the nowcasting error of macroeconomic variables such as GDP growth. Simulation results demonstrate that the non-Gaussian DFM consistently outperforms Gaussian alternatives, particularly when the degrees of freedom of the error distributions are small.
