December 2024
Trend cycle decomposition and prediction using Bayesian multivariate unobserved components
Mohammad R. Jahan-Parvar, Charles Knipp, Paweł J. Zelzen
Abstract:
We propose a generalized multivariate unobserved component model to decompose macroeconomic data into trend and cyclical components. Next, predict the series using Bayesian techniques. We document that full Bayesian estimates that account for state and parameter uncertainties consistently outperform out-of-sample predictions produced by alternative multivariate and univariate models. . Additionally, accounting for stochastic variation components in variables improves predictions. To address data limitations, we leverage cross-sectional information, use commonalities between variables, and consider both parameter and state uncertainties. Finally, we find that the optimally pooled univariate model outperforms the individual univariate specifications and generally performs close to the benchmark model.
Keywords: Bayesian inference, maximum likelihood estimation, online prediction, out-of-sample prediction, parameter uncertainty, sequential Monte Carlo method, trend cycle decomposition
DOI: https://doi.org/10.17016/FEDS.2024.100
PDF: Full paper
Return to top
Last updated: December 30, 2024
