A1324
Title: Bayesian trend-cycle decomposition and forecasting
Authors: Pawel Szerszen - Federal Reserve Board of Governors (United States) [presenting]
Charles Knipp - Federal Reserve Board (United States)
Mohammad Jahan-Parvar - Federal Reserve Baord of Governors (United States)
Abstract: Out-of-sample forecasting performance of a general family of unobserved component models applied to macroeconomic time series on inflation, output growth, and unemployment are studied. First, it is documented that fully Bayesian estimation accounting for parameter uncertainty dominates density forecasts produced with the common approach of replacing unknown parameters with maximum likelihood-based estimates. The findings are consistent for all studied macroeconomic time series and still hold when restricted to forecasting extreme events. Second, the optimal pooling of all studied models is found to outperform out-of-sample forecasts produced by any single model specification. The unobserved component model with stochastic volatility in the optimal pool receives the highest weights. Third, it is studied that in the event a joint information content of inflation, output growth and unemployment translates into their improved out-of-sample forecasts. It is found that the multivariate unobserved component models can further improve the forecasting performance of individual macroeconomic series.