We evaluate the out-of-sample forecasting performance of six competing models at horizons of up to three quarters ahead in a pseudo-real time setup. All the models use information in monthly indicators released ahead of quarterly GDP. We estimate two models – averaged vector autoregressions and bridge equations – relying on just a few monthly indicators. The remaining four models condition the forecast on a large set of monthly series. These models comprise two standard principal components models, a dynamic factor model based on the Kalman smoother, and a generalized dynamic factor model. We benchmark our results to the performance of a naive model and the historical near-term forecasts of the Czech National Bank’s staff. The findings are also compared with a related study conducted by ECB staff (Barhoumi et al., 2008). In the Czech case, standard principal components is the most precise model overall up to three quarters ahead. However, the CNB staff’s historical forecasts were the most accurate one quarter ahead.
Authors: Arnoštová, K., Havrlant, D., Růžička, L., Tóth, P.