How useful is external information from professional forecasters? Conditional forecasts in large factor models

This paper evaluates forecasts from a factor model estimated with a large real-time dataset of the German economy. The evaluation focuses on a broad cross-section of variables such as activity series including components of the gross domestic product and gross value added, deflators and other price measures as well as several labor market indicators. In addition to unconditional forecasts for these variables, we also investigate to what extent the forecast accuracy improves when we condition on professional forecasters' view on GDP growth and CPI inflation. We find that over the period from 2006 to 2017 the model's unconditional forecasts are broadly in line with autoregressive benchmarks for the majority of the 37 series that we focus on in the evaluation, in some cases performing somewhat better and in others somewhat worse. For a few variables capturing real activity and some price indicators, however, we find large gains in predictive accuracy that persist for forecast horizons of up to two quarters ahead. Conditioning on external information tends to improve the forecast accuracy in some instances but typically only for those series where the unconditional forecasts are already quite accurate. For around a third of the variables under consideration, the differences in forecast accuracy between conditional and unconditional forecasts are statistically significant for density forecasts; for point forecasts on the other hand we find no significant differences. From a methodological point of view, this paper proposes precision-based sampling algorithms to draw from the predictive density - unconditional or conditional on a subset of the system variables - in factor models and other models with unobserved components. Simulations show that these algorithms perform favorably compared to Kalman filter-based alternatives typically used in the literature.