An Unobserved Components Model To Forecast Austrian Gdp PDF Download

We nowcast and forecast Austrian economic activity, namely real gross domestic product (GDP), consumption and investment, which are available at a quarterly frequency. While nowcasting uses data up to (and including) the quarter to be predicted, forecasting uses only data up to the previous quarter. We use a large number of monthly indicators to construct early estimates of the target variables. For this purpose we use different mixed-frequency models, namely the mixed-frequency vector autoregressive model according to Ghysels (2016) and mixed data sampling approaches, and compare their forecast and nowcast accuracies in terms of the root mean squared error. We also consider traditional benchmark models which rely only on quarterly data. We are particularly interested in whether explicitly considering different regimes improves the nowcast. Thus we examine regime-dependent models, taking into account business cycle regimes (recession/non-recession) or financial/economic uncertainty regimes (high/low uncertainty) driven by global and Austrian economic and financial uncertainty indicators. We find that taking explicit account of regimes clearly improves nowcasting, and different regimes are important for GDP, consumption and investment. While the recession/non-recession regimes seem to be important to nowcast GDP and consumption, high/low global financial uncertainty regimes are important to nowcast investment. Also, some variables seem to be important only in certain regimes, like tourist arrivals in the non-recession regime when nowcasting consumption, while other variables are important in both regimes, like order books in the high and low global financial uncertainty regimes when nowcasting investment. In addition, nowcasting indeed provides a value added to forecasting, and the new information available in the first month seems to be most important. However, sometimes also the forecast performs quite well, and then it mostly comes from a mixed-frequency model. So monthly information seems to be helpful also in forecasting, not only in nowcasting. Finally, we do not find a clear winner among the different mixed-frequency models.