Forecasting Austrian GDP using the generalized dynamic factor model.

In this paper, a generalized dynamic factor model is utilized to produce short-term forecasts of real Austrian GDP. The model follows the frequency domain approach proposed by Forni, Hallin, Lippi and Reichlin (2000, 2003). The forecasting performance of the model with a large data set of 143 variables has been assessed relative to simple univariate time-series forecasts. The results show that the factor model can barely outperform the much simpler benchmark model, given the usuall levels of significance. Thus we followed a line of research proposed by Boivin and Ng (2003) and Watson (2000), who suggested that the use of a small data set may increase the forecasting performance. The main finding from our extensive out-of-sample forecasting experiment that we have conducted is that the best forecasting performance can be achieved with small data sets with a handful of variables only. These models perform signifi- cantly better than the large model. This result seems to contradict the basic idea of dynamic factor models, which have been constructed to exploit the potentially useful information of a large data set.