We use the concept of predictability as presented in Diebold and Kilian (2001) to assess how well the growth rates of various components of German GDP can be forecasted. In particular, it is analyzed how well different commonly used leading indicators can increase predictability of these time series. To this end, we propose an algorithm to select an “optimal“ information set from a full set of possible leading indicators.
In the univariate set up, we find very small degrees of predictability for all quarterly growth rates whereas yearly growth rates seem to be more predictable at short forecast horizons. According to the algorithm proposed, from a set of financial leading indicators the short term interest rate is included in the highest number of information sets and from a set of survey indicators the ifo-business expectation index is included in most cases. Conditioning on the “optimal“ sets of leading indicators improves the predictability of most of the quarterly growth rates substantially while the predictabilities of the yearly growth rates cannot be increased significantly further. The results indicate that there is clearly evidence that “complicated“ forecasting models are usually superior to simple AR univariate models.