Macroeconomic Forecast Evaluation Under Asymmetric Loss

This dissertation consists of three studies that address the analysis of macroeconomic forecasts under an asymmetric loss function. It begins by contextualizing the specific challenge to macroeconomic forecasters before establishing the methodical foundation for the subsequent chapters and providing an extensive literature review. The second chapter introduces the concept of a loss function and argues in favor of relaxing the standard assumption of symmetric loss, as, for example, it is usually difficult to assume that the costs related to a positive and a negative forecast error of the same magnitude are identical. Central to the studies presented in this thesis is the Elliott, Komunjer und Timmermann (2005, 2008) approach (EKT), which allows for an estimation of the degree of asymmetry in a forecaster's loss function that minimizes the loss function for a given series of forecast errors. The test is based on a GMM estimation. An overidentification test offers the possibility of simultaneously testing the forecasts' efficiency with respect to a set of instrumental variables. The first study analyzes employment forecasts of the German Council of Economic Experts and the Joint Forecasts of the leading economic research institutes. Herein, the joint focus lies on analyzing the institutions' loss functions with respect to their potentially asymmetric preferences and the forecasts' information efficiency. A broad spectrum of instrumental variables is considered and tests are conducted on each individual instrument and on the aggregate of the instruments. We determine whether or not the instruments contain unused information. Regardless of the instruments, both institutions appear to prefer underestimating employment growth. Using a Monte-Carlo analysis, the second study aims to reveal more about the finite sample properties of the EKT approach. The main focus here is on the accuracy and precision of the asymmetry parameter estimates and the size and power properties of the overidentification test. The results show that the degree of asymmetry is estimated quite well, with increasing precision for larger sample sizes. This still holds when the forecasting model is misspecified. For correctly specified forecasts, the size of the overidentification test is close to the nominal size of 5 percent. The test shows very good power against misspecified forecasts. The third study concerns itself with how the degree of asymmetry changes when higher moments of the forecast error are included in the loss function. Taylor series approximations of Linex-based loss functions containing a linear and an exponential part are used to include the higher moments. The study analyzes the growth, inflation and unemployment rate forecasts for the euro area provided by the European Central Bank's Survey of Professional Forecasters. Results show indications that the forecasters prefer overestimating growth and underestimating inflation and the unemployment rate. Including the second moment leads to a reduction of asymmetry in the loss function for all three variables. However, including third and fourth moments hardly changes the loss function’s form. This result can be interpreted in such a way that accounting for the forecasters' risk aversion suffices.