An outstanding question posed today by end-users like power system operators, wind power producers or traders is what performance can be expected by state-of-the-art wind power prediction models. This paper presents results of the first ever intercomparison of a number of advanced prediction systems performed in the frame of the European project Anemos. A framework for error characterization has been developed consisting by a measure- and a distribution-oriented approach. This comparison has given a perspective of the possibilities and limitations of the forecasts in the different test cases that were defined. At a second stage, the homogenous comparison process has permitted to evaluate the possibility of obtaining better performance by exploiting the merits of individual models through model combination. The paper presents the methodology and results from the combination approach.