We propose a multitask, regression-based approach for predicting future performances of soccer players. The multitask approach allows us to simultaneously learn individual player models as offsets to a general model. We devise multitask variants of ridge regression and ε-support vector regression. Together with a hashed joint feature space, the generalized models can be optimized using standard techniques. Relevant features for the prediction are identified by a modified recursive feature elimination strategy. We report on extensive empirical results using real data from the German Bundesliga.