Many applications of human-robot interaction (HRI) necessitate close physical collaboration. Accurate prediction of human intent can be utilized to allow robots to select actions and motions that are safer and more efficient. While a variety of human motion prediction approaches have been developed, they are often designed for specific types of tasks or motions, and thus do not generalize well. Consequently, it is not always obvious what method is appropriate for a given task, making human motion prediction difficult to implement in practice. Toward addressing this problem, we introduce the concept of human motion prediction ensembles, where optimal combinations of prediction approaches are generated by learning from task data. We argue that by learning what predictors and parameters work best directly from task data, we can generate a motion prediction technique that can achieve higher performance than individual methods, while also reducing the implantation overhead for the user. In this work, we describe a preliminary implementation of this concept, which combines a goal-based prediction method with a velocity-based motion projection method. We evaluate the performance of the combined predictor against that of the individual methods in terms of accuracy of prediction of human position over a range of look-ahead time values. The results indicate that the combined predictor, which is assembled based on learning from task data, outperforms the individual goal-based and velocity-based methods by 17.2% and 15.9%, respectively. We also describe avenues for future work, discussing possible ways of extending this approach to accommodate additional prediction methods.