Accurately predicting human movement trajectories is of critical interest in multiple fields, including humanexoskeleton interaction. In general, such predictions can be obtained from model-based approaches (e.g., optimal control theory) or from data-driven approaches (e.g., learning from human demonstrations). Data-driven methods generally require numerous demonstrations but avoid the computational burden of model-based methods. In this paper, we introduce a hybrid method mixing the strengths of these two approaches to enable the prediction of human trajectories in a recedinghorizon fashion when the number of actual demonstrations is very limited. First, we propose to use the stochastic optimal feedforward-feedback control framework to generate a large set of humanlike trajectories, including their variability. Second, we used the probabilistic movement primitives (ProMPs) framework to learn the distribution of these synthetic trajectories and make online predictions about the upcoming human trajectories from the observation of past movement data. Here, we evaluated our hybrid method on an existing data set composed of arm reaching movements in a parasagittal plane. Overall, our method proves to be advantageous when generalization is needed and demonstrations are lacking, such as in novel targets scenarios. The introduced method shows promise to efficiently predict realistic human trajectories on a given time horizon, even when limited or no human demonstration is available for the task at hand.