Animal movement trajectories represent a highly non-linear system of interest to ecologists. In animal systems exhibiting collective behavior, the non-linearity is accentuated by interactions and feedback between individuals. Furthermore, movement data typically are prone to missingness and therefore, good prediction models provide tools to accurately fill in the gaps in a manner that provides a measure of uncertainty. As these datasets can be quite large, the computational efficiency of the model is an important consideration. Neural-based machine learning methods, such as recurrent neural networks, can capture complex dynamics of animal systems through relatively simple frameworks, but often do not provide uncertainty quantification or computational efficiency. Here, we investigate efficient recurrent neural network implementations in a framework that can account for prediction uncertainty and which can account for collective behavior.