The cerebellum is thought to be able to learn forward models, which allow to predict the sensory consequences of planned movements and adapt behavior accordingly. Although classically considered as a feedforward structure learning in a supervised manner, recent proposals highlighted the importance of the internal recurrent connectivity of the cerebellum to produce rich dynamics (Rössert et al., 2015), as well as the importance of reinforcement-like mechanisms for its plasticity (Bouvier et al. 2019). Based on these models, we propose a neuro-computational model of the cerebellum using an inhibitory reservoir architecture and biologically plausible learning mechanisms based on perturbation learning. The model is trained to predict the position of a simple robotic arm after ballistic movements. Understanding how the cerebellum is able to learn forward models might allow elucidating the biological basis of model-based reinforcement learning.