A recent surge of experimental evidence points towards a significant role of the cerebellum in the development and maintenance of neocortical states through cortico-cerebellar loops. Notably, the re- sulting positive feedback loop is not limited to the motor domain to which the cerebellum is classically associated but has been shown to extend to various cognitive processes. However, it remains unclear what functional principles may underlie these cerebro-cerebellar loops. Here we model a neocortical area as a recurrent neural network which projects to, and receives from, a cerebellar feedforward net- work. Neocortical plasticity is modelled with biologically plausible temporal credit assignment, and cerebellar plasticity with a temporal window-specific learning rule used to predict neocortical feedback, in line with recent experimental observations. Our model captures cerebellum-driven cortical dynam- ics observed experimentally in both motor-based and working memory tasks. In a motor-based task we find that cerebellar feedback consistently improves the rate of learning and mitigates the need for neocortical plasticity. This is due to the predictive learning of our cerebellar model, which triggers a fast and reliable drive of early neocortical states. In working memory tasks, in which maintenance of representation is critical, the model achieves good task performance for all neocortical plasticity as- sumptions. Overall, we propose the cerebellum is an effective driver of neocortical dynamics with task relevant information, reducing the need for neocortical plasticity in both motor and cognitive domains.