Gossip Learning (GL) is a fully decentralized machine learning paradigm with the potential to enable highly scalability and to preserve user privacy. The majority of existing results however consider scenarios in which either each node communicates with all other nodes, or in which the connectivity graph is static, and they are therefore inapplicable in dynamic setups such as in VANETs. This work is a first attempt at designing and assessing GL schemes suited for scenarios with moving nodes with the application of predicting the trajectory of moving cars.