Real-world problems are often characterized by the large size of their input space, the noise which is added to these data and by the complexity of the underlying physical laws. These laws are often continuous and more information can be brought to the understanding of the problem if contextual hints are added. This contextual information yields close behaviors. This topology is itself often linked to the topology of the input space. This paper explores how recurrent models can improve prediction in a radio communication problem with such contextual information