The ability to predict the risk patterns of vehicles in a roundabout shows potential to improve the safety and efficiency of roundabout crossings by connected vehicles. Namely, exit patterns are relevant for vehicles seeking to enter a roundabout in the presence of incoming vehicles. Entering vehicles may take educated decisions on whether to enter the roundabout based on the likelihood of the incoming vehicles not to exit and cause a merging conflict. In previous work, a machine learning model was trained to assess the probability of a vehicle to exit a roundabout based on its observed position relatively to the next exit. Yet, the transferability of the knowledge of exit probability models was not investigated, i.e., whether the knowledge of an existing exit probability model can be accurately transferred in unseen roundabouts, both for model usage and training. In this paper, we compute a metric of similarity of exit probability models trained from eight real roundabouts. In turn, we identify the contextual features of two roundabouts which impact the similarity of their resulting exit probability models, and define three levels of context similarity, i.e., strict, moderate, and low. Lastly, significant accuracy improvements are obtained by constraining the knowledge transfer of exit probability models to roundabouts which feature a similar context. On the one hand, applying exit probability models on distinct roundabouts with a moderately similar context yielded an average accuracy of 80.4±4.6%, which is equivalent to the most accurate non-similar models. On the other hand, training a model for an unseen roundabout using exclusively training data extracted from roundabouts with a moderately similar context featured a 80±5% accuracy, which represents a consistent accuracy increase of 8.5±4.8% compared with knowledge transfer without context constraints.