Predicting the intentions of other vehicles in traffic is a frequently addressed challenge in autonomous driving. Due to the complexity and diversity of urban traffic, it is a major challenge to develop prediction models that are able to generate reasonable predictions for a broad range of situations. Commonly employed data-driven approaches encounter problems related to the lack of transparency of black-box approaches and poor generalizability due to overfitting. Meanwhile, most of the publications to date have focused on the modeling part, but investigations that provide transparency into the transferability of learned patterns and the effect of different settings on generalizability are rarely addressed. This paper addresses these challenges by presenting an advanced evaluation method providing insight into the ability of models to create plausible predictions even in exceptional situations. The proposed method is applied to investigate variations in the provided input information, varying diversity in training data, and different model parameters. Among other things, our results show that providing semantic contextual information and enriching real training data with synthetic samples contributes to better generalizability. Furthermore, the evaluation revealed weaknesses of commonly used metrics, as the exclusive use of displacement errors can be misleading in terms of generalizability and plausibility of results. In summary, this contribution paves the way for reliable predictions in urban traffic by providing valuable insights and a methodology for a critical evaluation of prediction models.