To date, accurate vehicle trajectory prediction is an unsolved and challenging problem, which is of paramount importance for autonomous driving. Various open research questions exist in different subdisciplines of trajectory prediction. State-of-the-art trajectory prediction approaches are either based on a one-shot or step-wise, autoregressive trajectory generation procedures, thereby indicating that there’s no clear winner among the two. One-shot approaches regress a full trajectory at once based on the past data, while the trajectory is sequentially generated for each step based on the past data and the previous predicted steps in autoregressive approaches. Due to their simplicity, one-shot approaches are less complex, but autoregressive approaches have more potential for self-supervision and the incorporation of the ground truth data in order to extract more information from training data. It remains an open question if the performance of one-shot prediction models can be improved by reformulating the one-shot prediction in an autoregressive prediction. The prediction of the future is inherently uncertain and error-prone, because of this system limits need to be defined for safety-critical prediction applications, as autonomous driving. Despite this fact, uncertainty estimation techniques for vehicle trajectory prediction approaches have not yet been explored. This work aims to fill the mentioned research gaps. Firstly, a novel autoregressive trajectory prediction model, the Multi-Branch Self-Supervised Action-Space Predictor (Multi-Branch SSASP) is presented. The Multi-Branch SS-ASP is derived from a one-shot model, which contributes to the first research gap. The Multi-Branch SS-ASP is based on branched overshooting, an additional training on predictions, starting at intermediate future steps through which the potential of the available training data is better exploited. It incorporates a novel interplay between context aggregation and context prediction, and an advanced strategy to combine multi-modal predictions of successive steps. In the context of the Multi-Branch SS-ASP, both known and novel uncertainty estimation techniques are investigated in order to contribute to the second research gap. The Multi-Branch SS-ASP achieves state-of-the-art results on the INTERACTION dataset with a simpler encoder compared to previous state-of-the-art models, which rely on complex graph-structureleaveraging encoders. Additionally, significant positive correlations between the prediction error and two novel uncertainty estimation metrics are found, which might provide a route for the determination of confidence in generated predictions.