There is a significant amount of recent literature in the domain of self-driving vehicles (SDVs), with researchers focusing on various components of the SDV system that have the potential to improve on-street performance. This includes work exploring improved perception of the SDV surroundings, or proposing algorithms providing better short-term prediction of nearby traffic actor behavior. However, in most cases, the authors report only aggregate metrics computed on the entire data, and often do not fully consider the bias inherent in the traffic data sets. We argue that this practice may not give a full picture of the actual performance of the prediction model, and in fact, may mask some of its problem areas (e.g., handling turns). We analyze the amount of bias present in traffic data and explore the ways to address this issue. In particular, we propose to use a novel off-road loss and standard bias mitigation techniques that result in improved performance. We further propose to avoid aggregate metrics and instead analyze performance on relevant
subsets of the data, thus better capturing actual model capabilities. Moreover, we propose to measure a novel off-road error to complement commonly used prediction metrics. Extensive analysis of real-world data suggests benefits of the proposed approach for improving the performance of SDV technology.