The production of electrical components for vehicles, battery systems, and power electronics components using ultrasonic metal welding is now well established. The industrial applications take advantage of the large joining surface and the associated excellent electrical and thermal conduction properties. However, this also prevents the weld quality from being monitored based on resistance measurements. For this reason, mechanical properties are used as a substituting quality criterion for process optimization and control. The type of mechanical testing depends on the component geometry, resulting in different quality features not necessarily comparable to each other. The ultrasonic metal welding process takes place in successive phases, which are characterized by the varying relative movements of the parts to be joined and the welding tools, prone to changes of material hardness, surface cleanliness, and geometric deviations. In this study, we investigate the influence of different mechanical quality features as target parameters for quality prediction on prediction accuracy and robustness. For this purpose, we consider the joint formation and the mechanical strength over the welding time for different quality parameters, fracture behaviors, and joint geometries. Supervised machine-learning quality prediction models are developed for the respective processes. The predictions are based on external and internal welding machine sensors. In addition to the sensor configuration, the importance of the individual process phases as a data source for the prediction accuracy is also investigated. We show that the overall process predictability strongly depends on the quality of the quality feature itself. Combining different sensor signals improves prediction quality. Furthermore, an analysis of the first process phases is sufficient for a robust quality prediction.