Next-generation networks are envisioned to be empowered by artificial intelligence with predictive capabilities. Predicting handovers in high mobility scenarios enables networks and applications to adapt ahead of time to improve the Quality of Service (QoS). In this paper, we present a two-step machine learning (ML) method, consisting of a classifier and regressor, that can predict the remaining time until a handover occurs. Our approach is validated on a dataset that was captured in a real cellular network. The results show that upcoming handovers can be detected with a recall above 90% and the timing of handovers with an error smaller than one second. Furthermore, we compare the importance of input features derived from radio conditions and user locations for the ML models and discuss deployment scenarios of our approach. In particular, our results suggest that cell-based models perform better than models trained for larger areas.