The Swedish Iron ore line is subjected to sub-arctic weather conditions, high axle loads and a large yearly tonnage. These operational conditions challenge the usefulness of existing deterministic deterioration models for railway infrastructure that are validated in other contexts. However, more frequent and higher precision of condition measurements, combined with statistically based prediction methods, may offer a viable solution. Here, we study remaining useful life predictions of railway track geometry properties based on recursively updated time series. We discuss how data-driven models are affected by measurement errors of track properties, such models’ ability to detect seasonal effects, and how they are affected by irregular sampling. The prediction abilities and uncertainty measures for different modelling approaches are also compared.