Computer simulation models have the potential to aid decision-making in many industries. They are particularly beneficial when applied to situations that would be too costly or impractical to experiment with in reality. In the railway domain, such applications may include making infrastructure changes or updating a timetable. A simulation is a representation of a system or object, not an exact copy. However, to be a meaningful planning aid, the output of a simulation needs to be sufficiently realistic. A review of existing railway simulation software identifies two features where established modelling techniques limit the realism of the output: train movements and signalling decisions. This thesis develops a stochastic simulation that models train movements and signalling decisions using supervised machine learning. Train movements are modelled using quantile regression. The models are applied to replicate the distribution of travel durations without requiring any underlying assumptions about the properties of the distribution. Probabilistic classification is employed to predict signallers’ actions, achieving accuracies of over 89%. Experiments were conducted by running the stochastic simulation for a week’s worth of activity on a section of the British network for 600 iterations. The outputs were compared to the true activity of the selected week. The true total travel duration of all the trains fell within the simulated minimum and maximum values achieved across all iterations. Moreover, the median total travel duration across all iterations settled to within 1% of the true value. The results demonstrate that machine learning-assisted simulation modelling can be applied to make realistic performance assessments in the railway domain. There is the possibility to use such a model to compare alternative train regulation policies or to investigate the performance of a new timetable.