Accurate prediction of track geometry degradation is essential for an efficient track geometry maintenance planning and scheduling. Track geometry prediction is a complex task as many quantitative and qualitative parameters affect track geometry degradation. Artificial neural networks (ANNs) have shown a great capability in prediction of such complex systems, although they are not complicated. In this paper, a three-layered feedforward network is employed to predict track geometry degradation for each track section with-in a track line for a period of two years. A set of influencing factors along with track geometry degradation history are used as inputs to the model. The weight and bias values in the ANN model are optimised using Levenberg and Marquardt optimization algorithm.  Data from the Swedish railway network are used to train and verify the proposed model. The relative importance of input parameters on track geometry degradation is determined by Garson`s algorithm. The results indicate that ANN is able to predict the future state of the track in next two years even in the existence of tamping activity within the degradation history.