This projects outlines how to choose a good eye movement dataset and evaluate a deep learning approach to eye movement classification. A thorough investigation of the annotation of the GazeCom dataset was performed. By looking at different feature distributions and event durations of eye movements it was possible to show that the annotations did not comply with the physiological properties of said movements. A similar investigation was performed on the Lund 2013 dataset and this showed that the features were in agreement with physiological properties. A 1D Convolutional Neural Network Bidirectional Long Short-Term Memory (1D-CNN-BLSTM) neural network was trained on the Lund 2013 dataset to classify fixations, saccades, smooth pursuit and post-saccadic oscillation. Different model parameters and eye movement features were tested. The best performing model was a multi resolution 1D-CNN-BLSTM but it as not outperforming the other neural networks by much. The biggest challenge was differentiating between fixations and smooth pursuit and this was not solved.