Saccadometry, through the high-resolution measurement of eye movements, provides an objective means of quantifying abnormalities in movement disorders such as Parkinson’s disease (PD) and progressive supranuclear palsy (PSP). Previous studies have found that saccades are altered in the presence of these diseases, and change with their progression and treatment. In this study, deep convolutional neural networks (CNNs) were used to encode and classify saccades based on shape alone, and were evaluated on existing sac- cadometric datasets from the ongoing Oxford Quantification In Parkinsonism (OxQUIP) study, comprising 8 PSP patients, 9 PD patients, and 10 healthy controls. A traditional analysis of saccade metrics was also conducted: in agreement with previous studies of similar size, there were no significant differences between groups in terms of prosac- cadic or antisaccadic latency, but PSP patients had elevated directional error rates. An autoencoder network produced 6-dimensional encodings of saccadic trajectories which represented them more faithfully than a best-fit sigmoid curve. A classifier network was then trained to predict patient groups from the 6-dimensional representations of their saccades, and had better-than-chance accuracy. Another network trained to differentiate PD patients from controls performed well on this task in 3-fold cross-validation (mean ROC-AUC 0.926). The present study thus provides proof of concept for the application of deep CNNs to raw saccadometric data, and demonstrates that it is possible for these networks to extract information about disease states. Future models based on those developed here may assist with diagnosis of movement disorders, monitoring of disease progression, and evaluation of candidate treatments.