Faster delivery of data is important for distributed scientific workflows. Predictability is equally (or even more) important.
Machine learning methods are applied to develop predictive models for data transfer times over a variety of wide area networks.
201,388 transfers, involving 759 million files totaling 9 PB transferred over 115 heavily used source-destination pairs between 135 unique endpoints, are used to build and evaluate these models. 
A detailed analysis of results is presented to insight into the cause of some of the high prediction errors.