Volumetric arc therapy (VMAT) is a type of cancer treatment where radiation is delivered continuously from a moving beam source. Due to potential delivery errors incurred by continuous delivery, pre-treatment dosimetric verification is regularly performed to measure beam deliverability by quantifying agreement between delivered and planned dose distribution. Using a dataset of VMAT beams and past verification results, we present a statistical learning approach to predict the deliverability outcome. A regression modelling pipeline with modular components of prediction models and resampling methods is trained and evaluated on three feature sets. Classification and outlier detection approaches are also applied after dichotomizing the deliverability outcome. The results show some evidence of learning, but performances are poor due to limitations of the dataset and the complex nature of the underlying process. We also propose a novel visual beam descriptor involving the Radon transform and present a preliminary image-based predictive pipeline.