The potential of using neural data to predict academic outcomes has always been at the heart of educational neuroscience, an emerging field at the crossroad of psychology, neuroscience, and education sciences. Although this prospect has long been elusive, the exponential use of advanced techniques in machine learning in neuroimaging may change this state of affairs. Here we provide a review of neuroimaging studies that have used machine learning to predict literacy and numeracy outcomes in adults and children, in both the context of learning disability and typical performance. We notably review the cross-sectional and longitudinal designs used in such studies, and describe how they can be coupled with regression and classification approaches. Our review highlights the promise of these methods for predicting literacy and numeracy outcomes, as well as their difficulties. However, we also found a large variability in terms of algorithms and underlying brain circuits across studies, and a relative lack of studies investigating longitudinal prediction of outcomes in young children before the onset of formal education. We argue that the field needs a standardization of methods, as well as a greater use of accessible and portable neuroimaging methods that have more applicability potential than lab-based neuroimaging techniques.