Although geomagnetic field data is available in real time, it is desirable to have warning of potential hazard at least 10 minutes in advance so as to provide stakeholders with the opportunity to take mitigating actions. Towards this end, we have begun the development of a data-driven, machine-learning enabled forecasting system whose ultimate goal is to provide a high-quality, uncertainty-quantified prediction of geoelectric fields. Our system combines modern approaches from non-stationary signal processing, information theory, ensemble forecasting, and feature classification with high-quality geomagnetic data sets and three-dimensional electromagnetic transfer functions. In this talk, I will provide an overview of our modeling effort and how discuss how physical insights have driven the development of our machine learning approach.