With the prevalence of technology found in modern society, the potential impact from strong geomagnetic storms cannot be underestimated. The primary drivers of these storms, coronal mass ejections (CMEs), are large explosions of solar material that are capable of being Earthward directed. Their effects have been well-studied in the literature, primarily by astrophysicists and others concerned with space weather. Because of the opportunity to obtain data about these phenomena, many studies have been done that build empirical models to predict if an impending CME will cause a strong geomagnetic disturbance. Two main types of data are typically used: data collected at the Sun and right before a CME's arrival to Earth. The former results in more timely predictions but less accuracy while the latter delivers better model performance but leaves much less time to prepare on Earth. Adequately dealing with this trade-off is still a growing area of research. In addition, creating and implementing advanced ensemble models for prediction and inference tasks based on machine and statistical learning theory have been lacking. Hence, this dissertation focuses on positing such models as well as present solutions for the trade-off problem. The first work presents a new ensemble model based on random forests (RFs) to classify geoeffective CMEs using both types of data. This approach not only makes competitive predictions but also provides a reliable way to investigate the importance of the predictor variables (or features). The second work establishes a two-stage meta-learning framework that uses the first type of data in the first stage and both types in the second. The postulated method focuses on the trade-off problem, which is not addressed in the first work. The third work seeks to improve initial CME classification by incorporating CME image data. These images are introduced into a convolutional neural network (CNN) to create a set of deep learning features that increase the predictive power of models that only use the first type of data. While the domain is specific to geomagnetic storm prediction, the methods proposed can be applied to a variety of prediction problems in other fields.