Predicting academic performance using trace data from learning management systems is a primary research topic in educational data mining. An important application is the identification of students at risk of failing the course or dropping out. However, most approaches utilise past grades, which are not always available and capture little of the student's learning strategy. The end-to-end models we implement predict whether a student will pass a course using only navigational patterns in a multimedia system, with the advantage of not requiring past grades. We experiment on a dataset containing coarse-grained action logs of more than 100,000 students participating in hundreds of short course. We propose two approaches to improve the performance: a novel encoding scheme for trace data, which reflects the course structure while remaining flexible enough to accommodate previously unseen courses, and unsupervised embeddings obtained with an autoencoder. To provide insight into model behaviour, we incorporate an attention mechanism. Clustering the vector representations of student behaviour produced by the proposed methods shows that distinct learning strategies specific to low- and high-achievers are extracted. [For the full proceedings, see ED615472.]