Our project is motivated by the early dropout and low completion rate problem in MOOCs. We have extended traditional features for MOOC analysis with richer and higher granularity information to make more accurate predictions of dropout and performance. The results show that finer-grained temporal information increases the predictive power in the early phases of the Pattern-Oriented Software Architectures (POSA) MOOC offered in summer 2013 by Vanderbilt University. As a next step, we plan to develop unsupervised learning methods with our extended feature set to define profiles that can be used for effective scaffolding and feedback.