This paper surveys the state of the art on prediction in MOOCs through a Systematic Literature Review (SLR). The main objectives are: (1) to identify the characteristics of the MOOCs used for prediction, (2) to describe the prediction outcomes, (3) to classify the prediction features, (4) to determine the techniques used to predict the variables, and (5) to identify the metrics used to evaluate the predictive models. Results show there is strong interest in predicting dropouts in MOOCs. A variety of predictive models are used, though regression and Support Vector Machines stand out. There is also wide variety in the choice of prediction features, but clickstream data about platform use stands out. Future research should focus on developing and applying predictive models that can be used in more heterogeneous contexts (in terms of platforms, thematic areas, and course durations), on predicting new outcomes and making connections among them (e.g., predicting learners' expectancies), on enhancing the predictive power of current models by improving algorithms or adding novel higher-order features (e.g., efficiency, constancy, etc.).

The advancement of learning analytics has enabled the development of predictive models to forecast learners’ behaviors and outcomes (e.g., performance). However, many of these models are only applicable to specific learning environments and it is usually difficult to know which factors influence prediction results, including the predictor variables as well as the type of prediction outcome. Knowing these factors would be relevant to generalize to other contexts, compare approaches, improve the predictive models and enhance the possible interventions. In this direction, this work aims to analyze how several factors can make an influence on the prediction of students’ performance. These factors include the effect of previous grades, forum variables, variables related to exercises, clickstream data, course duration, type of assignments, data collection procedure, question format in an exam, and the prediction outcome (considering intermediate assignment grades, including the final exam, and the final grade). Results show that variables related to exercises are the best predictors, unlike variables about forum, which are useless. Clickstream data can be acceptable predictors when exercises are not available, but they do not add prediction power if variables related to exercises are present. Predictive power was also better for concept-oriented assignments and best models usually contained only the last interactions. In addition, results showed that multiple-choice questions were easier to predict than coding questions, and the final exam grade (actual knowledge at a specific moment) was harder to predict than the final grade (average knowledge in the long term), based on different assignments during the course.