The selection of a model for academic risk prediction systems is usually based on the global performance of the model. However, this global performance is not an important factor for the end-user of the system. For the end-user, the performance of the model for his or her specific case is the most important aspect of that model. Given that the model is usually selected at design time, the end-user could end up with a sub-optimal prediction. To solve this problem, this work presents a conceptual framework to build adaptive multilevel clustering models for academic risk prediction. This frameworks allows the system to automatically select between several levels of hierarchical or semi-hierarchical features to create a clustering model to best predict the particular risk of each student. This conceptual framework is validated through its realization into an adaptive model to predict the risk of failing a course during a semester in a Computer Science program. In this study, the adaptive model consistently outperforms the prediction of the best-performing static model.