The prediction of academic failure is one of the most pressing concerns for researchers in the field of education. Research in the field of E-learning has predominantly prioritized the precision and accuracy of machine learning models designed for specific courses. However, there is a dearth of studies exploring the transferability and generalization of prediction models from a source course to other courses. By portability of machine learning models, we generally mean the ability of a model to be used or deployed in different environments and platforms. Many factors can affect the portability of a model. The objective of this study is to investigate the portability of models obtained directly from the Moodle logs of 15 courses hosted in the Moodle platform of Ibn Tofail University. The approach used aims at verifying whether the number and level of use of activities provided by the Moodle logs, and whether the use of numerical or categorical attributes affects the portability of predictive failure models in educational contexts. We used the KNN classification algorithm on a dataset of courses of the same level to obtain a model and test their portability to other courses by evaluating accuracy and loss of accuracy. The results show that the portability of the predictive models and their application to different courses is possible in some cases but with an acceptable loss of accuracy.