N°019_Vol.2_28

Introduction: Universities increasingly use educational data to move from retrospective reporting toward earlier forms of academic support. Grades, attendance records, assignment submissions, prior credits, and other traces can be assembled while a semester is still under way, giving teachers and academic services an opportunity to act before difficulties become irreversible (Long & Siemens, 2011; Romero & Ventura, 2020). However, the predictive performance achieved in a consolidated data set does not guarantee that the model would hold its accuracy when implemented in a new institution. A classifier can find patterns that are robust across different contexts, but it can also exploit local grading standards, program structures, or data collection methods that are not generalizable. This distinction matters in Kinshasa. Universities belong to the same national system of higher education, but vary in size, range of programs, number of students, administrative procedures, and degree of organization applied to academic knowledge.A model that looks satisfactory when all observations are mixed may hide uneven performance by institution. The practical question is therefore no longer only whether academic risk can be predicted, but whether the predictive relationship survives a change of university and whether the same variables continue to carry information in the new setting. A recent study conducted in Kinshasa showed that six-level academic-risk prediction can be developed from mid-semester educational records and linked to graduated support options (Tadiamba et al., in press). An important question nevertheless remains unresolved: whether the learned relationships remain stable when the model is transferred from one university to another. The present study addresses this gap by treating each institution as a separate validation domain and examining external performance, directional transfer, calibration, class-level errors, and institution-specific permutation importance. The research question is: to what extent does a fixed academic-risk model preserve discrimination, class balance, probability reliability, and explanatory structure when it is transferred among the University of Kinshasa (UNIKIN), National Pedagogical University (UPN), and the Protestant University of Congo (UPC)? Three objectives follow from this question. The first is to compare internal performance with performance obtained when an entire university is excluded from training. The second is to determine whether the dominant predictors are shared or institution-dependent. The third is to identify the classes and transfer directions in which errors become more pronounced. The approach relies on three operational hypotheses. The external macro F1 score for each institution will not deviate by more than 0.05 from the within-university cross-validation score. H2 indicates that ongoing evaluation, midterm performance, and already validated credits will continue to be the primary predictors across all three universities. The extreme categories of the ordinal scale (Risk and Excellence) will exhibit greater recall instability than the intermediate categories. These hypotheses are evaluated as empirical expectations rather than causal assertions. The contribution is institutional as well as methodological. The study methodologically transitions the unit of validation from an individual randomly selected student record to an entire institution, representing a more stringent assessment of portability. Institutionally, it provides evidence to determine if the universities of Kinshasa can initiate a unified predictive core or if entirely distinct models are required from the outset. The paper thereafter addresses methodology, results and analysis, as well as limitations and discussion.