Beyond Grades: Temporal Validation of a Multi-Output Model for Formative Academic Indicators in a Latin American Public University

Learning-platform records are often assembled after a course ends, obscuring whether predictors were available at an earlier decision point. We reconstructed student-course-period observations from pseudonymized assignment, assessment, learning-management-system, and forum events across nine official undergraduate periods (2020-1 to 2024-1) at a Mexican public university. For 7963 candidate observations, 16 predictors used events at or before 25%, 50%, or 75% of each period; five formative outcomes were defined on assignments due and examinations started after the cutoff. A shared-trunk multi-output neural network was evaluated with grouped cross-validation, course-held-out validation, and a retrospective 2024-1 holdout. Only future assignment non-submission showed a consistent aggregate signal (grouped cross-validation R2 = 0.217, 0.253, and 0.147). At 50%, leave-one-course-out R2 was 0.211, but median course-level R2 was −0.146. On the 2024-1 holdout, R2 was 0.144 (95% confidence interval [−0.328, 0.335]; n = 471), below a prior-submission persistence rule (R2 = 0.324). Other outcomes lacked stable performance, and prior examination accuracy approximated future accuracy better than the model. The findings support a limited retrospective association, not reliable course-level prediction, early-warning deployment, or intervention effectiveness. The contribution is a temporally explicit reconstruction and validation workflow for an underrepresented Latin American context.