Temporal Out-of-Distribution Validation of Calibrated Academic-Risk Early Warning Models on OULAD Learning Traces
Year: 2026
Type: preprint
Abstract: Confirmatory secondary-data study testing whether stratified random cross-validation overstates the performance and probability reliability of early academic-risk models when evaluated on later OULAD course presentations. The unit is a student-module-presentation record. The primary analysis uses VLE and assessment information available during presentation-relative days 0-30, a presentation-order expanding rolling-origin evaluation, and a matched random out-of-fold control on the same future test rows. Logistic regression is the confirmatory primary model and random forest is a robustness replication. We report AUROC, AUPRC, Brier score, ECE, calibration slope/intercept, paired bootstrap confidence intervals, and Holm correction. The study does not claim causal intervention effects, individual identification, cross-institution generalisation, or universal model superiority.
Author: Wang Qiushi
Language: English
Subfield: Computer Science Applications
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7204690044