Early identification of students who may experience unsuccessful academic outcomes can support timely academic guidance, provided that predictions are evaluated for both discrimination and reliability. This study evaluates a Logistic Regression workflow using the Open University Learning Analytics Dataset (OULAD). Student outcomes were converted to a binary target, with Pass and Distinction treated as successful outcomes and Fail and Withdrawn as unsuccessful outcomes. Models were evaluated at weeks 2, 4, 6, and 8 using student characteristics, enrolment information, and virtual learning environment (VLE) activity. A second configuration added assessment-related features. The best single-split result occurred at week 8 for the activity-plus-assessment configuration (accuracy = 0.7789, F1 = 0.7796, ROC-AUC = 0.8654). Its Brier score and log loss were 0.1463 and 0.4394, respectively. Across confidence thresholds, selective accuracy increased as coverage decreased; at week 8 and a threshold of 0.90, the activity-plus-assessment configuration retained 29.7% of cases with 97.1% accuracy. Across three random seeds, its week-8 mean ROC-AUC was 0.8629 (SD = 0.0029). These findings suggest that later cutoffs and assessment-related information can improve predictive performance in this dataset. However, the availability time of assessment scores must be verified to rule out temporal leakage, and results require external validation before practical deployment.