Global university rankings differ in methodology, yet their published positions are frequently compared across systems and monitored over time. This study examines two dimensions of ranking stability: cross-sectional consistency across ranking systems and temporal persistence within QS, while assessing whether machine-learning models improve predictive accuracy beyond a persistence baseline. The analysis included 57 universities from 20 countries and seven regions. Cross-ranking consistency among QS-2026, THE-2026, and ARWU 2025 was evaluated using pairwise rank correlations, bootstrap confidence intervals, and normalized percentile positions. Temporal predictability was assessed on a balanced panel of 40 universities with QS positions for 2021–2024 and 2026 using an expanding-window design, a persistence baseline, linear regression, Random Forest, XGBoost, and LightGBM. Pairwise correlations were positive within observed overlaps, but the ranking systems were not interchangeable. The persistence baseline achieved the lowest development MAE. On the final two-year holdout, Random Forest (MAE 5.073) did not significantly outperform persistence (MAE 5.150; p = 0.902), and the latest QS position was the only predictor with consistently positive importance. Historical persistence errors bracketed the 2024–2026 result, indicating that predictability was driven mainly by ranking persistence, whereas additional machine-learning complexity provided limited benefits in this sample.

The early detection of psycho-emotional risks remains challenging despite its growing importance. Most existing models rely on a single data type, mainly questionnaires or text, and operate as “black boxes”, limiting practical use. Psycho-emotional states are multidimensional, reflected in textual, structured, and temporal digital signals. Ignoring this complexity may result in information loss and lower prediction accuracy. This paper presents an explainable text-based computational framework for psycho-emotional risk classification using public datasets. Record-level structured information was used only when it was available within the same original observation and was not created by matching records across datasets. The study did not integrate observations across datasets or perform cross-dataset fusion; these aspects are considered potential directions for future research. Implemented in Python using PyTorch, the model was evaluated on two open-text datasets. The BERT-based configuration achieved accuracy of 0.593 ± 0.009 and an ROC-AUC of 0.860 ± 0.006, while RoBERTa-base improved the performance to 0.666 ± 0.005 accuracy and a 0.896 ± 0.004 ROC-AUC under five-fold cross-validation. The results demonstrate that classification quality depends on encoder selection, preprocessing, dataset characteristics, duplicate handling, calibration, and the way that available record-level information is represented. Therefore, the findings should be interpreted as record-level computational classification results rather than the clinical validation of psycho-emotional risk assessment.