Financial sustainability in higher education is increasingly fragile due to policy shifts, rising costs, and funding volatility. Legacy early-warning systems based on static thresholds or rules struggle to adapt to these dynamics and often overlook fairness and interpretability—two essentials in public-sector governance. We propose a university financial risk early-warning framework that couples a causal-attention Transformer with Multi-Objective Bayesian Optimization (MBO). The optimizer searches a constrained Pareto frontier to jointly improve predictive accuracy (AUC↑), fairness (demographic parity gap, DP_Gap↓), and computational efficiency (time↓). A sparse kernel surrogate (SKO) accelerates convergence in high-dimensional tuning; a dual-head output (risk probability and health score) and SHAP-based attribution enhance transparency and regulatory alignment. On multi-year, multi-institution data, the approach surpasses mainstream baselines in AUC, reduces DP_Gap, and yields expert-consistent explanations. Methodologically, the design aligns with LLM-style time-series forecasting by exploiting causal masking and long-range dependencies while providing governance-oriented explainability. The framework delivers earlier, data-driven signals of financial stress, supporting proactive resource allocation, funding restructuring, and long-term planning in higher education finance.

Introduction: Learning analytics models commonly produce predictive scores without specifying how uncertainty should inform prioritization when review capacity is limited. Methods: This study developed a reproducible methodological workflow integrating predictive benchmarking, global and normalized adaptive split conformal calibration, hypothetical capacity-constrained Top- K screening, threshold-relative analytical routing, and deterministic audit recording. The workflow was evaluated on a public synthetic student performance benchmark containing 6,607 records, 19 mixed-type predictors, and one numerical examination score outcome. Seven regression models were compared over five repeated 60%/20%/20% train-calibration-test partitions. Results: Regularized linear regression produced the most favorable observed mean performance, with a mean absolute error of 0.502 ± 0.016, a root mean squared error of 2.104 ± 0.149, a coefficient of determination of 0.713 ± 0.030, and a Spearman correlation of 0.976 ± 0.003. Global and adaptive conformal procedures both achieved complete-test empirical marginal coverage close to the nominal 90% level. Across all seven models, adaptive calibration generated record-varying intervals, reduced mean interval width and interval score, increased the mean split-wise minimum coverage across the examined encoded categories, and reduced mean coverage disparity, although several category-level estimates remained below 0.90. The common radius used by global calibration preserved the point-prediction rankings, whereas adaptive scores changed Top- K membership. However, the resulting changes in Precision@ K , Capture@ K , and FN@ K were numerically small at the reported precision and showed no consistent screening advantage. Training-quantile-derived robustness analyses supported the same interpretation. Discussion: Within this synthetic benchmark, adaptive calibration contributed primarily through uncertainty differentiation and interval efficiency rather than improved retrospective screening utility. The study demonstrated computational reproducibility and record-level traceability; it did not validate real student screening, establish category- or selection-conditional coverage, or support automated or causal intervention decisions.

High classification performance in financial-risk early-warning research may reflect recovery of a constructed contemporaneous rating rule rather than prediction of an independent future outcome. This study distinguishes these two forms of evidence through a controlled synthetic-panel experiment situated in the context of higher education institutions (HEIs). Task 1 is a contemporaneous positive-control audit in which the four components defining a deterministic three-class rating are supplied to the classifiers. Task 2 uses predictors measured at year $t-1$ to classify the rating at year t through strict rolling-origin evaluation under prespecified Weak, Moderate, and Strong temporal-persistence conditions. In Task 1, the four fitted classifiers achieved held-out Macro-F1 values of 0.9968–1.0000, demonstrating near-complete recovery of the disclosed rating rule but providing no prospective forecasting evidence. In Task 2, the best mean annual Macro-F1/Macro-AUC increased from 0.480/0.696 under Weak persistence to 0.549/0.747 under Moderate persistence and 0.633/0.824 under Strong persistence. The Weak–Moderate–Strong ordering was observed for both primary metrics across all four fitted classifiers, while higher model complexity provided no consistent advantage within the disclosed synthetic mechanism. These findings confirm contemporaneous rule recoverability and sensitivity to deliberately embedded temporal persistence only within the controlled experiment. They do not establish predictive validity, transportability, or decision benefit in real HEIs. The study contributes a reproducible task-to-claim approach that aligns target construction, predictor–target overlap, information timing, rolling-origin evaluation, probability quality, and model performance with the inferences that the resulting evidence can legitimately support. Real-world validation would require source-indexed longitudinal HEI data, independently adjudicated post-origin outcomes, verified information-availability dates, external testing, and prospective decision evaluation.