Background: Routine semen records contain repeated observations from the same bull, but their temporal structure is often underused when predicting subsequent semen quality. This study developed an explainability-oriented longitudinal prediction framework in which each predictor was traceable to reproductive information available before the target collection and compared transparent historical features with fixed machine-learning models. 
Methods: A retrospective longitudinal dataset comprising 57 semen collection records from seven artificial insemination bulls was subjected to predefined quality control. Eight observations were excluded and the first collection for each bull was unavailable for lag-based prediction, yielding 42 modelling observations. Semen quality was defined as high quality when total motility was ≥70% and lower quality when total motility was <70%. Longitudinal feature engineering generated immediate, rolling and cumulative historical reproductive predictors using only information available before each target collection. Simple historical indicators and Logistic Regression, Random Forest and XGBoost models were evaluated using Leave-One-Bull-Out Cross-Validation (LOBOCV). Explainability was addressed through transparent feature definitions, temporal traceability and comparison with parsimonious models; no post-hoc explainability method was used. Matthews correlation coefficient (MCC) was prespecified as the primary model-selection criterion. 
Results: The historical mean of previous total motility produced the strongest verified simple prediction, with 71.4% accuracy, 67.5% balanced accuracy and MCC = 0.391. Previous total motility achieved 66.7% accuracy and MCC = 0.296, whereas the three-collection rolling mean achieved 64.3% accuracy, 59.4% balanced accuracy and MCC = 0.211. Logistic Regression using previous total motility achieved MCC = 0.352, and adding previous progressive motility did not improve MCC. The combination of previous total motility and historical mean total motility produced MCC = 0.060. Under the fixed longitudinal machine-learning specifications, enhanced Random Forest and XGBoost models achieved MCC values of −0.247 and −0.224, respectively, while enhanced Logistic Regression achieved MCC = −0.407. Conclusion: Longitudinal feature engineering extracted a measurable predictive signal from historical reproductive records, but the strongest verified performance was obtained from transparent cumulative historical information rather than from increasingly complex machine-learning models. The findings support an interpretable decision-support pathway in which routinely accumulated semen records are transformed into temporally valid bull-level performance profiles, while emphasising that validation across substantially larger numbers of independent bulls is required before operational deployment.