AI-Driven Employee Turnover Warning for Talent-Retention Decision Support in Finance and Taxation Education: Model Development and Comparison

Employee-turnover warning in finance and taxation education must account for different job roles, repeated monthly records, rare departures, and uncertainty about when some fields became available. This study analyzed 1,397 de-identified employee records and reconstructed 6,935 monthly records for predicting voluntary turnover within 90 days. It compared 22 classifiers with the Hierarchical Risk Encoding XGBoost–CatBoost Ensemble (HRE-XCB), which combines Raw-XGB, HRE-XGB, and native CatBoost. Before model development, 278 employees were selected without reading their out comes. None of their earlier records entered feature screening, tuning, encoding, weighting, or cutoff selection. The search evaluated 152 base structures and selected weights of 0.65, 0.15, and 0.20. The June 2022 evaluation contained 278 separated employees and 24 departures. HRE-XCB achieved recall 0.7500, F1 0.3750, and PR-AUC0.3477. Among the equally tuned shortlist, random forest obtained the highest PR-AUC (0.3772), Extra Trees the highest F1 (0.4324), and histogram gradient boosting recall 0.7917 with the lowest Brier score (0.0752). A separate lower-tuning screen was led by preset gradient boosting (PR-AUC 0.4229). In 5,000 employee-cluster bootstrap samples, no tuned model had a PR-AUC difference from HRE-XCB whose 95% interval excluded zero. Component ablation showed that Raw-XGB alone had nearly the same PR-AUC as the complete ensemble (0.3481 versus 0.3477), while Raw-XGB plus HRE-XGB reached 0.3492. The study therefore supports employee-separated evaluation and decision-specific model choice, but not a universal-best algorithm or exact individual departure probabilities.