Human Capital Governance and Employee Attrition Prediction: A Theory-Informed People-Analytics Framework Using Synthetic IBM HR Data

Voluntary employee attrition creates recruitment, onboarding, productivity, and knowledge-continuity costs, which motivates people-analytics approaches that connect human-capital theory with interpretable prediction. This study presents a theory-informed proof-of-concept framework using the IBM HR Analytics Employee Attrition and Performance dataset, which contains 1,470 synthetic employee records rather than observations from a real organization. Six composite measures represent well-being, burnout risk, compensation fairness, governance, career stagnation, and engagement. The analytical pipeline combines logistic and ordinary least squares regression, regularized classifiers, Random Forest, XGBoost, SHAP-based model explanation, subgroup analysis, and an exploratory mediation specification. In the attrition model, the Well-Being Index, Governance Score, Engagement Score, and Compensation Fairness Proxy are associated with lower attrition odds within this synthetic dataset, whereas overtime and career stagnation are associated with higher odds. The theory-informed classification models achieve held-out AUC values from 0.756 to 0.770, while an un-engineered raw-feature logistic baseline reaches 0.778, so the composite indices are interpreted as a theoretically organized, governance-oriented representation rather than a source of superior predictive accuracy. The monthly-income model is treated only as a diagnostic illustration because the Compensation Fairness Proxy is constructed from Monthly Income and cannot provide an independent test of income prediction. The exploratory mediation results do not support an indirect well-being-to-attrition pathway through overtime under the Sobel test. The framework shows how theory-derived constructs, statistical inference, prediction, and explanation can be combined, while real organizational HRIS data, leakage-controlled preprocessing, construct validation, and stronger indirect-effect testing remain necessary for external validation.