Data-Driven Approach in HR-Engineering in the Context of Personnel Turnover Management

Background: High employee turnover imposes severe operational and financial costs on modern enterprises. In volatile labor markets, traditional reactive human resource management fails to anticipate workforce attrition before human capital depletion occurs. Objective: This study develops and evaluates an interpretable predictive decision-support framework using gradient boosting algorithms (XGBoost and CatBoost) to forecast voluntary employee turnover risk. Methods: Using an open secondary HR analytics benchmark dataset (N = 14,999) encompassing performance metrics, working hours, project loads, tenure, and compensation levels, continuous probabilistic risk scoring models were trained and calibrated. Performance was evaluated on a 70/30 train-test split using MAE, MSE, RMSE, and R² metrics alongside kernel density estimation (KDE) feature analysis. Results: Both gradient boosting models achieved high accuracy on unseen test data, with CatBoost (R² = 0.9359, MAE = 0.0392, RMSE = 0.1078) demonstrating slightly superior calibration and explanatory power compared to XGBoost (R² = 0.9305, MAE = 0.0436, RMSE = 0.1123). Non-linear distributions revealed that extreme monthly working hours, project overload or underload, and depleted job satisfaction constitute key risk indicators. Conclusion: Embedding continuous risk scoring into organizational dashboards enables management to transition from reactive exit interviews to proactive talent retention, providing a scalable decision-support architecture for contemporary business informatics.