AI-Enabled Attrition Prediction Using Calibrated Boosting

Employee attrition presents a significant threat to organizational continuity, as frequent turnover depletes valuable intellectual capital and incurs heavy recruitment costs. Although ensemble machine learning models often achieve high accuracy in predicting employee departure, they tend to generalize poorly and yield poorly calibrated probability estimates. In high-stakes human resources (HR) decision-making, miscalibrated models produce overconfident or underconfident risk scores that do not reflect true exit likelihoods, leading to suboptimal intervention strategies and resource misallocation. This paper proposes a unified approach combining a Gradient Boosting classifier with post hoc temperature scaling, to meet the dual needs of prediction strength and reliability. Experiments show that the proposed approach consistently improves probability calibration across all datasets; however, on the local dataset, the underlying predictive signal remains weak, so the resulting risk scores should be interpreted as modestly informative rather than strongly discriminative. Comparative scores with other calibration methods are also presented. The study underscores that well-calibrated risk scores are crucial for converting predictive outputs into actionable insights, allowing quantitative analysis to effectively support human judgment in workforce planning.