Employee turnover remains a critical challenge for organizations because it erodes human and social capital and generates substantial replacement costs. While the literature highlights job attitudes, social exchange, and organizational justice as key antecedents of turnover intention, HR departments still lack calibrated, decision-oriented tools for targeting scarce retention resources. This study develops and evaluates calibrated turnover risk models to support capacity-constrained retention triage. Using an anonymised HR dataset of employees in a large service organization, we estimate logistic regression and random forest models and then apply probability calibration, Brier score decomposition, and expected calibration error to assess probability reliability. We combine discrimination and calibration results with decision curve analysis under realistic capacity constraints and translate alternative operating points into workload and outcome metrics, such as alerts per day and true-positive and false-positive burdens. The findings show that a calibrated logistic model preserves high discrimination while substantially improving probability calibration and delivering higher net benefit than uncalibrated alternatives across plausible capacity ranges. Scenario analysis demonstrates how HR managers can select thresholds that align with available staff time and acceptable error trade-offs. The study contributes a calibration-first pipeline for turnover risk modelling and provides actionable guidance on integrating HR analytics into retention decisions in a transparent, capacity-sensitive manner.
