Вероятностные модели для прогнозирования кривой удержания сотрудников

Employees are important assets of any organization, and losing them might lead to decreased human capital and operational disruptions. Companies strive to apply predictive analysis to forecast potential employee churn to prevent it. Machine learning practices have been used for churn forecasting (Ozcan & Ozmen, 2021); however, they imply that big microdata is available to employees. Most employee churn studies were not accidentally based on synthetic rather than real-world data (Alamsyah & Salma, 2018). In addition, machine learning models lack full transparency. On the contrary, probabilistic models require little data and are based on fully transparent data-generating processes. The topic of probability-based modeling for projecting employee churn is barely covered in HR analytics. The shifted-beta-geometric model (sBG) (2007) and the beta-discrete-Weibull model (BdW) presented by Peter Fader and Bruce Hardie (2018) were previously applied only within the customer analytics field. In this paper, we test if these models are suitable for employee survival projection based on three cohorts of employees. The results of the models are compared with each other as well as with linear and logarithmic regressions. The key finding is that the BdW model outperformed other models thanks to the ability to capture the heterogeneity within the sample (e.g., in survival rates, in individual’s behavior) in continuous time-setting. Survival rates tend to vary along with time; thus, the Weibull distribution allows us to catch nuances in changing employee behavior on the individual level over time. Our research provides academically and managerially relevant conclusions, e.g., that probabilistic models applied in customer analytics and requiring little data can be used in HR analytics by practically any organization. Keywords: employee churn, churn prediction, probabilistic models, beta distribution, geometric distribution, Weibull distribution