A GMM-COX Hybrid Modeling Framework for Multidimensional Retirement Prediction in the Public Sector

Abstract

Forecasting retirements is a strategic challenge for human resource management in public administrations, where workforce planning and service continuity depend on reliably anticipating these departures. In practice, retirement decisions are often based on limited criteria, such as age or seniority, without systematically considering the various factors that can influence their occurrence. This work proposes a multidimensional predictive approach based on the use of a dataset integrating demographic, administrative, medical, behavioral, and legal variables related to civil servants. To enable the combined use of qualitative and quantitative variables within a homogeneous analytical framework, a probabilistic coding method for qualitative categories is introduced, ensuring the normalization and consistent integration of data into statistical models. The methodology adopted is based on a hybrid approach combining a Gaussian mixture model to identify latent risk profiles and a Cox regression model to estimate risk and probable time to retirement. This combination makes it possible to exploit both the latent structures present in the data and the explicit relationships between the explanatory variables and the event under study. Human resource planning in the civil service is a significant challenge, particularly for anticipating retirements and ensuring the continuity of public service. In practice, retirement decisions are often based on limited criteria such as age and seniority, which fails to consider all the factors that could influence this event. In this study, a multidimensional predictive approach was proposed by integrating demographic, administrative, medical, and behavioral variables from employee records. The methodology is based on a hybrid approach combining a Gaussian mixture model for profile identification and a Cox regression model for estimating risk and probable time to retirement, producing consistent results and satisfactory predictive capability. These results show that the integration of multidimensional variables and the joint use of machine learning and survival analysis methods constitute a relevant approach to improve the anticipation of departures and support decision-making in public human resource management.