A Machine Learning Approach to Predict Voluntary Separation from the Australian Army

Abstract

The Australian Army faces significant workforce growth challenges over the next decade. Our study uses machine learning (ML) with human resource data to predict individual voluntary separation decisions to support retention targets. We use a two-stage approach. Stage 1 predicts “who separates” through supervised classification. Stage 2 predicts “when separation occurs” through predictive survival analysis. We evaluate the best performing models, and which features were most important to predictive performance. In Stage 1 we find that tree ensemble and gradient-boosted model perform exceptionally well. The best performing model, Light Gradient-Boosting Machine, achieves an Average Precision of 0.99. In Stage 2, Random Survival Forest accurately predicts separation timing, achieving an Uno’s Concordance Index of 0.97. Unemployment rates and job-search duration consistently dominated feature importance analyses. However, we observe non-linear and non-monotonic relationships that contradict standard labor economic theory, suggesting broader structural factors or events may influence predicted separation risk. These findings show that ML is effective in processing large and complex datasets for predicting separation risk and identifying key drivers for further study and policy intervention.