Who Leaves? Individual-Based Predictive Modeling of Non-End of Active Service Attrition for Enlisted Marines

Abstract

Talent Management 2030 posits that the United States Marine Corps’ manpower system hails from the industrial era and calls for broad modernization. This thesis serves as a proof of concept designed to implement modern predictive machine-learning algorithms and techniques to an age-old military manpower problem. Current Marine Corps attrition modeling is conducted using historical averages and does not account for individual attributes of each Marine. This study employs two machine-learning models, a Random Forest classifier and a multinomial logistic regression with least absolute shrinkage and selection operator predictor selection. It uses individual, disaggregated data and compares the prediction results to current Marine Corps attrition modeling processes. Two key findings are reported. First, the Random Forest classifier models outperform the current trailing average models at predicting aggregate attrition. One caveat is that these models have difficulty at correctly classifying non-end of active service attrition at the Marine level, achieving an average of 45% correct individual classification. Second, even though the machine-learning models provide superior prediction, they may not be managerially relevant because of the opportunity cost of construction due to the current database structure, data systems, and capabilities employed by Marine Corps manpower entities.