A New Military Retention Prediction Model: Machine Learning for High-Fidelity Forecasting

June, 2019
IDA document: D-10712
FFRDC: Systems and Analyses Center
Type: Documents
Division: Strategy, Forces and Resources Division
Authors:
Authors
Julie Pechacek, Alan Gelder, Cullen Roberts, Joe King, James Bishop, Michael Guggisberg, Yev Kirpichevsky See more authors
IDA’s Retention Prediction Model (RPM) uses machine learning algorithms and 19 years of extensive personnel records to capture rich interactions in service characteristics and predict when individual servicemembers will separate from the military. Based on information about a service member’s career and characteristics observable up to at a given point, the RPM estimates the probability that a person will continue to serve for any number of future periods. The RPM uses a survival loss function developed specifically for analytic applications where the end state in a chain of events is not observable or has not yet occurred. The RPM’s person-level predictions can be aggregated by any desired population subset, including career field, cohort, unit, or demographics. The RPM produces person-level predictions that closely mirror actual attrition patterns. Testing on out-of-sample data, given two randomly selected servicemembers, one of whom separates from the military within one year, the RPM identifies the correct individual 88% of the time. Extending the time horizon to four years, the model is correct 80% of the time; for any number of years up to 18, the model is correct more than 78% of the time.