Forecasting Army enlisted ETS losses

Abstract

The Army currently uses time series models to forecast active-duty enlisted personnel losses. These time series models can provide accurate predictions but offer no insights into the underlying causes of loss behavior. In order to quantify the various forces that influence retention rates, a regression model is necessary. In this thesis, logistic regression is used to estimate end of term-of-service (ETS) losses. The model estimates the probability of reenlistment for soldiers with 12 months remaining on their enlistment contract. The model relies largely on individual soldier information such as pay grade, military occupation, and education, but also examines the impact of the civilian unemployment rate. Two models are developed. The first model includes 14 main effects. The second model includes the same 14 main effects plus 21 highly significant two-way interaction terms. Both models estimate the total number of personnel that reenlist in a seven-month test period fairly well, although the main-effects model results are more accurate. The two-way interaction model performs slightly better on most statistical measures of model effectiveness. Because the two-way interaction model is more complicated to produce, and does not generate results that are clearly better than the main effects model, this thesis recommends using the main effects model to complement the current set of time series models.