Forecasting Attrition from the Canadian Armed Forces using Multivariate LSTM

Using 15 years of monthly release volume of Canadian Armed Forces Regular Force members, we showed that a Long Short-Term Memory (LSTM) multivariate machine learning model performs better than univariate statistical models (Holt-Winters) for the same number of observations. For a 3-year forecast, our data supports a prediction error of less than 0.5%.