Forecasting Ammunition Demand on a Modern Battlefield

Your favorite major retail store (online or brick and mortar) currently has the ability to forecast demand for commodities such as toothpaste and toilet paper. They use these forecasts to make decisions about the best quantity of goods to deliver to distribution centers and retail outlets and when to deliver them. They minimize shipping and storage costs and avoid having too many or too few items on hand at the point of need. Military units on today's battlefields would benefit greatly from a similar scheme to forecast ammunition requirements and make sound delivery decisions. The forecasting is more difficult though, as ammunition demand is often the result of external, unpredictable enemy decisions. Production is slow and requires long lead times. Delivery is dangerous and unpredictable. Available data is often incomplete or seemingly irrelevant. Even though the system that controls the need for ammunition is hidden from our direct observation (friendly and enemy unit actions and counteractions), we can still predict future system behavior through the analysis of time series data.

This research explores several popular forecasting methods to determine their strengths, weaknesses, and overall applicability to predicting ammunition demand by US Army units in Afghanistan from 2010 to 2013. Autoregressive integrated moving average (ARIMA), Exponential Smoothing, Hidden Markov Models, and historic average estimation models are all used to predict future ammunition demand. The resultant forecasts may be used to feed a comprehensive logistics planning system to help military leaders make informed decisions about commodity delivery on the battlefield in order to decrease risk and increase reliability of logistics resupply. Results indicate that forecasting the outputs of a system as unpredictable as war is very challenging. The univariate exponential smoothing models forecast with the least percent error for near term forecast horizons, and their accuracy is shown to improve with bootstrap forecast aggregation. A novel alteration to residual bootstrap aggregation is presented that increases forecast accuracy by mitigating the large variance for such a stochastic time series as ammunition demand. This research is relevant not only for military sustainment planners, but for anyone who works with demand that varies over time across several echelons of product and consumer.