Using discrete event simulation to evaluate time series forecasting methods for security applications

This paper documents the use of a discrete event simulation model to compare the effectiveness of forecasting systems available to support routine forecasts of criminal events in security applications. Military and police units regularly use forecasts of criminal events to divide limited resources, assign and redeploy special details, and conduct unit performance assessment. We use the simulation model to test the performance of available forecasting methods under a variety of conditions, including the presence of trends, seasonality, and shocks. We find that, in most situations, a simple forecasting method that fuses the outputs of crime hot-spot maps with the outputs of univariate time series methods both significantly reduces modeling workload and provides significant performance improvement over the three currently used methods: naive forecasts, Holt-Winters smoothing, and ARIMA models.