Space–time autoregressive models and forecasting national, regional and state crime rates

Abstract

Abstract The recently advanced space–time autoregressive (ST-AR) model is used to forecast US, regional and state rates of violent and property crime. The disaggregate state (Florida) violent crime model includes murder, rape, robbery, and assault, while the property crime model includes burglary, larceny, and motor vehicle theft. In experimental forecasts, ST-AR RMSEs are compared to those for aggregate univariate AR(p) models, vector autoregressions (VAR), Bayesian VARs (BVAR), and two naive models that predict future crime rates either as the most recent rate or according to the most recent change in rates. The ST-AR model is of particular interest, given its efficient use of data, much like panel-data estimation. The ST-AR, BVAR, and AR(p) models outperform the other three approaches, but the ST-AR models are generally superior.