Exploring a Space-Time Autoregressive Moving Average (STARMA) model in spatial crime forecasting

Abstract

So far, a variety of methods is used by predictive policing for forecasting crime purposes. Depending on the type of forecasting, some methods work on forecasting the number of crime incidents while others on forecasting whether an area is possible to have crime incidents or not. At the same time, crime is not considered stable in space and most of the popular methods that are used do not take into consideration the parameter of time. This is an important issue not only for police departments, but also for a variety of fields (criminologists, geographers, society and academic society). Hence, the exploration of current Space-time autoregressive models, which are based only on past space and time crime data, becomes more and more necessary. With the present thesis the importance of a Space-time autoregressive moving average model for forecasting crimes is examined compared to simple methods that have been used till now, based only on past crime data. The study focuses on New York City and three crime types are examined. The selection of current and baseline methods (a conventional Kernel Density Estimation and a naïve approach), their parameters and the proposed method (spatiotemporal autoregressive moving average – STARMA) is done after an extended research literature. Likewise, a research about the evaluation of forecasting performance metrics is done. The selected methods are compared under the same methodological framework, where a threshold value classifies their outputs into two classes; hotspots and non-hotspots. The proposed method’s parameters are experimented to examine how the consideration of space and time affect its performance. In the end, the experiment that excels is compared with the baseline and the conventional method through spatial forecasting accuracy metrics. According to the results, all the three methods present important outputs and each method outperforms for the different examined crime types (all, property and violent). More particularly, the naïve Baseline method outperform for type crime ‘all’, the conventional KDE excels for crime type ‘property’ while the proposed STARMA method shows higher performance for the crime type ‘violent’. The study concludes therefore that the current Space-time autoregressive models are quite sensitive to the parameters of space and time and further research needs to be conducted to examine under which spatial and temporal resolution these models could exceed the baseline methods.