Comparison of near-Repeat, Machine Learning and Risk Terrain Modeling for Making Spatiotemporal Predictions of Crime

Abstract

The main objective of this study is to test and compare the prediction performance of three of the most common predictive policing methods. A near-repeat model, a supervised machine learning model, and a risk terrain model are tested and compared against each other using retrospective analysis of home burglary crime data from a Belgian city. Hotspot analysis is included as a baseline. Predictions are made for three different months (January, May and September 2017) to account for seasonal differences. Variations in spatial context (city center vs. suburbs) and the number of predicted risk locations are also tested. Prediction performance is measured using accuracy, near-hit rate, precision and F1-score. The results show that there are some notable differences in prediction performance between the model types across the tested variations. In general, the ensemble model tends to be the most consistent high performer across all tested variations. Also notable is that hotspot analysis is not clearly outperformed by the other methods. The different methods have their own strengths and weaknesses and optimal prediction performance crucially depends on the specific location context. More comparative analyses of predictive policing methods in different contexts are needed to gain a more complete picture. Future research could also focus on how combining methods can help improve crime prediction performance.