RoboCop: Crime Classification and Prediction in San Francisco

In this paper, we employ machine learning and other statistical techniques to the problems of classifying and predicting crimes in San Francisco. Drawing upon existing research in the field to approach these two problems, we employ Random Forest and VAR(p) models, respectively. For the classification problem, our results across all 39 crime categories demonstrate the difficulty of the fully-specified crime classification problem, as we achieve a maximum 39-way classification accuracy of 31.84%. Although our results are perhaps inappropriate for daily or weekly use in any police organization, the time series model performs adequately at forecasting crime incident averages in the coming weeks and months. With more data and the use of a time series model already developed by these authors for discrete time series, our results might be improved upon further.