The term predictive policing refers broadly to the use of data analysis to inform the allocation of police resources. Not surprisingly, the vagueness of this definition allows many different data analysis programs to be called predictive policing. Here I define predictive policing as a formal process that (1) uses data to assign explicit probabilities to future crime events in space and time, (2) presents crime event probabilities in a useable framework to law enforcement decision makers, and (3) leads to resource deployment patterns conditioned on crime probabilities. For predictive policing to be effective it also must hold that (4) the accuracy of predictions be evaluated and (5) law enforcement be willing to act on probabilistic information. Here I outline the behavioral and mathematical architecture underlying our approach to predictive policing. I then review the results of the Los Angeles Predictive Policing Experiment, a real-time single-blind field deployment of predictive policing in multiple divisions of the LAPD. The experiment establishes a predictive accuracy six-times random and more than two-times that achievable by a dedicated crime analyst. The experiment also underscores the critical importance of officer ‘buy-in’ for successful real-world deployments.