Dominik Gerstner from the Max Planck Institute for Foreign and International Criminal Law evaluated the predictive policing project in Baden-Württemberg. The software predicts where break-ins may occur soon, on the basis of reported residential burglaries and defined parameters.
Interview: Yvonne Nestler Photo: Max-Planck-Institut für ausländisches und internationales Strafrecht
Mr. Gerstner, what effect does it have from your point of view that the analyses of the Precobs system only include reported burglaries?
Gerstner: Through dark field research, we know that most residential burglaries are reported. That’s why this isn’t a problem. Problems arise, however, in that the burglaries usually take place when no one is home. When the residents are gone for a longer period of time, the incident is oftentimes discovered and reported very late. This makes it difficult to predict regionally limited and small crime sprees.
What do the police think of Precobs?
Gerstner: The police who operate the software see it as a useful addition to existing resources. However, opinions differ among the cops on the ground who respond to Precobs alerts. Because the system challenges established routines and makes decisions on behalf of officials. In addition, the preventive benefit is not directly noticeable among patrol officers, of course.
You rated the project’s success only as moderate. Do you think that predictive policing can yield better results if the algorithm is self-optimizing with machine learning?
Gerstner: It is hard to say – mainly because there are virtually no research results so far. Also, machine learning, like predictive policing, is a broad field. In the case of residential burglaries, we know which parameters can explain spatiotemporal patterns. However, there are a great number of cases that cannot be explained. Chance plays a major role here, but machine learning may also help in these cases. However, it should be remembered that predictive policing is a process: What happens with the forecasts must be measured and incorporated again and again into the models. This is laborious – and miracles cannot be expected either. But it would be exciting to explore this.