Page 13 - Case study: Crime prediction for more agile policing in cities – Rio de Janeiro, Brazil
P. 13
Globally, there is still comparatively mixed evidence of the accuracy of crime prediction, its impact
on clearance rates, whether it improves response times or even leads to significant reductions
in crime. The only way to really gauge the impacts of crime forecasting is to conduct statistical
evaluations that isolate the effects of the measure, including randomized control trials (RCTs). The
annex provides a summary of the findings from a number of the most well-known instances of
crime prediction implementation.
3. Conclusions
CrimeRadar is an innovative AI-based solution that utilizes historical city crime data to predict crime.
The solution is adaptive in that it is able to incorporate to new incoming data and make adjustment
accordingly. The prediction algorithm can be applied to other cities as long as the crime data is
available for them, hence CrimeRadar is by design transferable to other cities (subject to data
consistency and translation where needed).
Limitations of crime prediction: For all their promise, predictive crime analytics are not a panacea.
For one, certain types of crime - including domestic and interpersonal violence - is not easily
amenable to predictive models since they are seldom concentrated in specific locations, and cannot
be readily attributed to specific profiles of victims. While predictive algorithms may reduce certain
forms of human bias by reducing subjectivities, they ultimately rely on often flawed crime data with
systematic reporting biases.
Furthermore, predictive policing experiences, when subjected to closer scrutiny, have registered
a host of challenges . They are often costly owing to data storage, lacking transparency in relation
9
to the underlying algorithm, and having on occasion led to the violation of basic rights and civil
liberties . Without high quality data and due care in the way they are built, predictive algorithms
10
can unintentionally reproduce and exacerbate societal prejudices . Identifying biases in data sets
11
is complex, requiring deep knowledge in statistics, mathematics, and programming. As with most
policing technologies, successful application requires a comprehensive approach. It depends not just
on institutional leadership and the technical capacity of law enforcement agencies to incorporate
predictive tools into routine operations, but also the development of minimum standards for
responsible development, auditing and evaluation.
Designing and deploying predictive tools: Crime forecasting is by definition a mixed method,
involving a host of integrated tasks. These include time-series modeling, intensive data mining, hot-
spots analysis, and socio-temporal assessment applied to historical crime data. It is important to
stress that predictive policing goes beyond basic online mapping tools that track crime.
Statistical methodologies include the near-repeat theory and crime hot-spot analysis. These
approaches assume that once a particular violent or property crime occurs in a particular location,
it is likely to occur again in that same area. Meanwhile, the risk terrain model is more focused on
geographical analysis, seeking to identify risk factors and features of crime-affected locations, such
as insufficient public lighting and potential escape routes.
Crime prediction for more agile policing in cities – Rio de Janeiro, Brazil 7