How does predictive policing work PredPol

Predictive policing

Predict crimes before they happen? The Californian start-up PredPol is already doing this.

A city in the middle of the US state of New Jersey. It is early evening and a police car is on patrol in a neighborhood with a high crime rate. Two officers get out and start a conversation with a passer-by; they want to know what the current situation is like. As they go back to the patrol car, suddenly a car rushes by; a person shoots at the passerby. The officers react immediately, take the victim to the hospital - at the same time the suspect is arrested. The calculation here is simple: if the police had not been on the patrol with foresight, firstly they would not have saved the victim and secondly they would not have been able to arrest the perpetrator. Jointly responsible for it? An algorithm based on machine learning and artificial intelligence (AI). “Since then, the officials were finally convinced of PredPol's predictive policing software,” recalls Brian MacDonald, CEO of the California start-up PredPol, which was founded in 2012.

Predictive policing describes the phenomenon of using software systems to predict future crimes - namely the type of crime, the crime scene and the period of the crime. Crimes should thus be prevented before they are committed. Police authorities across the USA and Central Europe are now relying on this type of crime prevention. Crime rates have actually decreased, according to some reports. The Bavarian Interior Minister Joachim Herrmann reported a 42 percent reduction from October 2014 to March 2015 in Munich. However, it seems more than unclear to what extent the machine contributed to this or whether this can be attributed to other reasons. Predictive policing is also viewed critically by some experts (such as John Hollywood from the American non-profit organization RAND Corporation) and data protectionists: Is it allowed to put oneself in the hands of an algorithm with such a sensitive topic? What if wrong predictions are made? Are personal data used for this?

In any case, the fact is that predictive policing solutions are being used more and more around the world. In terms of reach, the start-up PredPol is the largest provider. The company emerged from a research project between the University of California, Los Angeles (UCLA) and the Los Angeles Police Department (LAPD) that started in 2007. The team around the anthropologist Jeffrey Brantingham (UCLA), who specializes in crime patterns, and the computer scientist George Mohler rummaged through heaps of criminological data sets in order to identify behavioral patterns and put them on a mathematical foundation. After a few departments in Los Angeles, the one in Santa Cruz (around 558 kilometers from Los Angeles) signaled a specific interest in testing it. An organizational framework was needed - PredPol was born in 2012. Brantingham was there, as was Mohler, and the initial funding came from various large business angels. Since then, the company has financed itself from its own revenues - CEO MacDonald does not want to reveal how high these are. “We don't get any venture capital and grow organically. Above all, PredPol has a clear mission: to make societies safer by reducing crime. ”MacDonald's sentence does not sound trite, rather he jokes in this context about the often overused saying in Silicon Valley:“ We're making the world a better place. "

The algorithm is based on two scientifically sound criminological theories - they start with the behavior of the perpetrator. “Repeat Victimization” means that it is very likely that a criminal who broke into a household in one night will try it again in a timely manner. Because this “success” has proven to the perpetrator that it works and motivates him to make another attempt. "Near repeat victimization" describes that a follow-up offense in the immediate vicinity is most likely 48 hours after the first crime. The environment is also included, such as parking garages or shopping centers within which the offenses take place. “For each police department, we collect data that is three to five years old. The software automatically works these through for significant crime types and periods, ”explains MacDonald. The historical data are linked with those from real time and thus result in the (data) picture for the future. The officers are presented with risk areas in the form of red boxes on a screen; within a radius of 150 by 150 meters and a time window of up to twelve hours. Depending on which PredPol customer the software is used for, different crimes are predicted: burglary and vehicle theft, robbery, robberies, gun violence.

Particularly important - and in contrast to related instruments such as the Chicago Heat List, which listed 400 people who are most likely to be involved in a crime in the future: Victims and perpetrators are not spat out. “The police officers are given instructions for their shift on where it makes most sense to patrol. The decision whether or not to do it is up to them, ”says MacDonald. Without human decision-making authority, the technology cannot be used, said the former Vice President of Sales at Vidcie, a Silicon Valley company within which he implemented cloud-based security solutions in federal agencies and local law enforcement agencies.

At first glance, PredPol seems to work, or better: to be accepted by the police departments. Because meanwhile more than 50 cities in twelve US states use it; including Atlanta and Norcross as well as Seattle. The only customer outside the USA has been Kent, Great Britain, with 2,000 police officers since 2013: "After initial skepticism, the system has become part of their job," says MacDonald. Nevertheless, the question arises: How can the success of PredPol be measured? “The most objective parameter is certainly the reduction in the crime rate - but that also depends on the offense and the region. In a department in Los Angeles, this dropped 53 percent in five months. Of course, this can also have other factors: The criminal, who usually steals cars on Tuesday evenings, moves away or is caught. Or the police are increasingly relying on other measures to convict criminals. We still see a strong correlation with our software, ”says MacDonald. Apart from in-house evaluations (according to a 21-month study from 2015), there is no evidence of the extent to which PredPol actually works.

Elsewhere, precisely this question is faced with a paradoxical starting point. On the one hand, the clear goal of such providers is to reduce criminal offenses. On the other hand, criminal law authorities will only increasingly rely on predictive policing if the prognoses are correct - and these will be appropriate if there are secondary offenses (maximum within seven days, note) in the areas of operation. In this respect, the crime rate does not decrease. The verifiability also becomes blurred when it is not possible to measure how many crimes have actually been prevented by the software. “That is quite difficult to convey. We started in 2011 with the goal of reducing the crime rate by one percent within the respective configuration period of six months, ”says Ralf Middendorf, IT specialist at the Institute for Pattern-Based Forecasting Technology (IfmPt) in Oberhausen. “PreCobs” is the name of the software developed for crime prognosis, which is used in Basel, Zurich, Stuttgart and Munich; further cities are being planned. PreCobs is particularly specialized in burglary. “In 2013, the Zurich City Police decided to carry out a pilot project. An official dealt with this topic there, ”says the sociologist and criminologist Thomas Schweer, managing director of the IfmPt. Continuous operation began in 2014; the given goal has already been achieved. According to a report by the Zurich City Police (2017), the number of burglaries fell from 3,511 in 2013 to 2,470 in 2016. Two of the findings on PreCobs were: "Apparently positive effect (deterrent) in previously heavily polluted areas." The proportion of NR (near repeats, note) has fallen in Zurich, but is still significantly high. "

At the turn of the millennium, managing director Schweer was often on patrol with the police, he analyzed their working methods: “The officers always found out what happened the day before. They then marked the most risky areas with pins on wall maps. ”Something more modern was needed, not a“ crystal ball ”, but a computer that never tires of forecasting.

But even this is not immune to mistakes, continues Schweer. After all, it can happen that an officer forgets to enter the “modus operandi” (behavior) in the computer system after a case has been closed. Because in addition to the crime scene, the time of the crime and the offense, PreCobs also feeds in the booty and the behavior of the crime. From this, the machine in turn gains its - in this case imprecise - knowledge for the future. “The officer on duty has to decide whether to ask his colleague again.” The colleagues from PredPol see it similarly: “We do not promise 100 percent forecast accuracy. The prognosis is based on human behavior. If a robbery happens every Friday between four and five in the afternoon, but not one, it is already affecting this technology, ”says MacDonald. As at the institute in Oberhausen, PredPol does not use any personal data that could be used to infer the identities. For example, two to three blocks of houses are displayed in the red areas, but no exact addresses.

MacDonald cannot predict the future of his own company, but he does indicate a direction: “We are banking on international expansion. We are already in contact with police departments in Asia and South America. But large industrial companies also rely on our algorithm to make their own area safer. In the future it could be insurance and logistics companies. "

PredPol
The Californian start-up was founded in 2012 by co-founders Jeffrey Brantingham (UCLA) and George Mohler (formerly Santa Clara University). Brian MacDonald, formerly Vice President of Sales at Vidcie, has been CEO for two and a half years. The company offers software technology to predict and prevent future crimes. The basis here is formed by human behavioral patterns that are placed on a mathematical foundation. The police officers are shown risk areas in the form of red boxes on an interface (see example on the left). These mark a radius of 150 by 150 meters and are valid for eight to 12 hours, depending on the police shift. The individual points indicate offenses such as robbery or robbery that have already happened. PredPol's core market is the USA. In the future, the start-up also wants to be active in South America and Asia.

Illustration: Valentin Berger

This article appeared in our January 2017 issue of “Forecasting”.

Niklas Hintermayer,
editor