Data-driven prevention

Between the numerous police databases, the service has access to a wealth of information on criminals in each force area. However, according to Detective Chief Inspector Richard Evans of West Midlands Police (WMP), this intelligence is not always used as thoroughly as it could be, which can lead to inefficiencies in how offenders are managed. After deciding it wanted to make better use of its intelligence, WMP hired a team of data scientists and created an analytics laboratory for them to work in. They were tasked with building an algorithm to assess the likelihood of any particular person becoming an influential driver of criminality or a significant co-offender. This target was chosen because the force strongly believes co-offending is a significant pathway into crime. With the experts acquired, the force had to decide which data sources were most appropriate for this project. In the end, it settled on a wide range of systems, integrating information from the crime recording system, stop and search data, the Police National Computer, its intelligence management system, the Corvus integrated offender management (IOM) system, and the organised crime gang tracker. In total, the ensuing terabyte of data contained 10.4 million names – some identical – which were matched down to 7.1 million connected to 1.4 million entries, and then again to 5.2 million identities and 1.98 million entries. WMP then developed interim nominal tables – essentially timelines for every person mentioned in the data – which explained their journey to becoming criminal influencers. For an example of the algorithm’s power, in one case it successfully matched 108 nominal entries containing 27 different combinations of first names, surnames, etc. In essence, something the force’s systems had conventionally represented as 27 different records were turned up on a single search result. “At the moment we have an algorithm that we are confident in, which has the potential to predict people who are influential in driving co-offending and to identify people as they drive towards being involved,” Det Chief Insp Evans told the Society of Evidence Based Policing conference in March. “These are people who have a high impact downstream, and this gives us the opportunity to intervene with people before they make that transition.” Some of the immediate results were unsurprising. Analysis suggested that the younger a person is when they co-offend for the first time, the more likely they are to commit other crimes in the future. Thirty five per cent of the people identified who co-offended before the age of 23 became reoffenders, which Det Chief Insp Evans described as “not much of a shock”. However, other results were more unexpected. According to the scientists, 76 per cent of children aged under 11 who commit a violent offence will have co-offended in the past. In practical terms, this means the algorithm can look at a child who has committed or been subject to a crime, and if they have co-offended, more than three quarters will commit another violent crime. Similarly, 52 per cent of people who have committed any co-offending crime before age 11 go on to commit a section 20 wounding. This drops to 29 per cent if they are 23 or over. In context, a child who merely shoplifts with a friend has more than a 50 per cent chance of becoming a violent offender. The next step was to identify the core influencers driving co-offending. This led to a small group of nominals who shared a number of key factors: they were three years older than the people they associate with; they co-offended in a wider network; and on average, they had committed 42 crimes, of which ten involved co-offending. This subsection – just five per cent of the total number of offenders identified – accounted for 40 per cent of all co-offending crimes over five years, as well as 23 per cent of solo crimes. Based on the target group, WMP identified 1,349 key performance indicators (KPIs) that could lead to people co-offending. Traditionally, KPIs focus on ar