20. Avon and Somerset Police uses modelling to collect data about offenders, through the TRAP system.^[1] This identifies suspects from a cohort of subjects, who are given a score indicative of their risk of carrying out child sexual exploitation offences. A score is based on a number of factors which are common to identified suspects, such as having “sexual offences, violent offences, antisocial behaviour, drugs, criminality with children, links to MISPERs [missing persons], links to gangs and organised crime groups and trafficking”.^[2] Officers within Operation Topaz then consider whether to flag an individual based on that score and other intelligence, before deciding whether and how to intervene based on the views of the analyst, the researcher, the police sergeant and partners, as well as the score.^[3] From TRAP, a weekly ‘offender’ list is circulated between multi-agency partners, which includes the reason why each subject has been identified.^[4]

21. Bristol City Council uses its Think Family Database to create a profile of children at risk of sexual exploitation. This brings together 35 different datasets about children from education, housing services, the police, welfare systems, the number of episodes of children going missing and mental health concerns.^[5] The ‘Insight’ team, funded jointly with the police, used that data to create a predictive risk model to analyse the extent to which children are at risk of sexual exploitation.^[6] The model produces an initial score for each child based on known risk factors (such as episodes of going missing). Their risk level is then analysed, based on wider information.^[7]

22. The Think Family Database also produces a weekly, automated list of children at heightened risk of sexual exploitation, which is provided to Avon and Somerset Police’s Operation Topaz. This list is then analysed alongside flags on the Operation Topaz system to create a ‘victim list’ which is shared with partner agencies at multi-agency meetings every four weeks.^[8] In June 2020, for example, more than 1,000 children in the force area were identified as at heightened risk of sexual exploitation.^[9] Avon and Somerset Police told us that this allows them to understand who may be vulnerable to sexual exploitation so that it can support them proactively, rather than reactively when they come to the attention of professionals.^[10]

23. The use of predictive analytics to identify children at risk in Bristol and elsewhere has been subject to press comment and academic review.^[11] Concerns have been raised about the quality of the data inputted into the model, the risk of reinforcing the errors and biases of those making the original records and the potential to focus on factors linked to socio-economic and racial discrimination.^[12] Bristol City Council responded to many of the issues raised and was clear that the model identified vulnerability but did not replace professional judgement.^[13] It described the model as ‘mirroring’:

“it’s not predicting you will be sexually exploited … it’s saying you are demonstrating exactly the same characteristics and behaviours as someone who was sexually exploited”.^[14]

24. Bristol’s statistical analysis of the model in September 2020 found it to have ‘Very Strong’ precision (“how accurately the model found those who go on to be sexually exploited”) and ‘Very Strong’ recall (“the % of young people that were sexually exploited that the model found”).^[15]

25. However, research published in September 2020 by What Works for Children’s Social Care was more cautious about predictive models in other areas. It found that, of predictive models used in four different areas, none had sufficient precision to be considered a success. They missed four out of every five children at risk (false negatives) and of the children the models did identify as being at risk, they were wrong six out of ten times (false positives). The improved collection and use of data is critical to the response to child sexual exploitation but it is important that agencies do not over-rely upon it. Caution must be exercised; predictive analytics on its own may produce too many misleading assessments. While it may be a helpful supplement, predictive analytics should not be used as a principal tool.