Pre-crime to the rescue : sometimes a correlation without causation is enough

Tech firm PredPol - short for predictive policing - claims its data analytics algorithms can improve crime detection by 10-50% in some cities. It takes years of historic data, including the type, location and time of crime, and combines this with lots of other socio-economic data, which is then analysed by an algorithm originally designed to forecast earthquake aftershocks. The software tries to predict where and when specific crimes will occur over the next 12 hours, and the algorithm is updated every day as new data comes in.

"PredPol was inspired by experiments run by the University of California in collaboration with the Los Angeles Police Department," says PredPol co-founder and anthropology professor Jeff Brantingham. "That study demonstrated that algorithmically driven forecasts could predict twice as much crime and, when used in the field, prevent twice as much crime as existing best practice."

Prof Brantingham says machine learning allows PredPol to analyse data, draw conclusions and make connections between large amounts of data that human analysts simply could not cope with.

More than 50 police departments across the US use PredPol software, as well as a handful of forces in the UK. Kent Constabulary, for example, says street violence fell by 6% following a four-month trial. "We found that the model was just incredibly accurate at predicting the times and locations where these crimes were likely to occur," says Steve Clark, deputy chief of Santa Cruz Police Department. At that point, we realised we've got something here."

Sceptics say this is pseudoscience, because crunching crime data to make informed decisions on police deployment is nothing new. Many forces have traditionally used "hot spot analysis", where past offences are recorded and overlaid onto a map, with officers concentrating on those areas. But PredPol and others working in this space, such as Palantir, CrimeScan and ShotSpotter Missions, say that traditional hot spot analysis is just reacting to what happened yesterday, not anticipating what will happen tomorrow. AI and machine learning can spot patterns we've never noticed before.

I'm not sure what they mean by "specific crimes". I presume they mean type of crimes, not specific incidents. If so, it doesn't seem outlandish to me that there could be non-obvious patterns in data that give reasonable, broad indications of what might happen next. Obviously it would be ludicrous to suggest it could be anything like Minority Report's psychics, but a more general crime forecast ? "Tomorrow : cloudy, scattered showers with outbreaks of racial violence" ? Doesn't seem overly-preposterous. The thing about correlation not equalling causation is that sometimes you don't need to understand the cause to get useful information.
https://www.bbc.com/news/business-46017239