Sorry for hijacking this thread, but just wanted to share this whitepaper I stumbled on. It’s on a piece of software called Hunchlab that is used by the US police to predict crime and plan deployment.
Some technical details are noteworthy, like their use of time-related features that also make sense for us given the observed temporal dynamics in our data, or the use of a Poisson model to calibrate the predictions. I also like how they offer to measure the performance of the software on the data of a particular police department (‘more [data] is better’, ‘cleaner [data] is better’).
But what I really like about this document is how they present prediction to (power) users: the name Hunchlab to me is a stroke of genius, being quite modest (e.g. in constrast to Precrime) and so not too threatening, but also how they relate the workings of the algorithm to how a police officer would go about trying to predict where crime is most likely to occur next.I think we can learn something here on how we communicate prediction to doctors.