Miranda Mowbray

Predictive analytics: you can't have it all

I will talk about limits to predictive analytics. The main application I'll talk about is recidivism prediction – predicting whether a convicted criminal is likely to reoffend. However what I will say applies to predictive analytics more generally. As Alexandra Choldecova pointed out in the wake of a controversy about the COMPAS recidivism prediction algorithm, except in trivial cases, it is not mathematically possible to maximize the accuracy of recidivism prediction while meeting some fairness requirements for groups with different underlying recidivism rates. It is a policy choice whether or not to accept reduced accuracy, at least in the short term, in return for meeting fairness conditions. I will discuss this and some other limits to prediction that require policy choices. To make these choices, we need informed discussion and collaboration between techies, lawyers and policy makers.
Just for fun, my slides will include 79 cats, a catbot, and Catwoman.