There’s an old xkcd where a guy standing on top of a giant trash heap of math symbols explains how machine learning works: you dump your data into this junk pile here, see, and answers fall out the other end. And if the answers are wrong? “Just stir the pile until they start looking right,” he shrugs.

Models like goals added (g+) are great at answering wildly complex questions like “How much did this left back’s whiffed tackle at the halfway line change his team’s mathematical probability of scoring next time it gets the ball?” but terrible at telling you how they did it. In that sense the model is sort of like the athletes it’s trained on, guys who get a face full of microphones after every game but, as David Foster Wallace once wrote, “usually turn out to be stunningly inarticulate about just these qualities and experiences that constitute their fascination.” What were you thinking when you derived that bizarre possession value? Well, Sebi, it’s not the result we wanted but we’re just trying to take it one calculation at a time. Thanks to the fans for believing in us.