average calibration error

todo

In calibration, the goal is to produce a model $f$ such that

$$v\approx \mathbb{E}[y|f(x)=v].$$

That is, when $f$ predicts the response as $v$, the true response is roughly $v$.