Platt scaling for one decision model's probabilities: sigmoid(A · logit(p) + B).
A decision model's raw probabilities are not calibrated. Among answers of 0.9, the share that is
right is not 0.9, and the gap differs by model. So a consumer that acts on a threshold fits these
two numbers per model and per question on labelled data, and applies them before comparing.
A > 1 sharpens and A < 1 softens; B shifts the whole curve.
Platt scaling for one decision model's probabilities:
sigmoid(A · logit(p) + B).A decision model's raw probabilities are not calibrated. Among answers of 0.9, the share that is right is not 0.9, and the gap differs by model. So a consumer that acts on a threshold fits these two numbers per model and per question on labelled data, and applies them before comparing.
A > 1sharpens andA < 1softens;Bshifts the whole curve.