4. Deciding who to act on#
A trained model gives you a score per instance. Turning that into an action — contact this customer, reject this application — needs one more decision: where to draw the line.
Scores, probabilities and calibration
What a score has to mean before a cost computed from it is trustworthy, and how to get there.
Threshold Tuning
Picking the cut-off itself, either as a probability threshold or as a fraction of the population to target.