Discrimination
Discrimination is the ability of forecasts to separate outcomes that happen from outcomes that do not.
Definition
Discrimination describes whether higher predicted probabilities are assigned to outcomes that happen more often than to outcomes that do not. It is about ranking and separation, not absolute calibration.
Why it matters
A forecaster can have reasonable calibration but poor discrimination if probabilities do not meaningfully separate cases. Strong discrimination supports better decision making even before cost modeling.
Common pitfalls
Ignoring calibration: Discrimination without calibration can still be misleading.
Small samples: Discrimination metrics can be noisy on limited data.