Why is logistic regression not guaranteed to stay calibrated simply because it optimizes log loss?
Classic MLLevel: IntermediateModel-specific calibration
Why is logistic regression not guaranteed to stay calibrated simply because it optimizes log loss?
Hints
- The fitted model class may be wrong.
- Training and deployment populations may differ.
Answer
Correct model specification, regularization and the evaluation distribution matter. The favorable calibration behavior in a well-specified setting is not a promise under misspecification or distribution shift.
Source and licence: scikit-learn 1.7.2 — calibration · BSD-3-Clause · answer: Quantoros editorial