A model says "fraud, 90% probability", but only 70% of such cases are really fraud. What is this problem called?
Options
- Class imbalance
- Underfitting
- Data leakage
- Poor calibration: the probabilities do not match the observed frequencies (Correct option)
Answer
Poor calibration: the probabilities do not match the observed frequencies
Why
A calibrated model is right about 70% of the time when it says 70%. A good ranking (high AUC) can come with overconfident probabilities, which matters when you use them in decisions such as expected costs.
Check it with a reliability (calibration) curve. Fix it with Platt scaling or isotonic regression, fitted on data the model did not train on.
Remember
A good ranking does not mean honest probabilities.