A model outputs the probability of fraud. You raise the decision threshold from 0.5 to 0.8. What usually happens?
Options
- Precision goes up and recall goes down (Correct option)
- Precision goes down and recall goes up
- Both precision and recall go up
- Nothing changes, because ROC-AUC does not depend on the threshold
Answer
Precision goes up and recall goes down
Why
A higher threshold means the model flags fewer cases, and only those it is most sure about. Usually that gives fewer false positives (higher precision) and more missed positives (lower recall). F1 moves depending on which effect is bigger.
ROC-AUC does not depend on the threshold, but precision, recall, F1 and the confusion matrix do.
Choose the threshold from the cost of each error, not by default at 0.5.
Remember
Higher threshold: precision up, recall down.