Reference · Glossary

Threshold

A **cutoff score** that turns a continuous model output into a yes/no (or class) decision.

When to use

Tuning precision vs recall — raise threshold to approve fewer high-risk cases.

When not to

When the model already outputs calibrated class probabilities and you use argmax instead.

Example

`pred = 1 if score >= 0.7 else 0` — stricter than 0.5, fewer false approvals.