What is Conformal Prediction in model evaluation?

Conformal prediction provides prediction sets with statistically guaranteed error bounds.

Conformal prediction is a framework that wraps around any model to produce calibrated prediction sets with formal coverage guarantees (e.g., 90% confidence means ~90% long-run correctness). It does not assume distribution form and works by using nonconformity scores on calibration data. Unlike probabilistic calibration, it provides set-valued outputs rather than single-point predictions, making it highly valuable for safety-critical systems.