It is a framework to give certainty Intervalls. Basically the prediction will be an Intervall with a guarantee, that the result is to X percent inside of that Intervall.

The Intervall itself can be a probability interval or even an interval of Intervalls. It is called a Set Prediction. For classification the output would be {'cat', 'dog'}, instead of just {'cat'}.

We want our conformal prediction method to have the following properties:

• Efficiency: have the smallest Interval possible
• Validity: Be correct
• Model agnostic

The loss function can then be changed to train for better Efficiency instead of just Validity (by adding a regulating factor).