Test points, fit and conformal band
Target, guaranteed and empirical coverage

Split Conformal Prediction Band around a Polynomial Fit (2D)

Fit a polynomial, hold out calibration points, take the right order statistic of the absolute residuals and wrap the fit in a prediction band that covers new points at least 1 minus alpha of the time, whatever the model. Change the calibration size, alpha, noise and model degree and compare the guaranteed coverage l/(n+1) with the measured one. A Blender-built plot table with numerical checks of the Beta distribution of coverage built in.

Made with 🟠 Blender