Hi everyone! One simple (or not) question(s)? How can I fix a bad calibration model on a test dataset without creating a new one? It depends on algorithm or not (LR, SVM, RF and so on)? What are the criteria for determining whether a calibration is bad (value of slope? value of average calibration error? Is it bad if error 5%?). I can’t find the answer to this question. I use R for model creation (package caret) and model calibration (package rms).
Please repost this as a reply at the end of RMS Describing, Resampling, Validating, and Simplifying the Model - #10 by Elias_Eythorsson and I’ll remove it here.