You can now rigorously measure and estimate how much your model's performance will degrade when facing distribution shifts, using a unified framework that works across different types of shifts and loss functions.
This paper addresses how machine learning models fail when training and test data distributions differ.