In short, it's when your training data come from different conditions than your real world data. This means that the probability distribution your model was tra

In short, it's when your training data come from different conditions than your real world data. This means that the probability distribution your model was trained on is different from the one you are using it on. This means that an model that did great in cross-validation could still be useless in practice.