"Only if the data is split in such a way that the test data differ significantly from the training data can it be determined whether the model can later handle novel data, so-called ...