But since my experience (always
risky when thinking about statistics which so often has provable results that seem
counter to intuition) tells me that one should be really excited to have a model that shows more skill in verification than over the data
against which it was calibrated, I have to wonder why discard those cases at all?