OP said elsewhere that the specific discount that produced the $10M sales + refund combo was only active at five stores. That's the sort of edge case that you can't count on being in the training data. I'm as big a fan of real-data testing as anyone, but there's always something not represented.