1% of transactions are fraud. A model that always predicts "not fraud" has what accuracy, in percent?
Answer: 99 %
It is right on the 99% of legitimate transactions and catches no fraud at all. On imbalanced data, look at precision, recall and PR-AUC instead.