Why is mean square error a bad measure of model performance? What would you suggest instead?