Everyone wants predictive algorithms to be accurate, but there are different ways to define accuracy. Is it better to have an algorithm that's rarely perfect, but also rarely off by a mile? Or to have ...