Rethinking the science of prediction