The algorithm versus the chimps: On the minima of classifier performance metrics

n this paper we seek the minima of performance metrics for binary classification to facilitate comparison between metrics and applications, and to assess the quality of inferential statistics made from non-probability samples. We use these minima to min-max normalize the performance metrics so that they can be interpreted as a percentage of the perfect classifier relative to the proverbial chimps at the zoo†guessing at random. We compare our results with the balanced metrics that have been introduced recently, which are corrected for bias due to class imbalance.