Ninety-Nine Point Eight Percent: Classification, Thresholds, Calibration and Rare Events

Deep End Era. Engineering with Sankalp · Oct 28, 2026 · 50m

A vendor shows a card fraud model that is 99.8 percent accurate, and Amara, Northgate Pay's new head of fraud analytics, points out that a model that never flags anything scores the same. Part two of The Deep End's Machine Learning series builds the classic classifiers: logistic regression as stacking clues on the log-odds scale (with Verhulst, Berkson and Cox), nearest neighbours as asking the neighbours, naive Bayes and Paul Graham's "A Plan for Spam", and support vector machines as the widest street between two neighbourhoods. Then we count. We go through the four boxes of the confusion…