Why Accuracy Misled a Lead-Scoring Model: Precision, Recall, ROC AUC and Cross-Validation

A developer who built a lead-scoring classifier in Module 3 of ML Zoomcamp initially reported 0.645 accuracy, which seemed acceptable. Module 4 of the course introduces ROC AUC, precision, recall and cross-validation to show why accuracy alone can give a misleading picture. The revised model reaches an AUC of 0.732, and the author argues that the threshold tradeoff becomes clear when plotted. The post also stresses that cross-validation standard deviation matters as much as the mean score. Finally, it shows how sweeping the regularization parameter C makes the bias-variance tradeoff visible, since a very small C produces a model that barely learns.