bank-full.csv test: random forest tops 8 models, but duration must go

The Bank Marketing dataset comes from a Portuguese bank's direct phone marketing campaign, and the target is binary: whether the customer eventually subscribed to a term deposit. The data has a strong class imbalance, with only about 11.70% of the 45,211 records labeled yes, which is why plain accuracy is misleading. The tutorial shows that simply predicting no for everyone would already give roughly 88.3% accuracy while finding none of the interested customers, so ROC-AUC and PR-AUC are used instead.

The exercise also draws a line between correlation and cause. Education level, job, housing loans, and call counts all show different sign-up rates, but the tutorial stresses that these patterns are starting points for hypotheses, not proof that any one factor drives the decision. To settle questions like how many times to call, it says randomized tests or past-campaign causal analysis would be needed.

Beyond the classifier comparison, the piece emphasizes workflow choices: preprocess after splitting the data to avoid leakage, use stratified sampling and StratifiedKFold for the imbalanced target, and reserve the test set for a single final evaluation. The most important design decision, it argues, is defining when the prediction is made and what information is genuinely available then.