Datasets and notebooks

Follows the mathematics workshop in order. The eight stores from the slides are the running example, so every formula is checked against an answer the deck already prints. London data enters twice, where eight rows are not enough: to show a sampling distribution, and to read $R^2$ out of sample.

Fits a rule by hand on four properties, so nothing about fitting stays hidden inside a library call. Then asks the harder question of the bank campaign: what evidence would let you claim the rule works on clients you have never seen?

Opens both models. Least squares on London prices, then the same grammar for a probability: the sigmoid, log loss, calibration, and a decision cut-off taken from the prices of the two mistakes rather than from a software default.

Fits models that can bend, and builds the testing discipline that stops them flattering themselves. It ends with two evaluations that reverse each other's verdict on the same data.

The answer column is taken away. Two questions replace it: which listings behave alike, and what few forces drive nine Treasury yields. Because nothing checks the result automatically, the day is mostly about designing the check.