Coding and AI classes in Finchley

From the Nomis API the learner collects four Census 2021 tables for all 1,113 output areas in Barnet. An area is labelled car-light if its share of households without a car is among the highest in the borough. The detector is a logistic regression fed three clues per area: how densely people live, how many households are a single person, and how much of the housing is flats. Its verdict on each area is always given without having trained on that area. The experiment is run four times, with car-light meaning the top half of areas, the top quarter, the top tenth and finally the top twentieth: just 56 areas where at least 56.4% of households have no car.

The ROC AUC hardly moves: about 0.91 in every row. Judged by that number alone, the model is equally good at all four tasks. Average precision, the summary of the precision-recall curve, tells a different story, sliding from 0.901 to 0.359. In practical terms, at the rarest setting the model must flag 209 areas to catch 45 of the 56 car-light ones, so 164 of its alarms are false and only about one in five is right. Even its 56 most confident picks are correct just 37.5% of the time. Nothing about the model changed between rows. What changed is how many negatives there are to be wrong about, which ROC ignores and precision counts.

A Finchley teenager who can turn a score into a count of false alarms will not be dazzled by a 0.9, and that habit is learned by coding the test. The longer argument is in why teenagers should still learn to code in 2026.