AI and programming classes in Wimbledon

Four Census 2021 tables for Merton come from the Nomis API. Only the 199 output areas inside the five wards listed above are kept, home to 23,671 households; 7,585 of those have no car or van. Three facts about each Wimbledon area go in (how crowded it is, how many homes hold one person, how many homes are flats) and a gradient boosting model has to say what share of its households are car-free. Forty settings are drawn at random, varying the number of trees, their depth, the learning rate and more, and each is scored by five-fold cross-validation. The whole experiment is repeated ten times with different draws.

The winner's own score, 6.14, is the number most tutorials would report. Nested cross-validation, which repeats the search inside each training fold and scores on the fold left out, gives 6.37, and the winner's score was the lower of the two in 8 of the 10 repeats. The gap is small here because forty settings of one model on 199 areas leave limited room for luck; search thousands of settings on a smaller sample and it grows. The last row carries a different warning. On the 455 Merton areas outside Wimbledon the chosen model's error rises to 7.04, because those areas are not like the ones it learned from. Nested cross-validation makes the score honest for data like the training data; it cannot promise anything about somewhere else.

Wimbledon teenagers who can run that check are much harder to impress with a headline accuracy figure, and the skill comes from building the test in code. The longer argument is in why teenagers should still learn to code in 2026.