Coding and AI classes in Farnborough

From the Nomis API the learner downloads four Census 2021 tables for all 318 output areas in Rushmoor. Of 39,339 households, 10,831 live in a flat of some kind, most of them, 9,138, in purpose-built blocks. The share in flats is the target. The inputs are population density, the share of one-person households, the share of households without a car, and a fourth column filled with random numbers that cannot carry any information. A random forest is trained on 70% of the areas and tested on the other 30%, and the whole process is repeated on 20 different random splits.

On unseen areas the model explains about 65% of the variation in the flat share. Two importance methods are then compared. The first is the forest's built-in score, based on how much each input helped split the training data. The second, permutation importance, shuffles one input at a time on the test areas and measures how much worse the predictions become.

The built-in score hands the random column 6.0%, half as much as density. That happens because the score is measured on the training data, where a forest can always find some split on noise that fits a few areas slightly better. Permutation importance, measured on areas the model never saw, gives the random column essentially zero: shuffling it changes nothing. Two more warnings emerge. Density's permutation score swings from -0.023 to 0.121 across the 20 splits, so with only a few hundred areas the smaller scores are shaky. And the no-car and one-person shares are correlated (0.66), so they partly stand in for each other; their scores describe this model, not which one causes flats.

Farnborough teenagers who can audit an explanation will not be talked round by a confident chart, and coding is where that audit skill is learned. The longer argument is in why teenagers should still learn to code in 2026.