Coding and AI classes in Carrickfergus

The learner downloads two NISRA Census 2021 tables at data zone level, the smallest areas NISRA publishes these for: age by single year and household size. For each of the 3,780 zones the target is the share of households with one person living alone; the inputs are the shares of residents in six age bands plus the number of residents. The 74 zones of the Carrick Castle and Knockagh electoral areas are held back as the test, where one-person households range from 6.2% to 68.3%. Every model trains on the other 3,706 zones.

A full-depth tree fits its training data closely, and that is its weakness: train it on a bootstrap sample, the same zones drawn at random with repeats, and the prediction for a single Carrickfergus zone typically moves by about 8 percentage points from one sample to the next. Averaging 50 such trees cancels much of that wobble and cuts the typical error from 10.07 to 8.55 points; scikit-learn's bagging with 10, 50 or 200 trees lands between 8.17 and 8.28, and separate 100-tree ensembles differ by only about 0.7 points on a zone. Bagging does little for the shallow tree. It was already fairly steady, and its problem is that three questions are too few to capture the pattern, a bias that averaging cannot fix: 50 averaged shallow trees still missed by 10.86 points.

A Carrickfergus teenager who knows when averaging helps, and when it cannot, will read AI claims with the right amount of doubt, and coding is where that instinct forms. The longer argument is in why teenagers should still learn to code in 2026.