Online coding and Python classes in Hayes

The learner downloads Census 2021 population density from the Nomis API for every output area in Hillingdon and uses ONS lookups to keep the 201 areas that belong to the five wards in the table above. Density runs from 386.4 people per square kilometre to 96,585.4; the next highest is 53,928.6. Python then computes four averages: the mean, the median, a 10% trimmed mean (drop the lowest and highest tenth, average the rest) and a 10% winsorised mean (replace those tenths with the nearest remaining value).

On the real data the averages roughly agree, but the two measures of spread do not: the standard deviation is more than twice the robust one, because it squares the distance of a handful of extreme areas. Next the learner simulates a common accident. A chosen number of middling values are multiplied by 100, as if a decimal point had been lost.

One bad value in 201 moves the mean by nearly half while the median moves by under 1%. The trimmed mean holds until the share of bad values passes what it trims, then collapses: at 25% corrupted it is almost as wrong as the mean, and only the median, whose breakdown point is 50%, still looks like Hayes. Outlier hunting shows the same effect. The classic rule, more than three standard deviations from the mean, flags 2 areas in the real data; the robust version built on the median and the median absolute deviation flags 7, because the extremes inflate the very standard deviation that is supposed to catch them.

That instinct for a fragile number is what lets a Hayes teenager audit an AI report, and it is built by writing the Python yourself. The longer argument is in why teenagers should still learn to code in 2026.