Build real AI projects

The curriculum points the same way

A Years 9 and 10 Digital Technologies elaboration describes "accessing and storing data from the Australian Bureau of Statistics in a format that is useful for analysis". Another asks students to consider "the complexities associated with training predictive models to capture events that occur with low probability".

Both are exactly the habits a real project needs: real public data, and humility about what a model can predict.

Why the baseline matters most

A model that scores 85% sounds impressive until a one-line rule scores 87%. Without a baseline, nobody can tell whether the AI added anything.

In our lessons, the baseline is written and scored first, before any model is trained.

Error grows with distance

For Australia as a whole, the model's average error rose from 8.9% one month ahead to 12.8% twelve months ahead. Its lead over the plain "this month again" forecast shrank from 4.5 points to 1.8.

Where the baselines win

Twelve months ahead, a simple baseline beat the model in 4 of the 8 states and territories. In the smallest markets the model sometimes did worse than repeating the latest month, because there was too little regular pattern to learn.

Some series barely forecast at all

In the ACT, approvals swing so much from month to month that every method's average error was above 95%. Saying so plainly is part of the project, and it echoes the curriculum's caution about rare, lumpy events.

What a learner writes up

The model is useful one to three months ahead for the national total and the larger states, and close to useless a year ahead in smaller ones. A one-line baseline is the right tool where it wins, and the report names those places.

At twelve months, "same month last year" and "this month again" are the same forecast, which is why those two columns match; noticing that is part of reading results carefully.

What we deliberately left out

The ABS also publishes the dollar value of approved building work. We used counts of dwellings only; projects at every age in our classes avoid money figures.

Seasonally adjusted series exist too. We used the original series so that the model had to learn the seasons itself.