Simple Victorian school demand forecasting from public data

What the forecast is for

School planners need to know how many children will live in each area, and in which year levels, five to fifteen years from now. That is the lead time needed to buy land and build a school. The usual chain is: forecast the children in each small area, then multiply by a capture rate (the share of those children who will enrol at a given school or sector), then split by year level.

We tested only the first link, using public data. We measured how accurately the school-age population of each area can be forecast, and which method does best. We did not forecast enrolments at any school, because that needs enrolment data that is not public.

Data used

• Estimated resident population (ERP) by area, age and sex, 2001 to 2025. ABS, Regional population by age and sex, 2025 (cat. no. 3235.0), from the ABS Data API, dataflow ERP_ASGS2021, retrieved 9 October 2026. Licensed under Creative Commons Attribution 4.0. Source: Australian Bureau of Statistics.
• Areas are Statistical Areas Level 2 (SA2) on the ASGS Edition 3 (2021) boundaries. Victoria has 522 SA2s. Most hold between 3,000 and 25,000 people, about the size of a suburb or a country town.
• Boundaries for the map are the ABS ASGS Edition 3 (2021) SA2 digital boundaries for Victoria, GDA2020, from abs.gov.au (file SA2_2021_AUST_SHP_GDA2020.zip), retrieved 9 October 2026. Licensed under Creative Commons Attribution 4.0. Simplified for the web, so borders are approximate.
• Ages come in five-year bands at this level: 0 to 4, 5 to 9 and 10 to 14 for children, and women aged 15 to 44 for the ten-year forecasts. The ABS does not publish single years of age for SA2s in this series.

The method in plain words

Three forecasts were made for every area, plus a blend of two of them.

• Naive. The number of children stays where it is now. A useful method has to beat this.
• Trend. A line fitted to the area's own history of children aged 5 to 14, from 2001 to the forecast year, starting from the latest count. The slope is damped, so it flattens out over time instead of growing forever.
• Cohort-survival. Children age into the next band. If an area had 1,000 children aged 0 to 4 five years ago and has 1,050 aged 5 to 9 today, the ratio of 1.05 sums up everything that happened to that group: families moving in and out, and a few deaths. The model applies the same ratio to today's 0 to 4 year olds to forecast the 5 to 9 year olds in five years, and the same for 5 to 9 into 10 to 14. This is the Hamilton-Perry method, a standard tool for small areas. Ten years out, the 0 to 4 band does not exist yet, so it comes from the area's ratio of young children to women aged 15 to 44, applied to a forecast of those women.
• Combined. The simple average of the cohort and trend forecasts.

For the cohort model, forecasts one and three years ahead sit on a smooth path between today's count and the five-year cohort forecast. Every setting was fixed before any forecast was scored, so nothing was tuned to the test results.

How the back-test works

A back-test picks a past date, makes a forecast using only the data available up to that date, and scores it against what happened. Here the forecast date moves one year at a time from 2006 to 2024. From each date, each method forecasts each area's children aged 5 to 14 at 1, 3, 5 and 10 years ahead. That gives 30,347 scored forecasts across 505 areas.

• Error is the absolute percentage error: how far the forecast was from the published figure, as a share of the published figure. The headline compares each method's median error across all the forecasts at a given horizon. At five years that is 7,463 forecasts (each area, each starting year), so half did better and half did worse.
• Small areas are left out. An area is scored only when it had at least 100 children aged 5 to 14 on the forecast date. That drops parkland, industrial and airport areas, and new estates before their first families arrive.
• Same cases for every method. A forecast is scored only when all four methods could make one, so the comparison is like for like.
• Growth groups use no hindsight. Each area is grouped by how fast its school-age population grew in the five years before the forecast date: growing fast (3% a year or more), growing (0.5% to 3%), or stable or falling.

Results

The cohort model beat the trend at every horizon on the median. One year ahead its median error was 1.5% against 2.0%. Five years ahead it was 5.5% against 7.8%, which is 29% lower. Ten years ahead it was 10.7% against 12.7%, 16% lower. Forecast by forecast, it beat the trend about six times in ten up to five years out, and in just over half of ten-year forecasts. The trend barely beat the naive forecast: its median error was within a few tenths of a point of "no change" at every horizon.

The trend line also leans low. Its median forecast was 2.7% under the actual figure at five years and 7.2% under at ten, because it damps growth that kept going. The cohort model's median miss was within about 1% either way.

Ten years ahead, the simple average of the two did best, at 9.9%. Averaging a method that overshoots in growth areas with one that undershoots cancels some of both errors.

In established areas, growing slowly or not at all, the cohort model cut five-year error by about a third: 4.6% against 6.9% for the trend. Those are most areas and most forecasts.

In fast-growing areas the cohort model loses its lead. At five years the two methods tie on the median, near 16%. At ten years the cohort model's median error is 44% against 29% for the trend, and some of its misses are enormous. When an estate goes from a handful of toddlers to hundreds in five years, the ratio between the two counts is huge, and the model projects that surge forward as if it will repeat. In Point Cook South, a forecast made in 2010 expected about 104,000 children in 2015. There were about 2,100.

Those few large misses dominate any total. Weighted by the size of each area, the five-year cohort error was 24% of all children, against 16% for the trend and 12% for no change at all. A planner who used the raw cohort model in growth corridors would have been badly wrong in the fastest-growing areas. Practitioners do not apply cohort ratios where the starting numbers are tiny. Those areas need a different method, set out below.

The map shows where this happens. The trend did better in pockets right across the state, but the areas where both methods failed sit almost entirely in Melbourne's new estates, to the west, north and south-east.

What public data cannot see

• Estates not yet built. Tarneit North had 8 children aged 5 to 14 in 2015 and about 3,050 in 2025. Every method based on past population forecast roughly the same small number. Only a dwelling pipeline (land releases, subdivisions and building approvals) can see this coming.
• Where children go to school. Population is not enrolment. The share of children at government, Catholic and independent schools, and which school they choose, varies by area and changes over time. School zones, new schools and travel time all move it. None of that is in ABS population data.
• Year levels. The ABS publishes five-year age bands at this level, so this test cannot split forecasts by year level. Single-year detail needs Census counts or enrolment records.
• Policy changes. Changes to school zones, school starting age, migration settings or housing policy can shift demand quickly. A model fitted to the past will not know about them until they show up in the data.
• Hindsight in the history. The ABS revises its estimates after each Census. This test uses today's revised series, so each past forecast started from better data than a planner had at the time. Real forecasts would have done somewhat worse. Estimates after the 2021 Census will be revised again after the 2026 Census.
• Small-area noise. SA2 estimates for single years are themselves modelled by the ABS. Small areas swing a lot from year to year, and the smallest were left out of the scores.

How this scales with agency data

With an education agency's own data, each gap listed above has a known fix. Each fix can be back-tested the same way as the methods on this page.

1. School enrolment census. Enrolments by school, year level and student home area give capture rates for each area and each school, and let the cohort model run on year levels directly: this year's Year 3 becomes next year's Year 4, scaled by the ratio seen in past years.
2. Dwelling pipeline. Lot releases, subdivision approvals, building approvals and precinct structure plans tell you where homes will be built and when. In new estates, forecast children from new dwellings, using the number of children per home observed in earlier estates at the same age. Hand over to the cohort model once the estate has settled.
3. Capture rates and travel time. Allocate each area's children to schools using observed capture rates, school zones and travel time, so a new school or a boundary change can be tested before it is made.
4. Accuracy testing every year. Store every forecast as it was made, on the data available at the time. Each year, score last year's forecasts against the new enrolment census and report the error by area type and horizon. If a method gets less accurate, the yearly score shows it within a year.

The back-test on this page is the template for that last step. It takes minutes to rerun when new data arrives.