Occupation Ceilings Are a Cap, Not a Target: The New 189 Prediction Method

The short answer

On 8 October 2026 we froze a new prediction for the next SkillSelect subclass 189 invitation round, and we changed the method behind it. The previous model was too optimistic for niche occupations that have a large ceiling. The new model does much better on the past rounds we can check. It is a reference, not a guarantee.

This post explains what the previous model assumed, why that was wrong, what the backtest showed, which occupations moved, and how far you should trust the new version.

What the previous model assumed

Every year the government sets an occupation ceiling for the 189 visa: the maximum number of invitations an occupation can receive. The previous model treated each ceiling as a quota that would be used up. It then shared each round among occupations in proportion to their ceilings. An occupation with a big ceiling and a long queue was given a big slice of the round, at a low score.

That sounds reasonable. It was not how the rounds behaved.

What the backtest showed

When we replayed past rounds blind and compared the result with the official minimum scores per occupation, the pattern was clear. The ceiling works as a cap applied after the number of invitations has been set. It makes sure no occupation goes over its limit. It does not mean every occupation is invited up to its limit.

In other words, the size of the round decides how far down the scores go. The ceiling only stops an occupation that has already used its places.

Who the old model hurt

Niche occupations with large ceilings. The previous model gave them too many invitations, so it set their cut-offs too low: 6–7 points too easy in every round we could test it on.

• Early Childhood Teachers had about 1,027 189 places in 2025-26, against a queue that grew from roughly 5,900 to 10,800 over the year. A model that shares places by ceiling hands them a large allocation at a low score. In reality they were invited at 85 in August and November 2025, and the June 2026 round invited none of them.
• Small health occupations (optometrists, surgeons, midwives) were predicted at 65 with the whole pool invited. The real cut-offs were 75–85.

If you were reading one of those predictions, it was telling you what you wanted to hear. We should have caught it sooner.

The new model and its five-round test

The new model is built on that finding: it works from the size of the round and treats ceilings as a cap, not a target. We tested it blind on the five rounds we can check, grading each occupation against the official minimum. “Exact” means the predicted score equalled the official one.

The June 2026 round is the weak one. Only 47% exact, and we are not hiding it. Even there the new method stayed within 5 points for 89% of occupations, against 59% for the previous method. Over the two years we can check, the new model stood out, and that is why we think it has some reference value.

What changed for you

The live freeze uses the 30 September 2026 EOI pool plus one month of new EOIs at the pace of the last three monthly snapshots, a first round of 8,400–9,400 invitations expected by 31 October 2026, and the 2026-27 estimated ceilings. It was first frozen on 8 October without the month of new EOIs and re-frozen on 9 October with it: the scores are identical in both, the dates of effect and expected invitations below are from the 9 October freeze. The predicted cut-offs:

• 85 points for most occupations
• 80 points for most health and education occupations
• 65 points for construction trades

Each occupation also carries a confidence label from our cross-check. High-confidence predictions were exactly right 71% of the time in the backtest. Medium means the cross-check landed within 5 points. Low means treat the number as a rough guide and keep your own margin.

Six popular occupations, as the 9 October freeze predicts them (date-of-effect ranges cover an 8,400 and a 9,400-invitation round):

What moved between our 2 October and 8 October predictions

Of the 209 occupations that had a predicted cut-off in both freezes, 126 moved up, 58 stayed the same and 25 moved down. The 9 October re-freeze kept every one of these scores, so the comparison below holds for the live prediction. 64 smaller occupations moved from 65 to 85. Those are the niche occupations described above: the previous prediction had each of them at 65 with the whole pool invited.

The two freezes also differ slightly in inputs, so this is the change between our two predictions, not the method change alone. For the most popular occupations, ranked by expected invitations:

Some of these moved against you. We would rather say so here than have you find out from the page. Every occupation is in the table below; the date of effect and expected invitations for each are on the prediction page.

All 253 occupations: the live prediction, and what it was on 2 October

Same detail as the prediction page, from the 9 October freeze: predicted cut-off, date of effect, expected invitations and confidence, plus the 2 October cut-off and the change. Tap a row on a phone for the rest.

Limits

• Only five rounds can be checked, because SkillSelect keeps 24 months of EOI data.
• The previous method could not be tested on the 2024 rounds, because no 2024-25 ceilings were published. So the comparison rests on three rounds.
• Four of the five rounds use monthly data without dates of effect, which makes the grading blunter.
• The size of the first 2026-27 round and the Department’s priorities for the year are not announced. Both can move the lines.
• The 9 October freeze adds one month of new EOIs at the pace of the last three snapshots. If the round is held earlier than expected, fewer arrive and the 8 October freeze, kept on record, is the closer answer; if it slips into November, more arrive than we added. Scores are the same in both; the dates of effect are what this assumption moves.

The test that counts

The first real pre-published test of the new model is the round expected by 31 October 2026. Earlier freezes stay on record, and every prediction is graded against the official result on the rounds page.

Common questions