The verdict
In August we published forecasts for every 189 occupation pool we track, built on the July SkillSelect snapshot. August's data is in, and this is the honest issue. The whole 189 pool landed within 0.6% of our forecast. Underneath it, only 40 of the 68 occupation pools stayed inside their published 80% ranges — 58.8% coverage against an 80% target, down from 97% in July. Our average error of 5.77% still beat the naive “no change” baseline's 6.25%, but the margin was thin. The cause is a single mechanism the model did not include — lapsed invitations from the June round flowing back into the queue — and we walk through it below.
The 189 pool
The 189 Points-Tested pool, graded against the August snapshot. Two vintages matured this month: the July forecast at one month out, and — as promised in the July issue — the June forecast at two months out.
| Series | Predicted | 80% range | Actual | Error |
|---|---|---|---|---|
| 189 pool — July vintage, h=1 | 181,793 | 172,660–191,409 | 180,651 | +0.6% ✓ |
| 189 pool — June vintage, h=2 | 183,582 | 171,805–196,166 | 180,651 | +1.6% ✓ |
Across all 67 June-vintage 189 forecasts at two months out, the model's average error was 6.96% against 9.01% for the naive baseline — a clearer win than at one month — with 39 of 67 inside their ranges (58.2%). The same coverage shortfall shows up at both horizons, for the same reason.
Big occupations
The eight largest 189 occupation pools. Early Childhood Teacher was our best call among the big eight for the second month running — off by 0.8%. Two of the eight missed, in opposite directions: Chef shrank through the floor, and Registered Nurses nec grew through the ceiling.
| Series | Predicted | 80% range | Actual | Error |
|---|---|---|---|---|
| Early Childhood Teacher (241111) | 12,730 | 12,090–13,403 | 12,628 | +0.8% ✓ |
| Chef (351311) | 11,704 | 11,116–12,323 | 10,982 | +6.6% ✗ |
| Software Engineer (261313) | 9,815 | 9,322–10,334 | 9,603 | +2.2% ✓ |
| Civil Engineer (233211) | 9,288 | 8,821–9,779 | 8,858 | +4.9% ✓ |
| ICT Business Analyst (261111) | 8,120 | 7,712–8,550 | 7,885 | +3% ✓ |
| Mechanical Engineer (233512) | 8,025 | 7,622–8,450 | 7,661 | +4.8% ✓ |
| Accountant General (221111) | 7,762 | 7,372–8,173 | 7,377 | +5.2% ✓ |
| Registered Nurses nec (254499) | 6,731 | 6,393–7,087 | 7,188 | -6.4% ✗ |
Case studies
The headline number was our best call of the month: the model landed within 1,142 EOIs of the actual 189 pool. That accuracy is real, but it flatters the month — the occupation-level errors below it were large and pointed in opposite directions, and they cancelled out at the total. We would rather say that plainly than let the 0.6% stand alone.
The point forecast missed by 5.2%, and the actual landed five EOIs above our range floor. It counts as in-band, and it should — but a save this narrow is a warning, not a win. Civil Engineer and Mechanical Engineer finished within 40 EOIs of their floors too. The mature, older pools all shrank a little faster than the model's trend expected as two-year-old EOIs kept expiring.
The biggest miss of the month, by a distance. The Carpenter pool went from 697 to 1,182 in one month — 370 EOIs above our ceiling. It was not a wave of new applicants: 477 Carpenter EOIs were sitting in “invited” status at the July snapshot, and the pool grew by 485. At 65 points, 220 were invited and 241 came back; at 70 points, 157 and 145. The same pattern explains Registered Nurses nec (649 invited, pool up 671) and most of the other 26 misses. What happened is explained in the errors section.
The six-month blind test
Same blind protocol as our methodology post and the July issue, shifted one month forward: we refit the engine on data up to February 2026 only, and asked for the next six months. March to May landed within 1.3% of the point forecast. Then the end-of-financial-year expiry wave and the June round pulled the pool 7–9% below the point forecast from June through August — and the deliberately widening range absorbed all of it. Six months out, the model said 196,394; the pool sat at 180,651, inside a range whose floor was 171,772.
Three blind-test case studies under the same protocol — production fit, history cut at February 2026, six months predicted blind. One strong, one flattered by its range, one honest miss:
Engineering Technologist (233914) is our methodology post’s flagship case, re-run blind for the second month in a row. The first four months landed within 0.7% of the point forecast; the June expiry wave then took the pool below the forecast, and the endpoint finished 5.2% high but inside the range. Six for six again.
Chef (351311) scores six for six, and we do not think it deserves the credit. The model kept expecting the pool to grow while it shrank every month from June onward, so the error climbed steadily to +14.2% at six months out. The actual finished 12 EOIs above the range floor. That is the range doing the work, not the point forecast — a trend miss that the widening band happened to hide.
Registered Nurses nec (254499) is the honest one. March to May landed within 3%, then the June expiry wave cut the pool from 7,667 to 6,261 in a single month — 18% — and it fell straight through the range floor. July stayed below. August’s rebound of 671, most of it returning June-round invitees, closed part of the gap but not enough. Three misses in six, all on the same fault line.
The queue at your score
The pool totals above grade whole occupations. What most readers actually use is the queue at their score — how many 189 EOIs sit at or above 65, 70, … 95 points in their occupation. The model beat the naive baseline at every score level again, but by a whisker (12.14% vs 12.64% across 571 matured threshold forecasts), and the errors roughly doubled from July's 5.86%. Coverage held up better than the pool totals — 89.0% of actuals inside the 80% ranges — because the threshold ranges are built from pooled residuals rather than a scaled average. The weak spot was again the 80-point bracket (19.8% error, 77.8% in range).
| Score | Forecasts | Our error | Naive error | In-range |
|---|---|---|---|---|
| 65+ | 131 | 9.34% | 9.93% | 93.9% |
| 70+ | 120 | 10.45% | 11.07% | 89.2% |
| 75+ | 109 | 14.32% | 14.72% | 82.6% |
| 80+ | 81 | 19.78% | 20.10% | 77.8% |
| 85+ | 63 | 14.35% | 14.57% | 92.1% |
| 90+ | 43 | 6.41% | 7.04% | 100.0% |
| 95+ | 24 | 4.72% | 5.67% | 100.0% |
| All | 571 | 12.14% | 12.64% | 89.0% |
If you hold 65 or 70 points in a trade, nursing or teaching occupation, the queue at your score is larger this month than the model assumed — and most of the growth is EOIs that were already there before the June round, back in the queue and eligible to be invited again.
Where the errors came from
The 189 round of 4 June 2026 issued invitations that are valid for 60 days. At the July snapshot, 4,600 189 EOIs were still in “invited” status — invited but not yet lodged. The 60-day window closed on 3 August 2026. At the August snapshot, 55 were left. Some of those EOIs lodged a visa application; roughly 3,500 did not, and an EOI whose invitation lapses returns to the submitted pool. Across every 189 occupation, the number of EOIs invited in July predicts the growth of that occupation's pool in August with a correlation of 0.80. Carpenter, Registered Nurses nec, Secondary School Teacher, Construction Project Manager, Social Worker, Solicitor — every large upward miss is an occupation where the June round invited deep into the 65–70 point queue and many invitees did not lodge.
The downward misses are the other side of the same month: the mature pools — Chef, Accountant, Civil and Mechanical Engineer — kept shrinking as two-year-old EOIs expired, faster than the model's damped trend allowed. Growth concentrated in the pools the June round had emptied; shrinkage concentrated in the pools it had not touched. At the total, the two cancelled almost exactly. That is why the whole pool graded at 0.6% while 28 of 68 occupations missed.
Two consequences. In July we wrote that the ranges “could narrow”; we withdraw that. A one-month range of roughly ±5% is too tight for the month in which a round's invitations lapse, and we will not narrow anything until the ranges have survived one. And the model now needs to know about the lapse itself: our 189 fit already adds back an estimate of each round's removals so the trend sees gross inflow, but it treats the removal as permanent. We are testing a return term — the same adjustment in reverse, timed to the 60-day window — on history before it touches a published forecast. If it holds up, it ships; if not, we will say so here.
What matures next
In early October we grade the July vintage at two months out and the August vintage at one month out — every month, wins and losses alike. The next 189 round is expected by 30 September, so the September snapshot will carry a round's removals as well as the lapse we just described. Method details: how we test our forecasts. The July issue: July 2026 scorecard. Live numbers: EOI trends.