How close are
our forecasts?
The current forecast is experimental. We have not yet established its error on independent student results, so there is no verified accuracy figure for the model you use today.
Historical benchmark: 16 October 2026 · Small sample
What this means for your forecast
Use your central estimate as a planning guide, with the range beside it. Neither is a guaranteed result.
- Live model
- Historical school scenarios with conservative upper-tail assumptions.
- What the range shows
- Variation across historical scenarios, not a validated confidence interval.
- What we are testing
- Whether corrections improve predictions on students the model has not learned from.
Explore historical model comparisonsPrevious engine and experimental corrections, not live-model accuracy
Explore the evidence
Lower error is betterFive lowest errors + previous engine · ATAR pointsScale: 0 to 2.23. Bars rescale for each view.
Select a bar to inspect its results. Use the full table below to explore every model.
These are small historical backtests, not a guaranteed error margin for your forecast. The live model now applies conservative course and school upper-tail assumptions. These historical scores do not validate that update.
Every model, side by side
Download results (CSV)Sorted by average miss. Select a model to read its results above.
Scroll the table sideways to see all three measures.
| Model | |||
|---|---|---|---|
| 1.21 | 0.85 | 1.53 | |
| 1.23 | 0.83 | 1.63 | |
| 1.28 | 1.12 | 1.61 | |
| 1.36 | 1.38 | 1.71 | |
| 1.41 | 0.80 | 2.07 | |
| 1.51 | 1.53 | 1.70 | |
| 1.52 | 0.41 | 2.27 | |
| 1.53 | 1.63 | 1.89 | |
| 1.54 | 0.94 | 2.07 | |
| 1.59 | 1.35 | 1.98 | |
| 1.81 | 1.13 | 2.55 | |
| 1.87 | 1.16 | 2.68 | |
| 1.92 | 1.66 | 2.69 | |
| 2.10 | 0.98 | 3.25 | |
| 2.17 | 1.51 | 2.95 | |
| 2.23 | 1.48 | 3.00 | |
| 2.23 | 1.48 | 2.98 | |
| 2.29 | 1.76 | 3.16 | |
| 2.33 | 1.69 | 3.26 | |
| 2.34 | 1.84 | 3.27 | |
| 2.34 | 1.84 | 3.27 | |
| 2.34 | 1.84 | 3.27 |
What these tests
can tell us.
Promising is not the same as proven. Here is how to read the numbers without overstating them.
How do we keep the test fair?
Each student is held out while the model learns from the others. All subjects belonging to that student stay in the same test fold. Parameter tuning, scaling, missing-value handling and PCA happen only inside training folds, using nested cross-validation.
The baseline uses historical scenario years strictly before the student’s HSC year, not the student’s own year’s school results. This differs from the five-year forecast currently displayed on the site. Historical source publication dates have not been independently verified.
How is flagged data handled?
A submission with inconsistent rank and mark information is excluded from every comparison on this page. The original record is retained privately for review rather than deleted. Excluding flagged data does not remove the uncertainty from a small sample.
What about the external calibration data?
Many imported achieved-ATAR records have no exact recorded ranks. The usable subset is tested separately in a direct-input benchmark. Those records do not have the matched baseline and true subject marks needed for the correction comparison shown here.
That separate small-sample test holds out account or source-thread groups. Its best average ATAR error was 5.65 points using RBF SVR. It is not directly comparable with this correction table. Forum or calculator estimates are not treated as actual outcomes.
Why not switch to the lowest-error model?
We searched 21 model families on a small sample of website submissions. Selecting the best of many candidates can make the winner look better than it will on new students. We need more verified achieved results and a test on untouched students before replacing the live model.
All records have equal training weight in this benchmark. HSC marks from one-unit courses are normalized to a 100-point scale. HSC and ATAR corrections are tested separately; better mark errors do not automatically mean better ATAR errors.
How often does the range contain the actual ATAR?
We have not established that yet. This test measures point-estimate errors, not how often actual results fall within the forecast range. The range is not a validated confidence interval, and we do not claim a percentage confidence.
Your actual result can make the next test better.
Finished your HSC? Contribute your ranks, final marks and ATAR to help us measure what works.