The evidence

A school is asked to trust a projection about a student before anyone can check it.

The exam has not been sat. By the time it has, the decisions that mattered were made months ago. So the question worth asking any platform offering this is not how accurate it says it is. It is what was tested, what was found, and what changed as a result.

These are ours. No form, no email address, no summary version. The limitations are in the same documents as the method, because a methodology paper without them is marketing.

01

What actually predicts an HSC result

Every forecasting product makes a choice about what to feed the model. This sets out the choices we made, the evidence behind each one, and the inputs we deliberately refuse to use.

  • Why prior attainment carries the signal, and what the NSW Department's own modelling reports
  • Why we calibrate at subject level rather than whole school
  • Why we use no socio-economic measure, and why adjusting for prior marks does not make a model neutral
  • Why we do not use gender, though it is a larger factor than socio-economic status
Read the paperPDF · 7 pages
02

The maths behind the projection

Which statistical model we fit, why we chose it over the obvious alternatives, and what it is still wrong about. Written for a Head of Mathematics who intends to argue with it.

  • Isotonic regression, and why not least squares, splines, quantile regression or a mixed-effects model
  • How the confidence range is built, and why treating courses as independent understates it
  • Leave-one-year-out testing, and why the folds are years rather than students
  • A bias we found in our own engine, and the standard fix that turned out to remove none of it
Read the paperPDF · 7 pages
03

How the engine is tested

A projection is only worth what the testing behind it is worth. What we test, what the testing has found, and what we changed as a result.

  • Correlated errors, and a confidence range that was too narrow
  • A selection effect that nudged every projection upward, and the correction for it
  • An interval that returned impossible numbers at the sample sizes schools actually have
  • What we have not solved, stated plainly
Read the paperPDF · 4 pages
04

How Avantus protects student data

We handle data about real children. What we read, what we refuse to read, how one school's data is kept from another's, and what we have not finished.

  • The narrow slice we take, and the categories we decline even where your system would expose them
  • How database-level isolation keeps one school from seeing another's students, and what happens if it fails
  • How a connection is authorised, and how you revoke it without asking us
  • What we have not finished, including independent penetration testing
Read the paperPDF · 7 pages

Why these are not behind a form

A school cannot check a projection before the examination. It can check whether the people who built it went looking for their own mistakes, and said what they found. That is the only thing available to judge in advance, which is why we publish it openly rather than trading it for a contact detail.

Every change described in these papers made our own numbers less flattering. A wider range and a lower projection are harder to sell than the versions they replaced. We would rather a school held a number it can rely on than one that looks impressive in a demo.

If you are being shown a projection by anyone, ask them for the same thing.

See it on your own results. A demo is a guided walk-through of Avantus, with no commitment. We show you the projections, the range around them, and where the engine says it does not yet know enough.

If it is not right for your school, we will say so. You hear from us directly, not a sales team.

Request a demo
or email hello@avantus.com.au