Astris HR Model

How the models work together

Four different methods each estimate the chance a placement will last 90 days. Astris HR merges their answers into one score, giving more weight to the methods that have proven more reliable on real results. The version that scores a match is always the same version that past outcomes have tuned. Figures on this page come from internal test data; a benchmark on real placements is still to come.

A note on the numbers

Please read this before the figures below.

The methods described here are currently built and tested on simulated data, because enough real placement results have not yet accumulated. Every accuracy figure on this page is an internal test resultmeasured on data the methods had not seen during training — not a claim about real-world accuracy. A benchmark on actual placements is still to come and will replace these figures once real outcomes are available.

The four methods

Each estimates the chance that the same candidate–job pairing lasts 90 days.

Shared input
Candidate+Job

All four methods read the same short list of retention-related facts — how quickly the person is learning the language, whether they can reach the job by transit, how close a same-background community is, how long since they arrived, and whether funded training matches the role. Characteristics such as race or national origin are never used to score anyone.

Timing method
share: tuned

Looks at how the chance of a placement lasting changes as the weeks go by, then reads off the odds of reaching the 90-day mark. It is the only method that models the whole timeline rather than a single yes/no answer. Internal test accuracy about 73 percent; real-placement benchmark still to come.

Pattern method
share: tuned

Builds up its estimate from many simple rules layered on top of each other, which lets it catch combinations of factors that a straight-line model would miss. Internal test accuracy about 81 percent; real-placement benchmark still to come.

Transparent method
share: tuned

A straightforward model whose reasoning can be read factor by factor, so you can see exactly why it landed where it did. It serves as the explainable reference. Internal test accuracy about 83 percent; real-placement benchmark still to come.

Learning-network method
share: tuned

A small network, built in-house, that discovers its own patterns in the same inputs and adjusts itself as it trains. Internal test accuracy about 80 percent; real-placement benchmark still to come.

Combined retention score
One 90-day estimate, blended from all four
Each method's share is tuned by results, and every method always keeps at least a small share.

This combined estimate becomes the retention part of the eight-part match engine. Because every method keeps a minimum share, no single method can ever be shut out entirely as the balance is tuned.

How the balance is tuned

Real placement results shift how much each method counts.

Step 1
Record

Whenever a placement reaches the point where we can tell whether it lasted 90 days, we save what each method had predicted alongside what actually happened. Only the most recent stretch of results is kept in view.

Step 2
Grade

For each method, we check how much worse the combined estimate would have been without it. That tells us which methods have been pulling their weight lately and which have not.

Step 3
Adjust

We nudge each method's share toward the ones that have been more reliable. The changes are kept small and gradual, and no method is ever pushed all the way down to zero.

Step 4
Confirm

Before a new set of shares replaces the old one, we wait until the evidence clearly favors it. We keep gathering results until the case is convincing rather than switching after a fixed number of placements.

Step 5
Apply

When Astris HR scores a new match, it uses the exact same shares the tuning produced. The version that scores matches can never drift apart from the version that results have tuned.

What each method contributes

Four ways of looking at the same retention question.

  1. Method 1The timeline view
    Timing method

    How does the chance of a placement lasting change week by week, and where does this specific candidate–job pairing stand at the 90-day mark? This is the only method that follows the full timeline instead of giving a single yes/no answer.

  2. Method 2The pattern view
    Pattern method

    Which combinations of support factors tend to lead to a placement reaching 90 days? This method is good at spotting patterns that only show up when several factors line up together.

  3. Method 3The transparent view
    Transparent method

    What does a model say when its reasoning is fully readable, factor by factor? Because you can see exactly how it weighs each input, it serves as the explainable reference.

  4. Method 4The self-learning view
    Learning-network method

    What does a small, in-house network discover in the same inputs on its own? It adjusts itself as it trains and can capture relationships the simpler methods do not.

A separate tool, not part of the score

Grouping candidates never rates an individual.

Astris HR also includes a grouping tool that sorts candidates into clusters who share similar support needs, so caseworkers can plan for groups rather than one person at a time. This is a descriptive tool only: it is not one of the four methods above, it never affects the combined retention score, and it never rates an individual for a match.

See Astris HR in your workflow

Run it against your candidate base or a sample JD. We'll set up a tenant in under a day and show you the matches your existing tools are missing.

Request a demo