Real labor data, refreshed monthly
Astris HR grounds its matches and forecasts in real public government labor data — official measures of job turnover, wages, and local demographics. That data shows up as retention forecasts, pay benchmarks, and local context. Nothing is a guess.
What we use
All of it is free public government data, updated automatically. Nothing is behind a paywall, and nothing is scraped from job-board sites.
| National job-turnover data (from the U.S. Bureau of Labor Statistics) | The official monthly measure of how often workers quit, leave, or get hired, broken out by 18 broad industries at the national level. Published early each month for the prior month. We use it as the baseline whenever more local, state-level turnover data isn't available for a job. |
| State-level job-turnover data (from the U.S. Census Bureau) | The same quit, separation, and hire rates, but broken down by state and industry — detail the national figures don't provide. For the same industry, tech workers in the Bay Area quit far more often than manufacturing workers in the Rust Belt. This is the primary source for our retention forecast; the national data is the fallback. |
| Local wage data (from the U.S. Bureau of Labor Statistics) | Typical annual pay for 95 detailed occupations across 60 metro areas, including the median and the low, high, and in-between pay levels. This drives the pay-context view and pay-percentile insight shown on every placement. |
| Local income data (from the U.S. Census Bureau) | Median household income for more than 410 metro areas. It replaces a hand-built table with live, real-world local cost-of-labor comparisons against the national figure. The hand-built table stays as a fallback. |
| State demographics (from the U.S. Census Bureau) | Median household income and the share of foreign-born residents in each state. This grounds the local-context view in the employer analytics dashboard. |
How the data stays current
Once a month, the system automatically pulls the latest figures from the government's public data services and saves them. The matcher uses these live figures whenever they're available, so the newest data takes over as soon as it's published.
The product ships with the most recent published figures already loaded (wage data from May 2024, turnover data from May 2025, income and demographics from 2022). It works with real numbers from day one.
Once a month, on the 1st, the system automatically checks the government's public data services and pulls in the latest figures.
When the model needs a wage or turnover figure, it uses the latest data if available and falls back to the starting data otherwise. Recently used figures are kept in memory for an hour so scoring stays fast.
Every update is recorded in an admin view — how many figures were refreshed, when, and the reason if anything failed — with the last-updated date shown for each source.
What it changes in the product
This data doesn't just sit in a table — it shows up where the case worker and employer make decisions.
| A retention forecast on every match | The forecast starts from real turnover in that state and industry (for example, roughly 1.8% of workers a month in Texas information-sector jobs), or the national figure when state data isn't available (about 1.6% a month). The candidate's own stability and fit then raise or lower that starting point. The forecast shows which source it used and for which period, so the estimate is traceable. |
| A pay-context view on every placement | Using real local wage data, the product shows the typical pay for the role in that metro area, a normal pay range around it, and where the offer sits compared with similar jobs in the same metro. It uses live local income data where available and a hand-built table otherwise. |
| Priorities that fit the role | What the match weighs most shifts by role. A driver role gives more weight to the commercial license and transport experience; an analyst role gives more weight to skills and language fluency. The role type is identified using categories grounded in the government occupation data. |
| A fairness monitor | The product compares the makeup of the group it selects against local population data, using the standard four-fifths test for adverse impact. This flags potential bias at the matching stage, not just after placement. |
How we test it
These results come from internal testing only. We built 70 sample profiles, each designed to have one clearly best-fit job, and checked whether the matcher finds it. These are not real placements, and an independent test on real cases is still pending — so treat the figures below as internal checks that the ranking works sensibly, not a measured accuracy claim.
For 60 of the 70 sample profiles, the matcher ranked the intended best-fit job at the very top. Internal check, not a real-world result.
The intended job appeared in the top 3 for 64 of the 70 sample profiles. Still to be validated on real placements.
On these sample profiles, the intended job always appeared in the top 5. Real cases will include variation these samples do not.
Standard measures of how well the ranking separates the intended job from the rest score well above chance, confirming the matcher orders the samples sensibly. These are internal diagnostics, not a benchmark accuracy figure.
When the matcher ranks the intended job first, it usually also labels it a strong match. Measured on sample profiles; testing on real cases is still pending.
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