The claim worth checking is not "we have a score." Anyone can produce a score. The claim worth checking is whether the score anticipated what actually happened afterwards. Ours has been graded against subsequent federal unemployment outcomes on 90,266 resolved observations, and the answer is: yes at six months, and only weakly at three. We publish both halves of that.
What is actually covered
Every county in the United States, including territories — 3,221 scored records. Not a sample, not the metros, not the ones with interesting data. All of them.
Scores are rebuilt against BLS Local Area Unemployment Statistics. The most recent period currently available anywhere is June 2026 — county-level LAUS runs roughly two months behind publication, which is a property of the federal release schedule, not of our pipeline. Rows in our database were last written the same night this post went up.
Where the country currently sits
The live distribution across risk tiers, queried tonight:
- CRITICAL — 322 counties (10.0%)
- HIGH — 483 counties (15.0%)
- MODERATE — 806 counties (25.0%)
- LOW — 1,610 counties (50.0%)
Those proportions are close to round because the score is a cross-sectional percentile rank — each county is ranked against every other county in the same scoring pass, and the tiers cut at fixed percentiles. That is a deliberate design choice. It means the tiers describe relative position: "CRITICAL" means top-decile risk compared to the rest of the country this month, not a fixed absolute threshold that half the map could drift across in a bad quarter.
The spread is real, not cosmetic
Within a single scoring pass we routinely see counties in the high 90s carrying unemployment above 8% alongside counties scoring near 50 with unemployment near 3% — roughly a 50-point score gap tracking a five-point gap in the underlying labour market. Small rural counties with a few thousand people in the labour force show up at the top end far more often than their population share would suggest, because a single large employer represents a much larger fraction of local employment.
The inputs driving a score are labour-force size and trend, the unemployment trajectory over several horizons rather than a single month, seasonal adjustment against each county's own history, and state-level initial claims. That is the whole list.
Why some CRITICAL counties carry an asterisk
The score measures how fast a county is changing, not how bad it is right now. That is deliberate, and it has a consequence worth stating plainly: a county whose economy swings hard every year — a resort town, a fishing port, a national-park gateway — can reach the top decile on movement that is entirely normal for the season.
So we flag them. Counties above the 90th percentile of seasonal amplitude nationally carry a * everywhere their tier or score appears. The cut is 2.03; the national average is 1.32 and the distribution is heavily right-skewed (median 1.15, maximum 11.06), which is why we use a percentile rather than a standard deviation. On the current pass that flags 318 counties — and 55% of the CRITICAL tier. We publish that proportion rather than quietly raising the threshold until the list looked cleaner.
* Denotes counties with high seasonal variation — normal seasonal swings can drive tier placement; interpret trend direction with extra caution for these counties. The flag changes no score, no tier boundary and no model weight. It is a reading aid.
Correction, 24 August 2026: this paragraph originally also claimed industry exposure derived from QCEW employment data — how concentrated a county is in the industries most exposed to automation — as a driver of the score. It is not one. An audit of the live scorer found that the field reserved for that signal is currently hard-coded to zero for all 3,221 counties, pending a component that has not shipped. The score is real and the validation record below is unchanged — both were always computed from the labour-market inputs listed above, not from the industry term — but the original sentence credited the model with an input it does not have, and that was wrong. We would rather correct it in place than quietly edit it out.
The validation record — including the part that is weak
A score that has never been graded is a guess with a number attached. Ours is graded monthly against what BLS subsequently published, and the record is public here:
- Counties in the top decile of the score saw unemployment actually rise 91.2% of the time over the following six months.
- Counties in the bottom decile: 15.4%.
- That is a 76-point separation, across 90,266 resolved six-month observations (9,026 in the top decile, 9,027 in the bottom).
- The gradient between those extremes is monotonic — each decile rises more often than the one below it, with no inversions.
Now the part most vendors leave out. The same model measured at a three-month horizon is weaker, and we say so. It beats the best naive baseline by 7.4 points out of sample — real, but nothing like the six-month result. We are not going to sell you a three-month signal on the strength of a six-month one.
Correction, 25 August 2026: this paragraph previously said the three-month edge "collapses by roughly two-thirds when tested on periods the model was never fitted to." That is no longer true, and the original statement was understating the model for the wrong reason. A defect was found in how the seasonal index was loaded, which left two of the eleven model inputs silently duplicating two others across most of the backtest history. After correcting it and re-running the entire backfill and validation from scratch, the three-month edge is +7.8 points in-sample and +7.4 points on the held-out period — it holds up rather than collapsing. Measured against the seasonal norm instead of the raw change, the same signal is right 69.6% of the time against a 55.9% baseline, a +13.7-point edge on 19,326 held-out observations. The six-month figures above were re-derived on the corrected data at the same time and improved slightly. We are recording the correction even though it moved in our favour.
The honest framing of what this is: a backtest over 2023–2026, not a live track record. The distinction matters and we hold ourselves to it. Separately, as of this month every future scoring run is executed against versioned point-in-time data, so from here forward the record is provably free of hindsight rather than merely believed to be.
Why there are no county names in this post
Because the specific list is the product.
That is the whole reason, stated plainly rather than dressed up. A ranked list of which counties are currently in the top decile, with the drivers behind each one, is what a lender, an insurer, a workforce board or an investor would actually act on. Publishing it free would be pleasant for readers and commercially foolish.
There is a second reason, and it is not commercial. A "CRITICAL" tier is a screening signal — it means a county resembles the ones that historically deteriorated. It is not a prediction about any individual county, and naming a handful in a blog post invites exactly the misreading we spend our engineering effort trying to prevent. We apply the same rule elsewhere: in our work on AI broker exposure we describe categories of relay operator and decline to name individual businesses, for the same reason.
If you want your county
We will run a specific county and send you what the score actually is, what is driving it, and where it sits in the national distribution. That is a real report on a real county, not a brochure.
This is most useful if you are a county or municipal official looking at your own jurisdiction, a lender or insurer with geographic concentration, or an investor or workforce organisation with a reason to care about a specific place.
The same discipline, again
We keep arriving at one question across unrelated domains: what is this number actually grounded in?
In patent due diligence, a high similarity score between a claim and a product feels like evidence and is not. In AI agent reputation, we measured a public registry and found eleven addresses writing two-thirds of all the feedback in it. And in county risk scoring, a headline accuracy figure means very little until you ask which horizon it was measured at, whether the model had ever seen that period, and what a coin flip would have scored on the same data.
That is why the weak three-month number is in this post next to the strong six-month one. A score you cannot check is just another number asking to be taken on faith.