Michigan's September workforce data: what the numbers say, and what they don't

Michigan's September workforce data is complete now — the state release landed October 15 and the county and metro revision followed on October 22 — and the headline is a number that did not move. The unemployment rate held at 5.0%. Underneath it, the labor force shrank by about 26,000 people for the second month in a row, participation fell again, and employers filed a fresh round of announced layoffs and closures that take effect mostly in December. This month's article runs on a new edition of the research study, built for exactly this data, and it reads the flat headline the way this series always does: the rate is the least interesting part.

Editor's note: this series' editions are now named for their data month, not their publication month. This is the September edition — September 2026 data, drafted after the October 22 county/metro revision completed that month's data, and published October 30. Earlier editions used publication-month naming; the dating changed so the name always tells you which month's data you are reading.

Evidence cutoff: October 29, 2026 (America/Detroit). Michigan labor figures come from the BLS public API series behind Michigan Pulse and from the October 29 edition of the research repository — its verified release is reports-2026.10.29.134835Z — which captures the Michigan Center for Data and Analytics releases of October 15 (state) and October 22 (regional), 2026, and the Michigan LEO WARN listing retrieved October 29, 2026. Projections are the Michigan MCDA long-term projections (2024–2034) as captured in the edition and the U.S. Bureau of Labor Statistics Employment Projections, 2025–35, released August 27, 2026. Education figures are from the Michigan Center for Educational Performance and Information (CEPI) and the Michigan Department of Education, cited in place.

The rate is not the story — again, and worse

Michigan's September seasonally adjusted unemployment rate was 5.0%, unchanged from August, against 4.2% nationally. Read alone, that is stability. Read with the rest of the release, it is not. Employment fell by 27,000 over the month and the number of unemployed rose by 1,000, so the labor force declined by 26,000 to 4,826,439 (BLS, preliminary). Participation fell three-tenths of a point to 58.5%, and the employment-population ratio fell four-tenths to 55.5%. Over the year, Michigan's labor force is down about 178,000 people (−3.6%) — while the national labor force grew 0.3% in September alone.

That is the same concealment the August edition flagged in August, one month further along: a shrinking labor force keeps the unemployment rate deceptively calm. The rate can hold at 5.0% while the workforce itself contracts by roughly 26,000 a month. Anyone sizing a training program, a hiring plan, or a plant expansion off the headline rate is working from the one number in the release designed — by arithmetic, not intent — to hide the change that matters.

The employer survey tells a smaller and genuinely different story, and the edition keeps the two apart: total nonfarm payrolls fell by 6,000 in September to 4,487,000 (seasonally adjusted, preliminary), while manufacturing employment gained 1,000 on the month to about 583,800 (seasonally adjusted, preliminary). The state's own summary is that Michigan's labor market cooled as payroll jobs dipped across the third quarter, with declines concentrated in other services and professional and business services. Household survey and employer survey measure different populations; we report both, labeled, and we do not subtract one from the other to invent a missing-jobs figure.

The trend board: where Michigan stands, and where it came from

Four series, read the way a ticker page reads a watchlist. "Month" is September against August 2026; "Since Sep 2025" is the change across the window the series snapshot carries. September 2026 values are preliminary. October 2025 is missing from the two household series — that month was not collected during the 2025 lapse in appropriations — and is left a gap, never filled. One more discipline this month: the August values below are the revised August values, because a change computed against a superseded first print is a fiction. Every change is arithmetic on the published BLS series, computed for this article — not a BLS-published delta, and not a forecast.

One reading rule before the numbers: the arrows describe the direction of the number, not whether the move is good news. An up arrow on the unemployment rate is not a gain. Read each row on its own terms, and read the moves against last month's report, with the causes checked rather than assumed:

• The labor force is still the falling line — and its September fall is unexplained. It has now fallen in every month of the window the snapshot carries, and September's −25,807 follows August's revised drop of about 25,800. We checked the obvious candidate causes before writing this: the October WARN filings (below) announce layoffs with mostly future effective dates, so they cannot explain a September household-survey decline; the state release itself offers no cause, reporting the fall alongside a cooled third quarter. So the honest label is the one this series uses when the checking comes up empty: unexplained — a measured contraction with no verified cause yet, which is precisely why it should drive questions (retirements? discouragement? migration? survey noise that revision will soften?) rather than conclusions.
• Manufacturing gained on the month and is slightly down on the year. September's roughly 583,800 (seasonally adjusted) is 1,000 above August on the state release's count and 1,600 below September 2025 — the recovery from January's trough has stalled at roughly last year's level. No official release attributes the September move to a specific event, and the announced auto-supplier closures in the WARN review take effect in December and later, so they are a forward signal for this row, not its September cause: the September move itself is unexplained.
• Payrolls and participation are still telling different stories. Nonfarm employment sits essentially where it stood a year ago (+2,000 across the window) while the household labor force fell by nearly 178,000. That divergence is a survey-concepts fact before it is an economic one — residents versus jobs at establishments, with different treatment of self-employment and multiple jobholding — and the edition warns against converting it into a headcount of anything.
• The rate held. 5.0% for September, 5.0% revised for August, 4.9% a year ago. Flat is the finding; the movement is in the denominator.

Warning signals, not forecasts

Three soft indicators, kept in the category this series has always put them in. Nationally, the U.S. unemployment rate rose to 4.2% in September. The final September national consumer sentiment index was 48.1, down from 51.7 in August (University of Michigan Surveys of Consumers, as captured in the edition — a national survey; the university's name is not Michigan geography). And Michigan's own labor-market office describes a cooled third quarter, with payroll declines concentrated in other services and professional and business services.

These are warning signals: examine orders, hours, hires, and postings before acting. They do not say Michigan will lose jobs in October, and none of them is a forecast — ours or anyone's.

What the WARN notices say

Every cycle, this series reviews every Michigan WARN notice filed since the prior cycle — announced layoffs and closings, labeled announced, never written as jobs already lost. Seven notices were posted to the Michigan LEO listing between the October 8 check and October 29, announcing 930 jobs in total:

Two further notices posted on October 8 itself — Lear Corporation's Sterling Heights facility closure (67 announced jobs; Automotive News reported on October 9, 2026 that the plant is to close in July 2027, with transfers to Orion offered) and HealthEdge's Detroit permanent layoff (53 announced jobs, effective December 1, 2026) — were captured in last cycle's snapshot and are not double-counted in the 930 above; both belong in the planning picture below.

Read carefully, the table says two things. First, the weight is automotive supply — Dakkota, Inalfa, and Woodbridge together account for 639 of the 930 announced jobs — which matches the manufacturing row's stall without explaining its September move, because almost every effective date here is still in the future. Second, the December cluster (Dakkota's December phase, Inalfa, Woodbridge, HealthEdge — about 500 announced jobs effective within the same three weeks) is a dated, named, local event that education and workforce agencies can actually prepare for, which is more than most labor data ever gives them. Announced is not completed: notices are withdrawn, reduced, and delayed in practice, and this series will report what actually happened when the employment data can show it.

The outlook board: on the rise, falling — as the official projections have it

Everything in this section is a published projection: a statistical agency's modeled outlook over a decade, quoted with its horizon. BLS itself cautions that its ten-year figures carry substantial uncertainty and should be read for direction and relative size. Axiovex issues no forecast of its own. Neither board's vintage changed this cycle — Michigan's MCDA capture and the national BLS vintage are the same ones the August edition verified — so the boards hold against last month's report, and they are reprinted because a standing board that silently disappears in a quiet month would be its own kind of distortion.

Michigan — Michigan Center for Data and Analytics (MCDA) long-term occupational projections, statewide, 2024–2034, as captured and verified in the research edition. Annual openings include replacement demand, not just growth — which is why the openings column matters as much as the percentage:

Nationally — BLS Employment Projections, 2025–35 (released August 27, 2026). Total U.S. employment is projected to grow 3.5% over the decade. The fastest risers: nurse practitioners (+41.0%), solar photovoltaic installers (+36.5%), data scientists (+34.6%), wind turbine service technicians (+29.5%), information security analysts (+21.0%), and industrial machinery mechanics (+17.8%). The fastest fallers: word processors and typists (−34.4%), data entry keyers (−25.5%), foundry mold and coremakers (−23.0%), and payroll and timekeeping clerks (−15.9%). Office and administrative support is the fastest-declining major occupational group (−4.0%, about 752,100 jobs), with BLS attributing the pressure in part to continued integration of automation tools, including AI.

Three disciplines before anyone acts on these tables:

1. Decline is not the same as no hiring. Michigan customer service representatives are projected to shrink 6.8% and still post about 8,680 openings a year — the largest openings figure on either Michigan table — because replacement demand dwarfs the net change. A program sized by the percentage column alone will make exactly the wrong cuts.
2. The two boards agree where it matters. Data scientists and information security analysts sit near the top of both the Michigan and national risers; data entry sits near the top of both fallers; industrial machinery mechanics rise on both.
3. Michigan's board governs Michigan decisions. The state projections embed Michigan's industry mix — that is why foundry work, a rounding error nationally, is a real line here.

Full tables: BLS fastest-growing occupations, 2025–35 and BLS fastest-declining occupations, 2025–35. Michigan tables are in the research edition's verified projection capture, linked in the method note below.

Education analytics: the pipeline behind the workforce

Workforce numbers are downstream of education numbers. This board tracks Michigan's education pipeline with the same rules — published state figures, trend shown, vintage stated. One disclosure up front, because it is itself a finding: no new Michigan education release was verified this cycle. The CEPI graduation release and the MDE teacher-preparation release that this board rests on are annual, and their next vintages have not landed; the board therefore holds against last month's report, at the vintages stated, and nothing on it is presented as a September measurement.

Sources: CEPI graduation and dropout data released February 20, 2026 (Michigan Department of Education); Michigan Department of Education teacher-preparation data (release of September 29, 2025); pupil membership from the January 2026 consensus revenue estimating figures published by the House Fiscal Agency — those are the estimates the state budget is built on, labeled as estimates, not an audited headcount.

Read against this month's trend board, the board's two opposing trends now have a third beside them. Completion is improving — the class of 2025 graduated at the highest four-year rate since Michigan adopted the federal formula in 2008, and the educator pipeline is refilling. Cohort size is shrinking — the pupil-membership figure the state budgets against fell by about 12,500 this year and is estimated to fall by about 10,100 more next year. And now the labor force those cohorts enter is shrinking too, by about 178,000 over the year. Three different systems, three different vintages, one combined planning fact: Michigan is producing better-finishing, smaller cohorts for a workforce that is itself contracting. Capacity decisions keyed to any one of the three alone will be wrong in a predictable direction.

Underemployment and underutilized degrees

A graduation rate says nothing about whether the degree gets used. That is the underemployment question, and this series now tracks it every cycle, thoroughly and with its definitions named — because a degree that lands in a job that never required it is a workforce outcome, not a footnote.

The cleanest national measure is the Federal Reserve Bank of New York's labor market for recent college graduates (ages 22–27 with at least a bachelor's degree, not enrolled): in the second quarter of 2026, recent graduates faced about 5.6% unemployment and 42% underemployment — underemployment in the New York Fed's definition meaning employed in an occupation that typically does not require a bachelor's degree. Two disciplines attach to that 42%: it is a national figure, and no honest Michigan-specific equivalent series exists to quote in its place — applying the national percentage to a Michigan university's graduates would be exactly the kind of borrowed number this series refuses; and underemployment in this definition does not by itself mean low pay or involuntary part-time work — it measures degree utilization, not job quality. Michigan's own broader slack measure is the U-6 rate of 9.1% for the July 2025–June 2026 window (excluding October 2025, which was not collected) — a window average, never a monthly reading — which captures marginal attachment and involuntary part-time work that U-3 misses, on a different population than the graduate measure. The two cannot be added, averaged, or substituted for each other; together they bound the problem from two sides.

The supply side of the mismatch is visible in the geography data below: Michigan's degree production is real, and so is its leakage. The University of Michigan — the state's only institution in the Census Bureau's Post-Secondary Employment Outcomes release — retained 41.0% of its employed bachelor's graduates in-state one year after graduation (computed from PSEO R2026Q2, cohorts 2001–2021 pooled; see the geography section). A state can simultaneously have employers reporting hard-to-fill roles, graduates working below their degrees, and two-fifths of its flagship's employed graduates building their careers in Michigan while the rest leave. Those are not contradictions; they are a matching problem — field, place, and first-job quality — and matching problems are the kind that bridges, placements, and honest outcome data can actually move. What would sharpen this section, and is disclosed as missing rather than estimated: a Michigan-true, field-by-field completions-to-openings comparison. We could not verify a Michigan completions-by-field trend from its published source this cycle, so that row stays off the board until the underlying IPEDS release can be cited directly. An empty cell is better than a borrowed number.

Talent geography

The maps behind this section live on the Signals page's geography subsection — the full picture at axiovexsystems.com/signals/ — linked, never embedded, with each layer's vintage stated as of drafting:

• County unemployment (LAUS) — September 2026, preliminary; refreshed and verified at drafting on October 29, 2026 (the refresh advances the layer from August to the newest published month).
• County industry mix (QCEW) — 2024 annual averages; last-good at its annual cadence.
• Institutions and completions (IPEDS) — 2024 survey cycle; completions 2023–24; last-good at its survey cadence.
• Graduate outcomes (PSEO) — Release R2026Q2, cohorts 2001–2021 pooled; last-good at its release cadence.

Two computed insights, each from the computations this series is allowed to run on those committed layers, each labeled computed with its inputs:

The county spread narrowed — and the regional release says why at the regional level. Computed from the LAUS county layer (September 2026, preliminary): the highest county rate is Oscoda County at 7.3% and the lowest is Grand Traverse County at 3.2% — a spread of 4.1 points, narrowed from 5.1 points in August (Oscoda 8.6%, Grand Traverse 3.5%). The direction is corroborated by the state's October 22 regional release, which documents September rate declines across Michigan's labor market areas (Northeast Lower Michigan, Oscoda's region, fell from 5.8% to 5.0%, not seasonally adjusted). The regional pattern is therefore checked, not guessed; a county-specific cause for Oscoda's own improvement is not verified, and none is claimed. County rates are not seasonally adjusted — a narrowing autumn spread is a reading to watch, not a verdict that rural labor markets healed in a month.

Where the rising professional-services work already concentrates. Computed from the QCEW county-industry layer (2024 annual averages): for professional and technical services — the sector that employs the outlook board's fastest risers, data scientists and information security analysts — the highest location quotient among Michigan counties is Oakland County at 1.85, on 93,104 jobs in the sector. A location quotient of 1.85 means the sector's share of Oakland's employment is 85% above the national share. The practical reading for program planners: the state's professional-services gravity is already in Oakland County, so a data or cybersecurity program sited elsewhere is making an access argument, not a proximity-to-employers argument — and should be honest about which one it is making.

Five decisions for education and industry

The new edition's five decisions stand — this cycle's evidence revises none of them — and the new boards sharpen two of them. They are standing guidance, labeled as standing:

1. Redesign narrow task training first. Standalone typing, data-entry, and routine transaction-processing curricula need immediate review — while preserving transferable business knowledge and adding data quality, controls, exception handling, and customer problem-solving. The outlook board is the citation: data entry keyers are projected −26.4% in Michigan and −25.5% nationally; word processors and typists −34.4% nationally.
2. Retain foundations in AI-exposed professions. Software, accounting, legal support, engineering, media, and medical imaging face task change, not established decline. AI exposure alone does not establish occupation decline — and a 42% national graduate underemployment rate is an argument for better matching and first-job quality, not for fewer degrees.
3. Prioritize Michigan demand and replacement needs. Match program capacity to verified regional demand and openings — industrial maintenance, cybersecurity, software, data science, nursing, electrical work — not to percentage growth or national projections alone. Size by the openings column, not the percentage column.
4. Build skills bridges before layoffs. This cycle, this decision has a calendar. The WARN review names the workplaces and the dates: roughly 500 announced jobs with effective dates between December 1 and December 31, 2026, concentrated in automotive supply in Detroit, Auburn Hills, and Lansing. That is a five-to-eight-week runway — enough to assess incumbent skills, secure real receiving roles, and start paid supervised learning before separation, if the work starts now instead of in January. Real receiving roles, competence verified through work samples, outcomes tracked as wages and sustained employment — not course completions.
5. Plan by graduation cohort. Short credentials for 2027, associate pathways for 2028, bachelor's programs for 2030, longer pathways beyond. Decision horizons, not occupation expiry dates — and now read alongside all three boards: the cohorts are smaller, finishing better, and entering a labor force that shrank by 178,000 in a year.

The worked example stands — and the current vintages agree

The August edition used radiology as its worked example: official projections do not establish radiologist obsolescence, and the prescription was to preserve clinical preparation while adding AI evaluation, imaging informatics, and workflow assurance. The current projection vintages — the same ones verified in the new edition — support leaving that exactly as written. Michigan projects radiologists at +5.1% (390 → 410) and radiologic technologists and technicians at +6.4% (7,460 → 7,940, about 440 openings a year) over 2024–34, and the national vintage projects diagnostic medical sonographers at +13.9% for 2025–35. Imaging is a growth field with an AI-task-change problem, not a declining field — the distinction this series exists to enforce.

The pipeline behind the numbers

How this month's article was produced, stated plainly — including the leg that did not run.

The study was executed in the research repository. Because September's state release (October 15) and regional revision (October 22) advanced the newest Michigan reference month to September 2026, a new edition was produced: the October 29 edition, adding the September state and regional data and the full WARN review to the August edition's carried evidence base — 32 registered sources, 205 observations, and 36 projection rows, verified end to end with python3 scripts/reviewed_evidence.py, its append-only research ledger validated (head a4cf84b6f3008e18e117d66ec0b27baafbc9b2b06d6b1b93db332bce6534af57). The edition was published through the repo's reviewed release path as release reports-2026.10.29.134835Z; its downloadable assets were re-downloaded and verified against the release's published SHA-256 checksums — every downloadable file and every ZIP member matched (the ZIP's own entry is the sole exception, as a file cannot contain its own checksum) — and its manifest confirms version 2026.10.29.134835Z and source commit 5f7e8fb0d5cc80c1b87986c321e88305f2a3c323.

One leg did not run, disclosed as always: the study's Docker-based collection pipeline (collect, analyze, build) could not run on the workstation used for this edition, which has no Docker runtime; the step-by-step run record is kept with the edition's methods documentation in the research repository. The edition is therefore built from reviewed public sources retrieved directly — the BLS public API, the MCDA releases, and the LEO WARN listing, each hash-recorded in the edition's source registry — and the figures above are labeled readings from those published sources and the verified edition, not a fresh database ingestion. The external data pack snapshots committed in the website repo (JOLTS state, FRED, DOL claims, Census, BEA, College Scorecard, EIA) were read as evidence inputs; where a pack vintage was too old to state as current — state JOLTS reference period December 2025, for example — its figures were not used, and that restraint is the disclosure. The geography layers were checked layer by layer: the LAUS refresh advanced the county layer to September 2026 (preliminary) at drafting, and the QCEW, IPEDS, and PSEO layers are cited at their last-good vintages in the geography section — never silently as current.

That is the standard: every number traceable, every revision shown, every gap named, announced never written as completed.

This study's data and method are published in this article series. The full reports, evidence ledger, and source registry are held in the AXIOVEX/michigan-workforce-intelligence research repository, which is private; the underlying reports are available on request from [email protected]. The current edition is the October 29 edition (September 2026 data); its verified release is reports-2026.10.29.134835Z.