Behind the headline numbers
Headline statistics can hide opposite movements inside them, two parts moving apart: a split. On 13 Oct 2026 we fixed which twelve to watch, what our explanation expects of each and what the ordinary explanation expects. They are judged only on data published after that date, and we report all twelve every quarter, including the ones that go against us.
Fixed on 13 Oct 2026. Only data published after that date are judged. The first yearly verdict is due on 15 Nov 2027. The twelve, what each explanation expects of them and the rules for judging them have not changed since. The rules as fixed are public, with the code.
The twelve at a glance
The latest year of each, not judged: no full year of data published since 13 Oct 2026 is in yet.
- 1. Care jobs and care payMoved our way
- 2. How far jobs are splitting across industriesWithin the normal range
- 3. Graduates against high-school leaversNo clear change
- 4. Software job postingsNo clear change
- 5. Hiring against unemploymentNo clear change
- 6. Output per hourNo clear change
- 7. Labour's share of company incomeMoved our way
- 8. Dividends and interest against payNo clear change
- 9. Transformer prices against other goodsNo clear change
- 10. Land values against payNo clear change
- 11. Investment in computing against housingMoved our way
- 12. Saving, allowing for stock-market wealthNo clear change
Over the latest year, not judged, three of the twelve read our way, none the other way and nine showed no clear change. Our explanation fits four, the ordinary explanations two and neither six.
How easy it is to find a split
US payroll figures divide jobs into twelve main groups of industries. There are 2,047 ways to split them into two sides. For each, we compare how fast the two sides' jobs grew over a year with how far apart they normally grow, and count it when they have grown apart by more than twice the usual spread of such moves in normal times.
In a typical month from 1992 to 2019 without a recession, 0.4% of the 2,047 had grown apart that far, and in only 1 month in 20 was it more than 26.6%. In recessions the typical figure was 54.0%. In Sep 2026 it was 7.5%, against 23.2% a year earlier.
At present the share is within its normal range. Even in a typical normal month about nine of the 2,047 had grown apart that far, so anyone looking for a split will find one. Care against all other jobs ranked 684th of the 2,047 by how far it had moved beyond its usual range. That is why the list was fixed in advance and is judged only on data published afterwards, and why all twelve are reported. Splits also fade: of the 14 normal months above 26.6%, one was still above it a year later.
The twelve, one by one
Each is a gap between two numbers that a headline figure lumps together, or between a figure and what would normally go with it.
1. Care jobs and care pay
Latest year, not judged: moved our way Fits our explanation.
In the year to Sep 2026, health care and social assistance added 520,000 jobs, while all other jobs together fell by 24,000. Hourly pay in private education and health services fell 1.8% relative to pay across all private employers, below its usual range of −0.6% to +0.7%. Care's share of all jobs grew faster than its average in normal years.
Shaded: US recessions.
Hourly pay of employees who are not managers, in private education and health services together. The pay reading counts only while care's share of jobs is growing faster than its average in normal years. A shift of care jobs toward lower-paid work, such as home care, would also lower the average; that has not been ruled out.
- Our explanation
- Care takes in people whom other work no longer needs. Care jobs grow faster than usual while care pay falls relative to other pay.
- The ordinary explanation
- An ageing population. More older people need more care, so care jobs and care pay rise together.
2. How far jobs are splitting across industries
Latest year, not judged: within the normal range Fits the ordinary explanation.
In Sep 2026, 7.5% of the ways of splitting US jobs into two groups of industries had grown apart over a year far beyond their usual range, down from 23.2% a year earlier. In normal months from 1992 to 2019, 19 in 20 were below 26.6%; in recessions the typical figure was 54.0%. It is within the range of normal months, and it fell in a year without a recession, the direction the ordinary explanation expects.
Shaded: US recessions.
- Our explanation
- Work keeps moving between industries outside a recession, so the share stays above the range of normal months.
- The ordinary explanation
- The business cycle and the end of the 2021–22 hiring boom. The unusual moves of the past few years fade, and the share falls back into its normal range without a recession.
3. Graduates against high-school leavers
Latest year, not judged: no clear change Fits neither explanation.
In Sep 2026, 2.6% of Americans aged 25 and over with a bachelor's degree or more were unemployed, against 4.3% of those who finished only high school. The graduates' rate was 0.62 times the high-school rate, against 0.64 a year earlier: a change of −2.9%, within its usual range of −7.6% to +8.6%.
Shaded: US recessions.
Americans aged 25 and over; three-month averages. Judged in a recession too, when the ordinary explanation expects its clearest move.
- Our explanation
- AI reaches office work first, so graduates lose ground: their unemployment rate rises relative to the high-school rate without a recession.
- The ordinary explanation
- An ordinary slowdown. Less educated workers lose their jobs first in a slowdown, so the graduates' rate falls relative to the high-school rate.
4. Software job postings
Latest year, not judged: no clear change Fits neither explanation.
In Sep 2026, postings for software developers on Indeed stood at 77 and all postings at 103 (1 Feb 2020 = 100); a year earlier they stood at 65 and 103. Relative to all postings, software postings rose 18.6% over the year, within its usual range of −35.4% to +27.6%. That is the direction the ordinary explanation expects, though not yet beyond the usual range of this short and volatile series. They peaked at 229 in Feb 2022 and were lowest at 62 in Mar 2025.
Shaded: US recessions. The usual range comes from this measure's own history from 2021 to 2025, which includes the fall itself.
Job postings on Indeed. Published only since 2020, so reported beside the count of the other eleven, not in it.
- Our explanation
- AI writes more of the code, so postings for software developers keep falling relative to all postings.
- The ordinary explanation
- The end of the hiring boom and a tax rule. The fall was the end of the 2021–22 hiring boom, deepened by a tax rule that from 2022 spread the deduction for software development over five years. The rule was reversed in July 2025, so postings recover.
5. Hiring against unemployment
Latest year, not judged: no clear change Fits our explanation.
Hires were 3.3% of jobs in Aug 2026. With unemployment at 4.1%, hiring at its 2001–2019 relationship with unemployment would have been 3.9%. Over the year the shortfall narrowed by 0.07 points, within its usual range of −0.14 to +0.13 points. Interest rates were cut over the year.
Shaded: US recessions. The moves of 2004, 2010 and 2020–2023 run off the scale.
The rate unemployment implies comes from the relationship between the two from 2001 to 2019. A rate cut means the Federal Reserve's target at least half a point lower than a year earlier.
- Our explanation
- Firms need fewer new people, so hiring stays below what the level of unemployment would normally bring.
- The ordinary explanation
- Interest rates and uncertainty. High interest rates and uncertainty hold hiring back, and the shortfall closes within a year of rate cuts.
6. Output per hour
Latest year, not judged: no clear change Fits the ordinary explanation.
In the year to Q2 2026, output per hour in the US business sector grew 2.2%: output grew 2.5% and hours worked grew 0.2%. That is within its usual range of +0.5% to +3.2%. From 1990 to 2019 it grew 2.0% a year on average.
Shaded: US recessions.
Nonfarm business sector.
- Our explanation
- Machines take over tasks, so output per hour grows faster than its usual range.
- The ordinary explanation
- Lower immigration. Fewer immigrants slow the growth of hours worked, while output per hour grows within its usual range.
8. Dividends and interest against pay
Latest year, not judged: no clear change Fits neither explanation.
In Aug 2026, Americans received dividends worth 13.4% of their pay and interest worth 15.2%, against 13.6% and 15.3% a year earlier. Relative to pay, dividends fell 1.3% over the year, within its usual range of −3.7% to +15.1%.
Shaded: US recessions.
Pay is wages, salaries and benefits. Rent is reported in the public results but not judged: most of it is the rent home owners are counted as paying themselves, which overlaps the land measure.
- Our explanation
- Owners take a larger share of income, so dividends rise relative to pay.
- The ordinary explanation
- Higher interest rates. Only interest rises relative to pay, because rates are higher.
10. Land values against pay
Latest year, not judged: no clear change Fits neither explanation.
In Q2 2026, land in the United States was worth $30.1 trillion, 1.84 times a year's pay of all US employees, against 1.88 a year earlier. Relative to pay, land fell 2.1% over the year, within its usual range of −5.8% to +11.6%. The real interest rate did not rise over the year.
Shaded: US recessions.
Federal Reserve estimates: homes at market value less what their buildings would cost to replace, plus business land. A fall counts for the ordinary explanation only in a year when the real interest rate rose.
- Our explanation
- Land and sites are what stays scarce, so land values rise relative to pay.
- The ordinary explanation
- High real interest rates. Rising real interest rates lower what land is worth, so land falls relative to pay.
11. Investment in computing against housing
Latest year, not judged: moved our way Fits our explanation.
In the year to Q2 2026, real US investment in information processing equipment and software grew 16.8%, while real investment in housing fell 3.3%. Relative to housing it rose 20.8%, above its usual range of −1.3% to +18.9%.
Shaded: US recessions.
Information processing equipment and software, against residential investment, both after inflation.
- Our explanation
- Building the machines draws investment ahead of housing, and the gap widens while the boom lasts.
- The ordinary explanation
- Interest rates. Housing revives when rates are cut, and the gap narrows.
12. Saving, allowing for stock-market wealth
Latest year, not judged: no clear change Fits neither explanation.
In Q2 2026, Americans saved 4.4% of their disposable income, against 5.8% a year earlier. Their stock-market wealth was 3.1 times that income, against 2.6 a year earlier, above anything from 1960 to 2019 (at most 2.0). Fitted on those years, the relationship between the two would put the saving rate at −0.4%. The measure uses an effect taken from published studies instead: saving 1 point lower for each unit of that ratio, about one cent more spent a year for each dollar of stock-market wealth. On that effect, stock-market wealth implies a saving rate of 6.4%, and saving was 2.0 points below it. Over the year the gap fell 0.9 points, within its usual range of −1.3 to +1.0 points.
Shaded: US recessions. The moves of 2020–2022 run off the scale.
The level of the implied rate is set so that the gap averages zero from 1960 to 2019. Our explanation expects higher saving only once investment in computing falls relative to housing; until then, a year with no clear change fits neither explanation.
- Our explanation
- Later, owners save more of the income that moves to them, so saving rises above what stock-market wealth explains.
- The ordinary explanation
- The wealth effect. Saving moves with stock-market wealth and nothing else.
Adding up
Over the latest year the 11 measures with a history before 2020 give a count of +3: those that read our way less those that read the other way. Software postings, published only since 2020, are reported beside the count: over the latest year they showed no clear change. In the 13 years from 1990 to 2019 in which all 11 can be measured and no recession, or rise in unemployment of half a point or more, touches them (2004–2007 and 2011–2019), the same count ran from −2 to +2, and was 0 in a typical year. The latest year is reported, not judged; only the yearly verdicts from 15 Nov 2027 count.
How the twelve are judged
- Each measure's change over a year is compared with its usual range: the average change in years from 1990 to 2019 without a recession (from 1993 for graduates and 2002 for hiring, when their data begin), give or take its usual spread. Software postings, published only since 2020, are compared with their own history, which includes the fall itself. Beyond that range is a move our way or the other way; inside it is no clear change. The split of jobs across industries is judged instead on its level, against the range of normal months.
- Once the list is fixed, only data published after that date are judged, a year at a time and without overlap. Each year on 15 November we add up the moves of the measures with a history before 2020. The year supports our explanation if the count is higher than in 9 normal years out of 10, no recession, or rise in unemployment of half a point or more, touches it, and our explanation fits at least twice as many measures as the ordinary ones. It favours the ordinary explanations if the moves they expect came true in at least twice as many measures as ours. It shows nothing unusual if the count is no higher than in a typical normal year, or nothing moved either way. Otherwise it is inconclusive. A year touched by a recession, or by a rise in unemployment of half a point or more, gives no verdict. The verdict given on 15 November stands; later revisions to the data are reported beside it.
- We drop this line of evidence after two consecutive judged years that favour the ordinary explanations or show nothing unusual. Applied to the normal years from 2004 to 2019, that would have happened in 6 of 12 pairs of consecutive years. It would not disprove the argument; it would mean these numbers cannot tell the explanations apart.
- A split that continues is weak evidence. From 1990 to 2019, when one of these measures moved clearly beyond its usual range, it moved the same way the next year 59% of the time, and clearly beyond the range again only 17% of the time. Applied to the 13 normal years, these rules found nothing unusual in eight, an inconclusive result in four, a result for the ordinary explanations in one and a result for ours in none.
Reported, not judged
Our explanation makes no call on these that differs from the ordinary one, or the data are only a stand-in.
- The biggest companies against the rest. The Nasdaq-100 weighted by company size against the same 100 companies weighted equally: the ratio rose 51.6% between Dec 2019 and Sep 2026, rising 2.6% over the last twelve months. This is a stand-in. The usual measure, the S&P 500 against its equally weighted version, is not in FRED's free data, and the Nasdaq-100 holds only 100 companies, most of them in technology. Share prices also move with expectations and interest rates, not only with income.
- Large companies against small. Nasdaq's indexes of large and small US companies: the ratio rose 37.4% between Dec 2019 and Sep 2026, falling 1.6% over the last twelve months. Smaller firms borrow more at floating rates, so interest rates move this one directly.
- Computer imports. Imports of computers, peripherals and parts were worth 72% of US investment in information processing equipment in Q2 2026, against 33% on average in 2015–2019 and no more than 39% before 2020. Much of the equipment is imported, so a fall in this investment would land partly on the countries that make it.
- Electricity prices. Electricity prices for US consumers relative to all consumer prices: up 0.4% over the year to Aug 2026, up 10.7% since Dec 2019. Our explanation does not expect power to keep a lasting premium: electricity is produced, and grid connections are scarce only until more are built.
- Consumer spending. US consumer spending was 67.5% of GDP in Q2 2026, against 67.4% on average in 2015–2019. Our explanation expects owners to save more only after the boom in AI investment ends, so this is not judged.
- Property income in Sweden. Swedish households' income from property (mainly interest and dividends) less the interest and other property income they paid was 6.1% of GDP in the four quarters to Q2 2026, against 6.0% on average in 2015–2019 and a high of 7.1% in Q2 2022. Before deducting what they paid it was 8.3%. From Statistics Sweden's household sector accounts.
The stock split
From Dec 2019 to Sep 2026 the Nasdaq-100 weighted by company size rose 51.6% against the same companies weighted equally, most of it since the end of 2022. The split has not run one way in 2026: smaller companies' shares did better than large ones from January to July, then fell behind again in August and September. Our explanation makes no call on whether the biggest companies keep these gains, and nothing here is a view on where share prices go. Much of what these companies own is produced capital, such as chips, data centres and software, which competition and new factories make cheaper over time; some of it, such as sites, power contracts and patents, is not. In our explanation the lasting income goes to the owners of what cannot be produced: land and sites, and legally protected rights.
Spending
In Q2 2026 the richest 1% of US households held 32.5% of household net worth, the highest share since the Federal Reserve's series began in 1989. The richest 10% held 68.9%, below their peak of 70.2% in Q1 2019. Households saved 4.4% of their disposable income in Q2 2026, against 5.8% a year earlier: spending is outrunning income. Spending by income is not splitting: in Sep 2026 card spending per household was 5.5% higher than a year earlier for lower-income households, 4.9% for middle-income and 4.8% for higher-income households. Our explanation expects owners to save more only after the boom in AI investment ends, so spending is not judged.
Wealth: Federal Reserve Distributional Financial Accounts, via FRED. Card spending: Bank of America Institute, Consumer Checkpoint, October 2026.
Data: FRED, Federal Reserve Bank of St Louis (US Bureau of Labor Statistics, US Bureau of Economic Analysis, Federal Reserve Board, Federal Reserve Bank of Cleveland, Indeed Hiring Lab, Nasdaq), NBER recession dates, and Statistics Sweden. Readings as of 13 Oct 2026: monthly data to Aug or Sep 2026, quarterly data to Q2 2026. The rules in full, the code and every reading are public at github.com/wilsoniumite/labor. Nothing here is investment advice.