Last hired, first fired, and first to know

Black unemployment has run at roughly twice the White rate for as long as the series has existed. That gap in levels is well documented. This paper asks a different question: whether the timing of a recession differs, so that Black workers’ labor market turns down before the aggregate does and recovers after it.

If it does, an indicator built on the headline rate is, in effect, waiting for a recession to reach White workers before it calls one. The paper traces that lag across every recession since 1982 and asks how much of it is about race itself.

Composition is what a group’s mix explains; the residual is what it does not

The composition piece prices Black workers’ own mix of income, education, sex, age and region at White workers’ unemployment rates within each of those cells. It answers: if Black workers faced White workers’ labor market, cell by cell, how much earlier would their recession still start?

The residual is what remains. It is not a direct measure of discrimination, since it also absorbs anything the cells leave out, but it is an upper bound on how much of the timing gap the observed characteristics can explain away.

The Racial Inequality of Recession

The full paper sets out the data, the turning-point framework and the decompositions, with robustness checks across smoothing windows, thresholds and reference groups.

The bottom of the distribution is the leading edge

Lower-income workers are concentrated in cyclically sensitive jobs and absorb hours cuts and early separations first, so their unemployment departs from its trough before the aggregate does.

Read from the bottom up, the income distribution is an early-warning system. By the time the headline rate has moved enough to trigger a conventional indicator, the poorest tenth of families has often been in a recession for months.

Separating the paycheck from the person

Low-income workers are also younger, less likely to hold a degree and more likely to work in industries that shed jobs early. The split below prices each decile’s demographic mix at economy-wide rates, so the composition piece is what that mix alone would predict and the residual is the head start that remains.

The Income Inequality of Recession

The full paper documents the decile series, the turning-point framework and the decompositions, and shows how the results change with the smoothing window, the threshold and the reference group.

Time, not only safety

A degree lowers the chance of losing a job in a downturn. It also changes when the downturn arrives. Workers without a high school diploma tend to see their indicator turn first, while those with a bachelor’s degree are often the last to register a recession and the first to leave it.

That ordering matters for policy timing: support keyed to the headline rate reaches the least-educated workers after their recession is already under way.

The Educational Inequality of Recession

The full paper sets out the education series, the turning-point framework and the decompositions, with robustness checks across parameters and reference groups.

A recession spreads across the map

The other papers in this series sort people: by income, by race, by education. This one sorts places. The Bureau of Labor Statistics estimates an unemployment rate every month for each of the country’s 3,136 counties, and every one of those series can be run through the same turning-point rule. Watched this way, a recession spreads across the map rather than switching on everywhere at once. Some counties turn early, some late, and some register downturns that never add up to a national recession.

Three patterns stand out. A handful of counties call recessions cleanly and almost never cry wolf. Counties that turn early tend to raise more false alarms, the same trade-off between speed and accuracy that governs national recession indicators. And places inherit the seasons of the industries they host, strongly enough to fake a recession even after standard seasonal adjustment.

A few places call recessions cleanly

Since 1993, 22 counties have crossed the recession threshold in each of the last three national recessions, 2001, 2007–09 and 2020, and at no other time. Many sit in North Carolina’s Piedmont, around Charlotte, Greensboro and Raleigh–Durham; others are suburbs of New York, Philadelphia and Omaha. Their mix of industries is close to the nation’s, so their labor markets move with the national cycle rather than with one plant, crop or tourist season.

Bellwethers are accurate but not early. On average they cross the threshold 2.5 months after the official peak. That is still well before the NBER announces a peak, which took four to twelve months after each of the last three recessions.

Early warning comes with false alarms

Some counties turn many months before the national peak. The cost is false alarms. A county whose labor market rides on a single plant, a commodity price or a tourist season will often cross the threshold early in a real recession, but it will also cross it when nothing national is happening. Pascal Michaillat calls the best attainable combinations of the two the anticipation–precision frontier: for any given lead, the most accurate a signal can be, and for any given accuracy, the earliest warning it can give.

For the typical county the penalty for speed is modest. Among counties that turned six months or more before the peak, 39 percent of signals were real recessions, against 44 percent for counties that turned close to it. At the frontier the trade-off is steep: every extra month of warning costs accuracy, and no county is both very early and always right.

Slow, but sure

The suburbs of Washington, D.C., show the trade-off in its purest form. Fairfax, Loudoun and Prince William counties in Virginia crossed the threshold in each of the last three recessions, but about 4 months after the official peak, later than most of the country. When they did cross, it was almost always real: 75 percent of their signals came near a national recession, against about 40 percent for the typical county. Across the whole metropolitan area, signals arrive later than in 74 percent of counties and are more accurate than in 91 percent.

The reason is likely the same on both sides of the trade-off. Many residents of Northern Virginia and suburban Maryland work for the federal government or for the contractors, consultants and services that depend on it. That work is insulated from the private-sector shocks that start most recessions, so a downturn reaches these counties late. And because only a broad, national downturn gets through that insulation, a signal from these counties is rarely a false alarm. The exception proves the rule: the only false alarms these three counties have sounded since 1993 came in late 2025, when the shock came from the federal government itself.

The District of Columbia proper behaves differently. Its unemployment rate is higher and more volatile, and only 33 percent of its signals were real. The accuracy belongs to the suburbs. At the opposite corner sits Elkhart County, Indiana, where most of the country’s recreational vehicles are built. RV sales are among the first things households cut when times turn, so Elkhart’s indicator crossed the threshold 11.7 months before the national peak on average. But only 55 percent of its signals were real: the same exposure to the private-sector cycle that makes it fast makes it jumpy.

A region can be in recession while the nation is not

Not every false alarm is noise. Some are real local recessions that never add up to a national one. Every Louisiana parish in the data crossed the threshold in 2014–15 after oil prices collapsed, and most of Wyoming’s counties in 2016, while the national unemployment rate kept falling. In November 2024, 55 percent of all counties were above the threshold, the highest share at any time since 1993 that was not during or just after a national recession, yet no national recession followed. Maryland’s counties were above it from early 2024 into 2026, and the District from early 2025 to early 2026.

Because these episodes are tied to places, policy can cause them and can also cure them. A state budget cut, a regulatory change or a federal decision about where jobs sit can start a local recession; state unemployment insurance, state spending and economic development can cushion one. A national indicator averages all of this away. A county view shows where a downturn is concentrated, which is the information a governor or a state legislature needs.

Places carry the seasons of the industries they host

A county’s labor market reflects the people who live there and, through them, the industries that employ them. Resort, farming, fishing and college towns have seasons that national data average out. BLS does not seasonally adjust county unemployment, and on the unadjusted series the seasons look like recessions: 39 percent of false alarms begin in January or February, after holiday hiring ends and outdoor work stops for the winter.

Standard seasonal adjustment helps but does not finish the job. After the same adjustment the site applies to every series, false alarms still cluster in the first months of the year (22 percent in January and February, against 17 percent if they were spread evenly), and the most seasonal fifth of counties, such as Skagway, Alaska and Mackinac, Michigan, still raise 2.3 times as many false alarms as the least seasonal fifth. Seasonal patterns that shift over time, as industries grow and shrink, escape a fixed seasonal adjustment. County data, and the sector data behind them, need seasonal adjustment designed for them rather than borrowed from national patterns.

The Spatial Inequality of Recession

The paper applies the turning-point rule to every county, maps the anticipation–precision frontier across places, and separates local recessions from seasonal and sampling noise. The statistics on this page use county unemployment rates from BLS Local Area Unemployment Statistics, January 1990 to August 2026, seasonally adjusted with STL and run through the six-month rule with a 0.5-point threshold. A signal counts as a real recession if it falls between 18 months before and 12 months after an NBER peak. They are recomputed every time the county data are refreshed. 313 counties are left out because the series attached to them in the source file could not be verified as their own.

What each measure adds

Each U-measure widens the one before it. U-1 and U-2 are narrower than the official rate; U-4 through U-6 add people the official rate leaves out: workers too discouraged to search, people who want a job but have not looked recently, and workers who are employed but working fewer hours than they want. U-4, U-5 and U-6 begin in 1994, when the redesigned Current Population Survey made them possible, so they cover the 2001, 2007–09 and 2020 recessions.

The intensive margin moves first

When demand weakens, firms usually cut hours before they cut jobs: shifts shrink, overtime disappears and full-time workers are moved to part time. Those workers are still employed, so the official rate does not count them, but U-6 does. Layoffs follow later, and the long-term unemployment captured by U-1 builds last.

The workers whose hours are cut first are disproportionately the same workers the rest of this site shows entering recession first: lower-income, less-educated, younger, Black and Hispanic workers. U-6 is, in effect, a composite of the leading edge of the labor market.

Discouraged workers, or shorter hours?

U-6 differs from the official rate in two steps: U-5 adds discouraged and other marginally attached workers, and U-6 then adds people working part time for economic reasons. The split below asks how much of U-6’s head start over U-3 comes from each step.

The answer is clear. In every recession since 1994, U-5 crossed the threshold in the same month as U-3: counting discouraged workers adds no early warning. All of U-6’s head start over U-3 comes from adding workers whose hours were cut. As the decomposition below shows, though, much of that is because adding them makes the series larger.

Scale, composition, or the intensive margin?

U-6 is about one and a half to two times the size of U-3, so the same 0.5-point rise is easier for it to reach. It also gives more weight to the groups that are most often underemployed. To separate these from genuinely earlier movement, the split below rebuilds U-3 and a U-6-style rate from 32 race, sex and education cells of the CPS microdata and asks three questions. How much earlier would U-3 fire if it were simply as large as U-6 (scale)? How much earlier if each group’s U-3 were scaled by that group’s own underemployment (composition)? And how much earlier does the actual U-6-style rate fire than that (within groups, the intensive margin)?

Most of U-6’s head start at a fixed threshold is scale. Composition adds little, and within groups part-time work does not consistently move before unemployment. When the threshold is scaled to each measure’s level (the Threshold scaling control in the full analysis below), U-6 and U-3 fire within a month of each other. The case for U-6 is less that it is faster than that it is larger: it shows how much more slack a recession creates, and for whom.

Who is underemployed, beyond their characteristics?

Two decompositions of the same CPS microdata, with the controls used across the site: demographics (age, sex, education, family, region, metro status, veteran status, nativity), then industry, occupation and sector, then income. The first asks who is underemployed but not unemployed: the gap between two groups in the share working part time for economic reasons, and how much of it their characteristics explain. The second asks whether a gap grows when the intensive margin is counted, by decomposing the same comparison with U-3 and with the U-6-style rate side by side.

The Inequality of Recession

The paper shows that, in a search across thousands of candidate recession indicators, a rule built on U-6 sits on or near the frontier that trades early detection against accuracy, and explains why: it captures adjustment on the intensive margin and the workers who feel a downturn first.

Broad measures pool the margins

Recessions reach some workers first. On this site’s other pages, unemployment among Black workers, young workers, workers without a diploma and lower-income households turns up before the national rate. A rule built on any one of those series also sounds more false alarms, because the samples are small, the series are noisy, and some shocks belong to one group or one industry rather than to the economy. A broad measure such as U-6 pools those margins without inheriting all of that noise. It reaches the workers who feel a downturn first because it counts more kinds of underutilization, not because it singles out any one group.

If breadth is the virtue, how far can it go? The ladder below adds four rungs to the official six, all built from the public CPS microdata and ordered from the narrowest to the broadest, jobless people first: everyone else outside the labor force who says they want a job (U-7); the own-account and the incorporated self-employed, two margins of independent and gig work (U-8 and U-9); and, at the top, workers whose degree is above what their occupation typically requires (U-10), the skills-mismatch measure in the research proposal. From U-3 up, each rung counts everyone the rung below it counts, and no one twice. The two narrowest official measures, U-1 and U-2, are slices of U-3, the long-term unemployed and job losers, so they can cross each other.

Bigger is not faster

Run through the same rule, the rungs above U-6 do not call recessions any sooner than U-6 does. U-6 itself crosses the threshold before U-3, by … in the Financial Collapse, because firms cut hours before they cut jobs. U-7 adds nothing to that head start: it crossed …. U-8 to U-10 are no faster: U-10 crossed …. With the threshold scaled to each measure’s size (the Threshold scaling control in the full analysis below), U-7 crosses … U-3 on average, …. Scaled to its volatility instead, U-7 crosses …, and ….

The classic three-month Sahm rule does not change the picture. With it, U-7 flags the Financial Collapse …, and it sounds …. How much the answer moves with the rule and with the definition of mismatch is on the data page.

The jobless rungs rise, the employed rungs shrink

The reason is in what each rung adds. The want-a-job pool is jobless, like the discouraged and marginally attached in U-4 and U-5, and since 1994 its twelve-month change has moved with U-6’s, with a correlation of …. The self-employed and the mismatched are slices of employment, and a recession cuts jobs: their parts of U-10 have moved against U-6, with correlations of … and …. Many mismatched workers hold the jobs a recession cuts first: retail sales, food service, driving, warehouse and construction work. When those jobs go, mismatched workers become unemployed, and U-10, which already counted them, does not rise. Mismatch then rebuilds in the recovery, as people take the jobs on offer rather than the jobs they trained for. From each NBER peak to a year after the trough, the mismatch part of U-10 …, and … in 2007–09, when the self-employment part … while U-6 rose … points.

… The broad rungs measure how much labor a recession wastes better than they mark when one starts. In April 2020 U-10 reached … percent and U-7 … percent, against … percent for the official unemployment rate.

A recovery in unemployment is not a recovery in utilization

Where the broad rungs earn their place is on the way out. Four years after the 2007–09 recession ended, … percent of that recession’s rise in the official unemployment rate was still there; for U-7 it was … percent. Four years after 2001 it was … percent for U-3 and … percent for U-7. … After 2020 every measure had all but recovered within two years.

Earlier warnings, more false alarms

The chart places every measure on the speed–accuracy frontier from the spatial page: how early its signal comes, on average, against the share of its signals that were real recessions. The narrow series sit low and to the right: unemployment among Black workers, young workers, workers without a diploma and immigrants turns up early but rings falsely more often. The official rate sits high and to the left, late but right every time since 1994, and so do U-1 and U-2, the narrowest official measures. U-6 buys part of the narrow series’ head start, and pays a similar price in false alarms; U-7 to U-10 add size without improving on it.

With three recessions and a handful of signals per series, one false alarm moves a measure by 15 to 25 points, so the pattern, not the ranking, is the finding. Several of the false alarms are recent: the broad measures crossed the threshold in 2024 and 2025, and, as of these data, the NBER has not dated a recession since 2020.

Degrees that go unused

Mismatch, the top rung, is not spread evenly. Among college graduates, … percent of Black workers and … percent of Hispanic workers hold jobs that do not call for their degree, against … percent of White graduates. For immigrants with a degree it is … percent. Among recent graduates aged 22 to 27, … percent of Black and … percent of Hispanic graduates are mismatched, against … percent of White graduates. Until U-10, the ladder counts most of them as fully employed.

What the restricted data would add

The public CPS can build the mismatch rung, but only coarsely. Occupations come in … harmonized categories, education stops at the highest degree, there is no field of study, and subgroup samples are too small for monthly mismatch series by race or by place. The restricted American Community Survey and LEHD files in the proposal would sharpen the level and the distribution of mismatch: detailed occupations, field of degree, local labor markets, and earnings records to test whether mismatched workers are in fact paid less, and move on when they can.

They will not sharpen the timing. The ACS is annual and LEHD quarterly, so monthly turning points still have to come from the CPS. The public-data ladder on this page is the monthly companion to the restricted-data measure.

The Case for U-7

The paper builds the ladder from public data, sets out the ten measures from U-1 to U-10, and asks what the four beyond U-6 add to the six the BLS publishes: a fuller count of wasted labor, a map of who bears it, and a recovery that runs longer than the headline rate shows.

Ten measures, one rule

U-1 to U-6 are the official seasonally adjusted series. U-1 and U-2 are slices of U-3, the long-term unemployed and job losers; they overlap and can cross each other. From U-3 up, each rung adds one more group to the numerator and, when the group is outside the labor force, to the denominator too, jobless groups first. A worker who fits two groups is counted once, in the lower rung: a self-employed worker who is part time for economic reasons is already in U-6, and a mismatched worker who is self-employed is already in U-8 or U-9.

The microdata cannot identify the marginally attached: the extract records why people outside the labor force are not looking, but not whether they could start a job or have looked in the past year. U-7 therefore replaces them with everyone who wants a job, a larger group that contains them, and takes the unemployed and the part time for economic reasons from the official series. Writing u = U/LF and m = MA/LF, U-5 = (u + m)/(1 + m), which gives m = (U-5 − U-3)/(1 − U-5), and with it the part time for economic reasons implied by U-6.

The want-a-job pool and the self-employed

U-7 widens the pool from the marginally attached to everyone outside the labor force who says they want a job: the people the CPS asks why they are not looking (WNLOOK). Since 1994 they have averaged … per hundred labor-force members, against … marginally attached. The BLS publishes a similar count, persons not in the labor force who currently want a job.

U-8 and U-9 follow the gig-work ladder in the draft code: own-account (unincorporated) and incorporated self-employment (CLASSWKR 13 and 14), counting the self-employed not already counted as part time for economic reasons. They are the weakest rungs as measures of underutilization, the incorporated self-employed above all: … percent of them hold a bachelor’s degree or more, against … percent of other workers, and … percent work in management occupations, against … percent. The draft’s other gig margin, multiple jobholding, needs the IPUMS variable MULTJOB, which the extract used here does not include; it is meant to replace incorporated self-employment as U-9.

Skills mismatch, step by step

Degree levels. Six steps of highest degree: no high-school diploma; high school, including some college without a degree; associate; bachelor’s; master’s; professional or doctorate.

What a job requires. For each of the … harmonized occupations in the CPS (OCC2010), the requirement is the degree at the occupation’s 75th percentile over the previous five years of data: three in four of its workers hold that degree or less. A worker is mismatched with a degree at least one level above it. This is a realized-matches definition, as in the proposal, rather than a job-analysis one: it asks what workers in the occupation actually hold.

A moving reference. Because requirements are re-measured every year, the rising education of the workforce does not by itself raise U-10. Each year’s update is phased in month by month, so the series has no January steps.

The 2003 recode. In January 2003 the CPS moved to the 2000 Census occupation codes, which shifts every occupation-based count. The level before 2003 is moved by the January jump net of the usual January change, ….

Counted once. Workers part time for economic reasons and the self-employed are left out of the mismatch count, because U-6, U-8 and U-9 already include them.

Do the microdata match the official series?

Rebuilt from the microdata and seasonally adjusted with STL, the unemployment rate stays within about … points of the official rate in a typical month (average difference …, correlation …), and part time for economic reasons within about … points of the official share implied by U-5 and U-6. The microdata use the CPS final weights and the BLS its composite weights and its own seasonal adjustment, so small gaps are expected.

How much the definition matters

The level of U-10 depends heavily on how mismatch is defined; the rungs below it do not depend on it at all. Requiring two degree levels instead of one cuts the mismatch part of U-10 from about … to about … points; the major-group crosswalk in The Case for U7.Rmd, which lets every sales and office job require only high school, puts it near …. Holding requirements fixed at their 1994–2026 levels lets rising education push the mismatch part of U-10 up over time, from … points in 1994 to … now, against … and … with requirements re-measured every year.

Timing depends on the choices less, but enough to matter for any claim that mismatch gives earlier warning. Across the definitions below, U-10 crosses the threshold in 2007–09 anywhere from …, with the main definition …. With the classic Sahm rule (a three-month average against the previous twelve months’ low), U-10 flags 2007–09 ….

The rule and the scoring

Seasonal adjustment. U-1 to U-6 are the official seasonally adjusted series. Each microdata component (the want-a-job pool, the self-employed, the mismatched, and every subgroup series) is a monthly ratio to the labor force adjusted as everywhere else on the site: a three-month centered average, then STL with a fixed seasonal pattern. October 2025, when the shutdown cancelled the household survey, is interpolated.

The rule. The site’s Modified Sahm Rule: a six-month average of the rate minus its lowest value over the past twelve months, with a threshold of 0.5 points. A recession is matched to the first crossing between 18 months before and 12 months after the NBER peak; a crossing outside every such window is a false alarm, and precision is the share of crossings since 1995 that were real. The full analysis on the U-7 page changes every one of these settings, including thresholds scaled to each measure’s level or volatility.

Files

Everything on these pages is built by data-source/U7_Ladder.Rmd from monthly tabulations of the IPUMS CPS Basic Monthly files, January 1994 to …, in data-source/u7/. The Rmd writes data/u7_series.csv, every measure and component, seasonally adjusted, and data/u7_summary.json, the numbers quoted here. With rebuild: true it recreates the tabulations from an IPUMS extract. Microdata: IPUMS CPS, University of Minnesota, ipums.org.

Are prime-age adults leaving the workforce?

…

…

The long decline is among men

…

That makes the answer depend on the yardstick. Measured against 2019, almost no prime-age adults are missing from the labor force today; measured against 2000, about … million are, all of them men. During the pandemic the count against 2019 reached … million.

Fewer out because of illness or disability

…

Drawing the labor force more widely

Each U-measure divides by the labor force, sometimes plus a group of people outside it. The panels below redraw that boundary six ways and recompute every measure:

• As measured. The official labor force: people with a job, and people without one who looked for work in the past four weeks.
• Everyone who wants a job. Adds every person outside the labor force who says they want a job, whether or not they have looked recently.
• Prime-age participation at its 2019 level, and at its 2000 level. Adds the adults aged 25 to 54 who would be in the labor force if participation in each age group (25–34, 35–44, 45–54) and sex were back at that year’s rate, counting only groups below it.
• Prime-age adults only. Every measure for ages 25 to 54 alone, as measured, which takes retirement and schooling out of the picture.
• Every prime-age adult. Counts every adult aged 25 to 54 as in the labor force, so U-3 becomes the share of the age group without a job.

People added to the labor force count as jobless from U-3 up. U-1 and U-2 count particular kinds of unemployment, the long-term unemployed and job losers, so the people added enter only their denominators.

The broader rungs move less

…

How these numbers are built

Data. Participation, employment, reasons for being out of the labor force and flows all come from the IPUMS CPS Basic Monthly files, January 1994 to …: civilians aged 25 to 54, final weights, adjusted with the site’s STL procedure. Flows link the same people in consecutive months (the CPS person identifier, with the same sex, an age the same or one year older, and the next rotation month) and are weighted by the earlier month. The CPS cannot link June to September 1995 to the months before, when it changed its household identifiers, so those months are interpolated, as are October and November 2025 around the shutdown, which cancelled the October survey.

Reasons for being out. The categories do not overlap: first everyone who wants a job; then, by the CPS labor-force status, people unable to work because of illness or disability, retirees, and everyone else (in school, caring for family, or another reason, which this extract does not separate).

Prime-age measures. The microdata do not identify discouraged or marginally attached workers, so the official counts are shared out by age in proportion to the people who want a job (for the discouraged, those not looking because of the job market). U-7 to U-10 use the same want-a-job, self-employment and mismatch definitions as on The Case for U-7; for ages 25 to 54, the mismatched self-employed are separated from other mismatched workers in the national proportions.

Everything here is built by data-source/U7_Ladder.Rmd, which writes data/u7_lfp.csv and data/u7_lfp.json.

Does a rule built on American data travel?

Claudia Sahm built her rule on the United States, where unemployment responds to output quickly and reliably. The version here is the one in Appendix B of the paper: a five-month average of the unemployment rate, less its lowest value over the twelve months ending five months earlier, so that the comparison window never overlaps the average, with an alarm at 0.6 points. Every country is scored against recessions dated the same way, from its own quarterly real GDP; the United States keeps the NBER’s dates. Applied unchanged, the rule’s accuracy varies as much as the labor markets do. … …

Accuracy here is the AUC, which needs no threshold: the chance that a month inside a recession has a higher indicator than a month outside one, so that a perfect signal scores 100 and a coin flip 50. The threshold matters for everything else, how many recessions are caught, how early and with how many false alarms, and the charts below let you change it.

The alarm arrives after the recession has spread

… Unemployment lags output everywhere. What differs from country to country is how far it lags, and whether it moves at all.

On time in some countries, late or never in others

…

As in the appendix, a recession counts as caught if the alarm sounds within nine months of its start or its end. Some of those alarms were already on when the window opened, left over from an earlier downturn: in a double dip, the alarm from the first recession is often still sounding when the second begins. They count as caught, but their timing says nothing about the second recession, so the medians on this page leave them out.

The rule works where Okun’s law bites

How much of a recession reaches the unemployment count is measured by pass-through, the Okun coefficient: how much the unemployment rate rises over a year for each point of GDP growth lost. … …

… The OECD’s indices score the law on the books, and the way firms actually adjust can differ from it. Short-time work schemes such as Germany’s Kurzarbeit let firms cut hours rather than jobs, whatever the dismissal rules. Dual labor markets such as Spain’s concentrate the adjustment on temporary contracts that are simply not renewed, even where permanent jobs are well protected. Both change how much of a recession reaches unemployment, and neither shows up in an index of dismissal law.

… The appendix reports a stronger gradient for these twenty countries; the difference comes from revisions to the data since it was written and from one correction to the recession dating, described in the notes below.

The same collapse, very different unemployment

… …

…

Calibrate to pass-through

An unemployment-based recession rule does not travel across borders by default. Where pass-through is weak the alarm is late and noisy, and moving the threshold only trades misses for false alarms: …

The same logic runs through the rest of the paper. Groups of workers whose unemployment responds strongly to output signal recessions early, and groups whose unemployment barely moves signal them late or not at all. For comparisons across countries, the useful grouping is by how much of a recession reaches unemployment, not by income level or by legal tradition.

Country by country

Sorted by accuracy, at a threshold of 0.6 points. Timing is the median number of months from the start of each recession to its first alarm, leaving out alarms left over from an earlier episode. Click a row to see that country above. The euro area, 21 members since Bulgaria joined in January 2026, is shown for reference and is not counted among the peers.

Three peers that count it

Only three of the peers publish unemployment by race, ethnicity or population group in their labour force surveys: the United Kingdom, by ethnic group, every quarter since 2001; Israel, for Jews and the rest of the population, every month since 2012; and Canada, by racialized group, every month since 2022. The panels apply the turning-point indicator to each group’s unemployment rate and mark when each group’s alarm first sounded, in the recessions those series cover and in Canada’s slowdown since 2023. Switch to the unemployment rates to see how far each group’s rate rose.

Data and methods

Unemployment. Monthly unemployment rates, seasonally adjusted, as a percent of the labor force: Eurostat’s une_rt_m for the members of the European Union, Norway, Iceland and Switzerland; the OECD’s harmonised unemployment rates for the United States, Canada, Japan, Korea, Australia and Israel; and, for the United Kingdom, whose Eurostat series ends in 2020, the Office for National Statistics series MGSX, for people aged 16 and over. The ONS publishes rolling three-month averages, labelled here by their middle month, so the British indicator is smoother and a little later than the others.

Recessions. For every country but the United States, dated with the Harding and Pagan (2002) version of the Bry–Boschan algorithm on the log of quarterly real GDP: Eurostat’s namq_10_gdp in chain-linked volumes, the OECD’s Quarterly National Accounts outside Europe, and the ONS series ABMI for the United Kingdom. A peak or trough must be the highest or lowest point within two quarters on either side, each phase must last at least two quarters and each full cycle at least five. A recession runs from the first month of the quarter after the peak to the last month of the trough quarter. One check is added to the appendix’s code, from the canonical Bry–Boschan rules: a phase counts as a recession only if GDP is lower at its trough than at its peak.

The rule and the scoring. Appendix B’s configuration: a five-month average, a twelve-month low lagged five months, and a threshold of 0.6 points. Months are scored from January 1995, once the indicator can be computed from data after that date, which puts the first scored month in September 1996 wherever the data go back that far. A recession counts as caught if the alarm sounds between nine months before it starts and nine months after it ends; an alarm episode that touches no such window is a false alarm. Two departures from the appendix’s code: a recession counts only if it reaches the months that can be scored (the appendix also counts recessions that ended before a country’s unemployment data begin, as misses), and a value exactly at the threshold sounds the alarm. Switzerland’s monthly series begins in 2010, Israel’s in 2012 and Iceland’s in 2003, so each is scored on fewer recessions than the rest.

Pass-through and institutions. The Okun coefficient is estimated for each country from 1995 on, by regressing the four-quarter change in its quarterly average unemployment rate on four-quarter growth in real GDP, in log points (Ball, Leigh and Loungani 2017). Employment protection is measured by the OECD’s indicators, version 4, for regular contracts, temporary contracts and collective dismissals, averaged over the years since 1995; the OECD does not cover Bulgaria, Croatia, Cyprus, Malta or Romania. Ireland’s GDP is swollen by the accounts of multinational firms, which weakens its Okun relationship.

Code and data. Everything on this page is built by data-source/International_Turning_Points.Rmd, which reads the source files in data-source/intl/raw/ and writes data/intl_series.csv, each country’s unemployment rate and indicator by month, and data/intl_summary.json, the recession dates, scores, pass-through, institutions and the 2008–09 and 2020 episodes. A scheduled workflow downloads fresh data from Eurostat, the OECD and the ONS twice a month and rebuilds both files. When you change the threshold, the page recomputes the alarms in your browser; at 0.6 points its numbers match the file’s.

A recession is not a switch

The NBER’s Business Cycle Dating Committee decides, after the fact, whether the economy as a whole was in a recession and when it began and ended. That is a judgment about the aggregate, and it is binary: a month is either in a recession or it is not.

Workers do not experience the aggregate. A downturn reaches some industries, places and groups of workers months before it reaches others, deepens unevenly and lingers for some long after the official trough. The same indicator, run on different groups, shows that a recession has a composition (who is in it), a distribution (how deep it is for each) and a continuous shape (how much of the labor force is in it at any moment).

Never called, but not ruled out

The Dating Committee never declared a recession in this period. But the indicators built to detect recessions in real time were not unanimous. In July 2024 the Sahm Rule reached its 0.5-point threshold. The Michez Rule of Pascal Michaillat and Emmanuel Saez, which requires unemployment to rise and job openings to fall together, entered its recession band in early 2024; in the working-paper version of “Has the Recession Started?”, they put the probability that the US was already in recession at about one in two with August 2024 data, and dated a possible start to March 2024.

Read as a yes-or-no question, the answer was no. Read as a distribution, the more useful question is: for whom was it a downturn, and how much of the labor force was in one?

The aggregate held; parts of the labor market did not

Run separately by race, income and education, the same turning-point rule shows which groups crossed into a downturn during these years even though the headline rate never did so convincingly. Pick a dimension below; the table counts the months each group spent above its threshold since January 2022.

Was it who these workers are, or something else?

A group can show a deeper downturn because of its mix: younger, lower-income or less-educated workers are more cyclical wherever they appear. The split below asks how much of each group’s gap over the period is predicted by that mix, priced at the reference group’s rates within each cell, and how much remains. The measure is continuous: the average level of the indicator over the period, or the number of months spent above the threshold.

Recessions are compositional, distributional and continuous

The vibecession is a useful test case because the aggregate answer and the lived answer diverged. A binary dating rule has to wait until a downturn is broad enough to move the whole economy. A distributional reading can say, month by month, how much of the labor force is in a downturn and who they are, which is closer to what people mean when they say the economy feels like a recession.

None of this says the NBER was wrong. It says that the aggregate is one summary of a distribution, and that indicators built on marginal workers carry information the aggregate averages away.

How the grid is built

The rule. Every rate row uses the site’s turning-point indicator, the Modified Sahm Rule: the six-month average of an unemployment rate less the lowest value of that average over the past twelve months, with the alarm at 0.5 points. A cell darkens as the indicator approaches the threshold and turns the deepest blue when the alarm is on. Region rows are different: they show the share of the region’s counties whose own indicators are above the threshold, as on the spatial page, and turn the deepest blue when most of them are, which counts as the region’s alarm.

The data. Labor utilization, race and ethnicity, gender, age, education, birthplace, industry and occupation are the Bureau of Labor Statistics’ unemployment rates from the Current Population Survey, by way of FRED. Where BLS publishes a seasonally adjusted series it is used as published; the rest are adjusted with the STL procedure used across the site. Industry and occupation refer to the unemployed person’s last job. Family income deciles come from the IPUMS CPS microdata behind the income page, and county rates from the BLS Local Area Unemployment Statistics behind the spatial page. The grid updates after each BLS and CPS release.

Reading it. The rows overlap: every worker appears in several of them, so the strip on top, the share of indicators signaling a recession, is a count of signals rather than of people. Small groups rest on small samples, and their indicators swing more and cross the threshold more often. Rows begin when their data do: industry and occupation in 2000, birthplace in 2007, and every indicator about a year and a half after its series starts.

More than a forecast

Turning-point rules are usually judged as forecasts: how early do they call a recession, and how often do they raise a false alarm? That question matters, and the indicators on this site are evaluated against it. But the same rules answer a second question that I find more interesting. Applied to the unemployment rates of different groups of workers and different places, they show who is in a recession, when it began for them and how long it lasted.

That is a distributional finding, and distributional findings tend to raise questions of policy. Economics is rightly cautious about prescription, and nothing here is meant as a verdict on how anyone should do their work. What follows are a few general observations about where this kind of analysis might be useful: to the committee that dates the business cycle, to the statistical agencies that measure the labor market, to the researchers who will study recessions in the future, and to policymakers who want support to reach people when they need it.

With gratitude, and a request for a little more detail

The National Bureau of Economic Research’s Business Cycle Dating Committee deserves to be congratulated, both for its work and for the transparency it already provides about the data it examines and the techniques it uses. Its chronology of peaks and troughs, in the tradition of Burns and Mitchell’s Measuring Business Cycles, is the reference point for nearly all research on recessions, including this work. The committee names the monthly indicators it considers, among them real personal income less transfers, nonfarm payroll employment and industrial production; explains each determination when it announces it; and maintains a page of frequently asked questions about its procedure.

That page is also candid that the committee works by judgment. A recession is “a significant decline in economic activity that is spread across the economy and that lasts more than a few months,” and “there is no fixed rule about what measures contribute information to the process or how they are weighted in our decisions.” The peak is “the month in which a variety of economic indicators reach their highest level.” The committee identifies turning points, in other words, without specifying a function that would locate them: how the indicators are combined and weighted, over what window, and against what threshold.

Judgment has served the committee, and the country, well, and there are good reasons not to bind it to a formula. This research may help fill part of the gap from the outside. The indicators at the heart of this work are turning-point functions with every parameter stated, applied in the same way to any group, place or measure. The formulas are written out in full in the methodology, and any reader can change the parameters in the interactive tool.

An astute macroeconomist could also reconstruct a good deal of the committee’s thinking from its members’ own scholarship, which includes careful work on dating turning points with statistical methods, among it Stock and Watson’s (2014) approach to estimating turning points from large data sets. Still, it might be helpful, and a further service to the field, if the committee shared some notes on what its formula generally looks like. The notes need not be exact, and they need not bind its judgment. A general description of the turning-point functions it has in mind, the algorithms or benchmarks it consults, and how much they weigh in its determinations would help researchers interpret its decisions and build on them.

A continuous value and a category for each recession

A second suggestion concerns how recessions are recorded. The chronology is binary: a month is either in a recession or it is not. Yet the committee’s own definition is graded. Its procedure notes that “each of the three criteria—depth, diffusion, and duration—needs to be met individually to some degree,” and that “extreme conditions revealed by one criterion may partially offset weaker indications from another.” A binary record keeps the committee’s verdict but sets aside the gradations behind it.

One way to keep them would be to record each recession, or each month of one, with a continuous value rather than a 0 or a 1: an index of intensity running from zero to one that summarizes the depth, diffusion and duration the committee already weighs. Values near zero would mark contractions at the edge of the definition, and values near one the most severe. How much of the labor force is in a downturn in a given month, the reading this site applies to the years since 2022 in its comments on the vibecession, is one natural candidate, since a share lies between zero and one by construction.

Alongside that value, each recession could be assigned a category describing its origin, drawn from a small, documented set: financial, monetary, energy and other supply shocks, public health, and so on. Together the two would convey much more than a date. A recession recorded in the financial category with an intensity of 0.7 would describe a severe downturn originating in the financial sector; one recorded as a supply shock with an intensity of 0.3, a milder contraction following a disruption to supply. Each piece of information bears on how an episode should be studied and on which responses are likely to help.

Committee members have already shown how such gradations can be measured. Christina and David Romer’s (2017) narrative series on financial distress in 24 OECD countries grades distress every half-year on a scale from 0 to 15, in categories that run from a credit disruption to an extreme crisis, rather than recording crises as simply present or absent. We will likely never be able to measure every component of every past recession. But American statistical practice is closely watched and widely emulated around the world, and beginning this work, perhaps even back-dating continuous values and categories for earlier recessions, would be a lasting gift to future researchers, here and abroad.

Measuring the margins of the labor market

Much of this work is possible only because of the Bureau of Labor Statistics. The Current Population Survey, which it sponsors with the Census Bureau, yields unemployment rates every month for a wide range of groups, and the Bureau publishes six alternative measures of labor underutilization, U-1 through U-6, that this site studies in its pages on underemployment. The microdata behind these statistics are public. That openness is a public good, and it deserves more recognition than it gets.

Three further steps might be especially valuable. The first would build on those alternative measures toward a more expansive understanding of slack. The broadest of them, U-6, adds to the unemployed the people marginally attached to the labor force, who want and are available for a job and have looked for one in the past year, and those working part time for economic reasons. It still leaves out many people with unused capacity to work: people who want a job but have not looked for one in the past year, the self-employed at the margins of independent and gig work, workers whose jobs do not use their skills or education, and adults in their prime working years who have left the labor force, some of whom would return if suitable work were available.

Economists have proposed ways to capture more of this, such as the nonemployment index of Hornstein, Kudlyak and Lange (2014), which counts everyone without a job and accounts for differences in their attachment to the labor force. A seventh, more comprehensive measure of slack, published alongside the existing six, would give a fuller account of how much of the country’s capacity to work goes unused, and of who bears that loss, in recessions and between them.

The second would reach further into the margins of the labor market: for instance, statistics on unemployed people whose previous work was informal, paid in cash or “under the table.” Because such workers often sit at the edge of the labor market, their employment might be more cyclical than that of workers in formal jobs, and their marginality might show signs of weakness sooner. Measuring them, even occasionally and in a supplement to the survey, could give an earlier and fuller picture of how a downturn begins.

The third would be to publish unemployment rates by family income as a standard data series. The survey already asks households about their family income, and this work uses those answers to estimate unemployment by income decile each month. The results are among its clearest: recessions begin sooner and last longer for workers in the poorest families. Regularly published rates by family income would let researchers and policymakers follow this dimension of inequality as routinely as they now follow unemployment by age, sex, race and education.

From a stimulus trigger to targeted support

Claudia Sahm designed her rule as a policy instrument. In Recession Ready, published in 2019 by the Hamilton Project and the Washington Center for Equitable Growth, she proposed it as a trigger for automatic stimulus payments to households, so that support would arrive automatically and early in a downturn. A clear, timely signal tied to a response is the idea this work tries to extend.

Policymakers who want to soften recessions face questions that a single national date cannot answer. Some are spatial: recessions reach some counties months before others, and local labor markets follow patterns of their own. Some concern inequality within the country, where the timing of recession differs sharply by income, education and race. Some concern particular sectors and the consequences a downturn brings to them. The same turning-point logic, applied to granular data, can speak to each: telling a state or a county when its own labor market has turned, showing which groups a downturn has reached first, or marking when the workers of a particular sector need support.

Used this way, the turning point that Sahm proposed as a national trigger can inform a wider range of multivariate, socioeconomic and spatial problems, from narrowing educational disparities to shoring up hard-hit sectors. The hope is a modest one: that better-timed and better-targeted information leads, over time, to better support for workers when employment falters, from unemployment insurance to the programs that help them back into work.

Offered with appreciation

Recession detection is a field built by careful people: the members and staff of the Business Cycle Dating Committee, the economists and survey staff of the Bureau of Labor Statistics and the Census Bureau, Claudia Sahm, Pascal Michaillat and Emmanuel Saez, and many others whose work this site leans on. These comments are congratulatory of that work. They suggest only that the tools the field has built for dating and measuring recessions can also help us see how recessions are distributed and, perhaps, help us respond to them a little better.

The committee, the rule and the data

A mentor beyond measure

I owe my deepest thanks to Trevon Logan, who has been a fantastic mentor in every sense of the word. His generosity with his time, his sharp and honest reading of every draft, and his steady belief in this project have shaped it from its earliest form to the papers it has become. It was his encouragement to decompose the gaps, to ask how much of each group’s recession is composition and how much is something more, that gave this work its center. Any scholar would be fortunate to have a mentor like him; I am grateful beyond words that I do.

The rule, and the guidance behind it

To Claudia Sahm, whose rule is the foundation of everything on this site: thank you for providing so much guidance to a young scholar venturing into a subfield that has been written about far less than it deserves. Your patience, your openness and your willingness to engage with a newcomer’s questions made it possible for me to build on your work with confidence. This project stands, quite literally, on what you created.

Thoughtful, forthcoming and transparent

Thanks to Pascal Michaillat for being an incredible mentor. He has been unfailingly thoughtful, forthcoming and transparent, and remarkably generous in meeting with me and answering so many of my questions as this research developed. His own work on recession detection set the standard I have tried to meet, and his candor and care have made me a better economist.

Thanks to Steven N. Durlauf of the University of Chicago for some of my earliest and most thoughtful feedback on this topic. He encouraged me to think deeper, broader and more ambitiously, and he would not let me settle for a simple statistical question when the work called for deeper theory.

Thanks to Emmanuel Saez for seeing the value in this research, for encouraging me to pursue further papers in this field, and for recognizing the promise of the spatial research. I am grateful for his lectures and for the guidance he gave in office hours during the Stone Center Summer Fellowship at the University of California, Berkeley.

Thank you to William A. Darity Jr. for being so incredibly helpful and so incredibly giving, for seeing the value in this research, and for inviting me to speak on it at the Journal of Stratification Economics Conference. Above all, thank you for creating space for me in the study of inequality. I will always be thankful.

Thank you to Miles Corak for coming to my early brown-bag seminars, when this research was in its infancy, and for pointing out that, whether or not I knew it yet, what I was really trying to do was characterize the economy. That advice, like so much of his advice, was sage.

Thank you to Matthew J. Baker for letting me pick his brain about the microeconomic theory behind this work.

Thanks to Justin Wolfers for helping me think clearly about the international panel, about how to index workers and the quality of the labor market across countries, and about how to respond to journal feedback.

To Wim P. M. Vijverberg, thank you for teaching me the econometrics I know, and for allowing me to bring this project into the classes you teach so well.

Thank you to Janet Gornick for always pointing me in the right direction, even when a research question sat outside her own area of expertise.

Thank you to Aaron Sojourner for his feedback on the general topic and for his early interest in what this work could mean.

Thanks to my professors at the CUNY Graduate Center. To Guillaume Haeringer: although you are a microeconomic theorist, you have been one of the most encouraging voices for this project, always supportive of the story I was telling with these macroeconomic statistics. To Thomas Herndon: thank you for our long conversations about the paper and for your counsel on how to navigate economics and academia. To Lilia Maliar: thank you for suggesting the machine-learning techniques in which you are an expert, to find which indicator among trillions was the right one, and for so generously offering your expertise. To Sebastiano Manzan: thank you for your insights into forecasting and how this work relates to it.

I also thank Matthew J. Baker, Miles Corak, William A. Darity Jr., Steven N. Durlauf, Janet Gornick, Guillaume Haeringer, Thomas Herndon, Lilia Maliar, Sebastiano Manzan, Aaron Sojourner, Wim P. M. Vijverberg, Bryan Weber and Justin Wolfers for their comments, insight and advice. Thank you to Marcus Cundiff of the City University of New York Graduate Center, and to Ricardo Martin, Angel Martinez, Joshua Velez and Ryan Lamidi of the City College of New York, for valuable research assistance. Thanks to the National Bureau of Economic Research Conference on the Economics of Race and Stratification, the Journal of Stratification Economics Conference and the American Economic Association Mentoring Program. Lastly, special thanks to Giorgi Ghviniashvili for his help building the interactive Modified Sahm Rule tool.

This work received no financial support from any outside agency or organization. The views expressed here, and any mistakes, are my own.