Has the Sahm rule ever given a false recession alarm?

Sahm wrote the rule as the switch for a policy. Her chapter of Recession Ready, a 2019 book from The Hamilton Project and the Washington Center for Equitable Growth, proposed that Washington send people lump-sum payments automatically once unemployment started to climb (Sahm, 2019). The payments would total 0.7% of GDP, with more in later years if the downturn deepened. Congress had sent such payments in 2001 and again in 2008, each time designed by hand once the recession was under way (Sahm, 2019). An automatic payment needs a trigger that fires early, and with hundreds of billions of dollars at stake, one that almost never fires outside a recession (Sahm, 2024a).

She built the trigger from the unemployment rate because it is closely watched, measured the same way for decades and published early in the following month, while GDP comes later and is revised often (Sahm, 2019). The three-month average smooths out the month-to-month noise, and measuring the rise from the past year’s low lets the trigger move with the jobless rate a healthy economy can sustain, which shifts over the decades (Sahm, 2019). She then spent several weekends testing thresholds in a spreadsheet: she wanted the smallest rise that fired early in a recession and not outside one (Sahm, 2024a). On the data available at each moment, lower thresholds of 0.3 and 0.4 points both fired outside a recession, in 1976 and 2003. Half a point did not, in the years from 1970 that she tested (Sahm, 2024a). Others named the rule after her, and she has written since that the checks, not the indicator, were the point (Sahm, 2022a).

A trigger was needed because the official date of a recession comes late. NBER’s Business Cycle Dating Committee names the month a recession began only once it is confident one has happened, and it leans most on payroll employment and on real personal income less government transfers (NBER, 2026a). Since 1980 it has announced each peak between four and twelve months after the fact (Fig. 3). The December 2007 peak, for one, was announced on December 1, 2008 (NBER, 2026b).

Forecasters already had signals built to warn rather than to confirm. Estrella and Mishkin (1998) found that, more than a quarter ahead, the slope of the Treasury yield curve was the clearest single predictor of recessions among the variables they tested. Engstrom and Sharpe (2019) later refined it into a spread between near-term forward rates. Sahm (2019) set such market measures aside for her trigger because they give more false positives, in part because the Federal Reserve responds to them. Hamilton (2011), surveying attempts to date turning points as they happen, found the task harder than many academics presume, because data are revised and economic relationships shift. He stressed testing a model on the data as they were actually released at the time, which is why this test also runs on the rates as first published.

On today’s rates the pattern Sahm found runs back to 1949 (Fig. 4). Since then the indicator has spent 223 months at or above 0.50. All but three of them fall within a recession’s wider window, which takes in the slow fall of unemployment after a recession and, with it, the 2003 crossing. The three months outside every such window are July, August and September 2024.

The rule outgrew its purpose after the pandemic. As the Federal Reserve raised interest rates in 2022, commentators began to quote it as a law of the economy. Larry Summers wrote that whenever unemployment rises by half a point within a year it goes on to rise by two, as Sahm (2022b) quotes him. Sahm answered that the rule was a historical pattern, not a law of nature, and that a cycle as unusual as this one could break it (Sahm, 2022a; Sahm, 2022b).

On today’s rates the indicator crossed 0.50 in all eight recessions since 1970. Each call came one to eight months after NBER’s peak, and not one came before it (Fig. 5). The typical gap was three months. Sahm drew the same line in 2019: the trigger gives little advance warning, and signals a downturn long before it is officially recognized (Sahm, 2019).

Lined up on NBER’s peaks, the 11 recessions since 1953 share one shape (Fig. 6). In most of them the indicator sat near zero in the year before the peak, then crossed 0.50 in the first months of the recession and kept climbing. Within two years of every peak it reached at least 1.5, the high of the mild 1990 recession. The 2024 panel, lined up on the July 2024 call, rises to 0.57 and turns back.

The one recession in which the crossing came first is 1960: the indicator crossed in November 1959, five months before the April 1960 peak. On the rates of the time Sahm dates that crossing to December 1959 and counts it as a false positive that a recession soon followed (Sahm, 2019; Sahm, 2024b). It came less than a year after the 1957 alarm had ended, so the count fixed in advance folds it into that earlier alarm, and the 1960 recession, which the claim does not cover, has no call of its own.

The first-published rates tell the same story for the recessions they cover. Their calls came in June 2001, April 2008 and April 2020, between two and four months after each peak (Fig. 7). Sahm (2019) reported the same months for the 2001 and 2007 to 2009 recessions from the rates as they stood at the time. The revisions since moved the 2008 call two months earlier, to February, and left the other two where they were.

The 2008 call shows what the speed is for. The Economic Stimulus Act became law on February 13, 2008, two months into a recession no one had yet dated, and the rebates reached households between May and July (Sahm, 2019). On the rates as first published, the rule’s call came with the April jobs report, published on May 2, 2008, as the first checks went out. NBER named it in December 2008 (NBER, 2026b).

The one recession without a call of its own is 1981. The alarm raised in February 1980 stayed on until April 1981. The indicator then spent six months under the line, through the July 1981 peak, and crossed again in November 1981. Under the 12-month count, 1981 is scored as missed. Sahm’s own chart treats November 1981 as a fresh trigger (Sahm, 2019).

The same count keeps 2003 off the record. In mid-2003, a year and a half after the 2001 recession ended, the indicator came back to the line after seven months below it. On the first-published rates it read 0.50 in June 2003 and 0.57 in July. On today’s rates it touched 0.50 in July and August. The count reads both as the tail of the 2001 alarm, so neither is a new call. With a fresh call at every crossing, the way Sahm’s chart counts November 1981, 1981 becomes a hit on today’s rates, and 2003 a false alarm on both sets of numbers. The verdict is the same under either count. On the rates as they stood at the time, revisions included, Sahm puts July 2003 at 0.47, just short of the line (Sahm, 2024b).

The indicator has also come close and turned back. Seven times since 1950 it climbed to 0.30 or more without reaching the line (Fig. 8). Two of those climbs, in September 1952 and October 1979, came within a year of a recession. The other five came to nothing, the two most recent in December 2023 and November 2025. On the rates as first published, November 2025 read 0.43, against 0.35 on today’s rates.

Revisions have also erased two near misses that looked close at the time. On the rates as they then stood, Sahm counts a reading of 0.47 in November 1967, and one of 0.50 in November 1976 that falls just short of the line before rounding (Sahm, 2024b). On today’s rates neither year came near: the highest readings were 0.23 in November 1967 and 0.20 in December 1976.

The jobs report of August 2, 2024 put the July unemployment rate at 4.3%, up from 4.0% in May (BLS, 2024). That carried the indicator to 0.53, over the line (Fig. 1), and it peaked at 0.57 in August. By October 2024 it was back under the line, after three months at or above it.

On today’s rates the call stands, by the narrowest margin the rule allows. BLS’s annual revisions have since trimmed the July and May rates by a tenth each, and the July reading lands exactly on 0.50 (Fig. 2). That still fires a rule that counts any reading at or above the line as a call. August’s 0.57 would have fired it regardless, so the false alarm does not turn on the revision.

After every earlier call, unemployment kept climbing. In the months after each recession call since 1953, the three-month average rose at least 2.23 points above the low the rule had measured from, the mildest case being 2001 (Fig. 9). After July 2024 it rose 0.88. The monthly rate peaked at 4.5% in November 2025 on today’s rates, and at 4.6% as first published, then eased.

Sahm had sketched this outcome in advance. In December 2022 she wrote that the rule would break if unemployment rose half a point and then stopped at around 4%, and that if it was ever going to break, it would be in 2023 (Sahm, 2022b). It fired in 2024 instead, and unemployment topped out at 4.5% on today’s rates.

No recession has been dated since. NBER names a peak only once its committee is confident a recession has occurred (NBER, 2026a), and its last announcement, on July 19, 2021, dated the end of the 2020 recession (NBER, 2026b). By September 2026 the unemployment rate stood at 4.2%, and the indicator was back at zero.

Revisions decide calls this close. Each January BLS revises five years of seasonally adjusted rates (BLS, 2026g). Of the monthly rates first printed since 1994, 57% still stand, and the rest have moved by a tenth or two. Across the first-published series that moves the indicator by 0.04 points on average, and at the readings that mattered by up to a tenth (Fig. 10). In 2008 the revisions raised the readings and pulled the call two months earlier, to February. In 2001, 2003 and every reading shown since late 2023 they pulled the first prints down, by as much as 0.08 in November 2025.

Every month BLS asks the unemployed in its household survey why they are looking for work. Job losers were laid off or let go, or finished a temporary job; job leavers quit; reentrants are coming back after time out of the labor force, and new entrants are looking for their first job (BLS, 2026c). In a recession most of the rise comes from job losers, because employers cut staff while people who might have started looking hold back (Sahm, 2024b).

In 2024 entrants added almost as much as job losers. Of the rise behind the July 2024 call, job losers added 0.28 points and entrants, new and returning, 0.25 (Fig. 11). As shares of the four groups’ total, that is 55% against 49%. At every recession call or crossing from 1980 to 2020, entrants made up at most 21% of the rise, and job losers at least 72%. Counted in points, job losers added less to the 2024 rise than at any recession call since 1970: 0.28, against at least 0.36 in 1974 and 2008. Sahm (2024c) found the same split a month after the call: entrants made up about half the increase, where in recent recessions most of it came from people who had been laid off.

The early 1970s come closest to 2024. Entrants made up 36% of the rise at the March 1970 call and 38% at the July 1974 call. Job losers’ shares in those two, 57% and 60%, were barely above 2024’s, though the gap between the two groups was still more than three times as wide (Fig. 11). Sahm (2024c) makes the same point: recessions have begun while the labor force was expanding, as in the 1970s. The nearest match to 2024 is the July 2003 crossing, which was followed by no recession: entrants made up 51% of its rise.

The new job seekers arrived as the country’s population grew at its fastest pace since 2001. The Census Bureau puts net international migration at 2.8 million in the year to July 2024, up from 1.7 million two years earlier (US Census Bureau, 2024). It made up 84% of the population’s growth. The survey’s reasons for unemployment do not say where people came from, so Sahm (2024b) put it cautiously: the steady share of entrants was consistent with immigrants adding to the supply of workers. When BLS brought the household survey up to date with the new population estimates in January 2025, its count of the labor force rose by 2.1 million at a stroke, while its rates moved little (BLS, 2025a).

New jobless claims, the quickest count of layoffs, were not rising. A worker who loses a job files a claim for unemployment insurance with their state, and the Labor Department adds up the new claims each week (US Department of Labor, 2026). In July 2024, new claims were 2.1% lower than a year before (Fig. 12). At the 1990, 2001 and 2008 calls they had been higher, by between 6.7% and 40.9%. In April 2020 they were more than ten times as high.

Employers’ own reports show the same. In July 2024 layoffs and discharges ran at 1.1% of employment, the same as a year earlier (BLS, 2026b). From 2001 to 2019 they ran that low in only 4 of 228 months. Quiet layoffs do not rule out a recession on their own, though. At the February 2008 call they were only 1.4%, and Sahm (2024c) notes that layoffs often pick up late in a recession, while hiring falls from the start. She read the 2024 pattern as a rise in the supply of workers, and not only weaker demand for them, pushing the rule over its line, while warning that a rise in entrants can carry some of both.

Each state’s own unemployment rate can carry the same rule. BLS publishes a seasonally adjusted rate for every state and Washington, DC, from 1976 (BLS, 2026d), and a state counts as over its own line when its own indicator reaches 0.50.

At the moment of a national call, on today’s state rates, the states give little away. In July 2024, 22 of the 51 were over their own line. At the recession calls of 1980, 1990 and 2001 the count was much the same, between 21 and 24. At the February 2008 call it was only 15.

The difference came after. In every recession since 1980, at least half the states were over their line within three months of the national call, and within a year the count reached at least 42 (Fig. 13). In 2024 it never passed 22, the count it first reached in June 2024, and by August 2026 nine states were over. The July 2003 crossing stalled the same way, at 24.

NBER’s own definition asks for this kind of breadth. A recession, in the committee’s terms, is a decline spread across the economy, judged on its depth, its diffusion and its duration together (NBER, 2026a). The count of states over their line is a rough measure of the diffusion. After the June 2001 call it climbed from 24 states to 48 by February 2002. After the July 2003 crossing it fell from 24 to none by February 2004. In 2008 it went from 15 states at the February call to all 51 by December, the month NBER announced the recession (NBER, 2026b).

After the 2024 call most states did cross their own line at some point, 41 of the 51, but a few at a time and not far. The typical state’s highest reading was 0.70 (Fig. 14). After the 1990 and 2001 calls the typical state climbed to 1.50 and 1.37. After the February 2008 call it reached 3.53. The 2003 crossing again looks like 2024, with a typical high of 0.53.

The 22 states of July 2024 included California, Florida, Illinois, Ohio and Michigan, and left out Texas and New York. A spreading alarm is how a recession shows up across the country. It cannot make the rule faster, but it is a check a reader can run within months of the next call.

The verdict is about the rule as Sahm wrote it in 2019, and a rule changed after the fact can always clear a false alarm. Fig. 15 scores nine other versions the same way as the main test, on today’s rates since 1970 and, where a first-published series exists, as first published since 1995. All were scored after the 2024 call.

The threshold sits close to the edge. At 0.60 or 0.70 the rule catches the same seven recessions with no false alarm on either set of rates. The 2024 reading peaked at 0.57, just under those lines. The price is speed: the typical call comes four months after the peak instead of three. At 0.30 or 0.40 it calls a month or so sooner and fires falsely in 2023 or 2024.

A shorter look-back helps too, because the 2024 rise was slow and built over more than a year. Measured from the low of the past nine months, 2024 stays quiet on today’s rates but not as first published. From the low of the past six months it stays quiet on both, but the calls in 1974, 1980 and 2008 come two to four months later. Using the single month’s rate instead of the three-month average makes the rule faster, with the typical call a month after the peak, but adds false alarms in 1986 and 2023; that noise is why Sahm averaged (Sahm, 2024a). The fresh-crossing count is the only version that catches all eight, at the price of 2003 as a second false alarm.

The version that fits the mechanism best uses only job losers. Applied to BLS’s unemployment rate of job losers, the same rule catches the same seven recessions and stays quiet in 2024, when it peaked at 0.30. It is slower than the rule in three of them, by one to three months, the most in 2008, when it called in May rather than February. Like every version that stays quiet in 2024, it is a hypothesis for the next downturn rather than a record. Even the rule itself was tuned on the years 1970 to 2018 (Sahm, 2019; Sahm, 2024a). Its record on data it had not seen is the seven years since: one recession, called in April 2020, and one false alarm.

The strongest case for setting 2024 aside comes from Sahm herself. A week before the July report she wrote that a recession was not imminent, and that shifts in who was looking for work, from the pandemic and from immigration, were likely making the rule overstate the weakening (Sahm, 2024b). In her words, the rule is “a recession indicator, not a recession forecast”.

As an indicator it is fast: NBER dates recessions after the fact, and it named the 2001 and 2007 peaks eight and twelve months after they had passed (NBER, 2026b). The jobs reports that carried the rule’s calls came out four and seven months sooner (Fig. 16). In 2020 the call beat NBER by a month.

Its defenders can add that this cycle broke more than one rule. In 2022 the first estimates showed the economy shrinking for two quarters running without a recession, which Sahm counted as the first such fall outside a recession since the Second World War (Sahm, 2022b). The inverted yield curve, the forecasters’ favorite signal, is among the tools Sahm (2024a) lists as having failed in these years. A pandemic that emptied the labor force and an immigration surge that refilled it are shocks no rule built on earlier cycles was tested against. After the call Sahm went further: tools built to read weak demand for workers, her rule among them, fall short while the supply of workers swings this hard (Sahm, 2024c).

Neither point changes the count. Sahm herself counts two false positives since 1959 on the rates as they stood at the time, in December 1959 and October 1969, each followed within months by a recession (Sahm, 2024b). Under the window used here a call in October 1969 counts as a hit, because the recession began two months later, and 1959 falls before the years the claim covers. The claim is that the rule has never fired without a recession, and under the count fixed in advance July 2024 is the first time it did, on both sets of numbers.

On the October 2026 report, the rule fires if the three-month average reaches 4.63% (Fig. 17). That is half a point above the lowest average of the past year, 4.13%. From a rate of 4.2%, the October report alone would need a jump to 5.6% to get there, so a new call would more likely build over several months. BLS publishes October’s rate on November 6, 2026 (BLS, 2026f).

The job-loser version is the hypothesis to test on the next call (Fig. 17). It read zero in September 2026, and its 2024 peak was 0.30. A new call carried by job losers, with that version also at 0.50 or more, would look like the recessions before 2024 rather than like 2024 itself.

Two faster signals sit beside it (Fig. 18). New jobless claims in September 2026 ran 10.4% below a year earlier, and every call since 1990 that came in a recession came with claims above their level of a year before. In August 2026, nine states were over their own line. In every recession since 1980, more than half were over within three months of the national call. BLS publishes September’s state rates on October 20, 2026 (BLS, 2026e). Layoffs, the slowest of the signals and often the last to move, ran at 1.0% of employment in August 2026, lower than in any month from 2001 to 2019 and well under the 1.4% of the February 2008 call (BLS, 2026b).

The verdict can still move one way. Under the count fixed in advance, the 2024 call becomes a hit if NBER dates a recession that began by January 2025 and ended no earlier than July 2024. The claim would then hold on the first-published rates, though not on today’s, where 1981 stays missed. Revisions cannot move it: the first-published false alarm is fixed, and today’s rates keep the 1981 miss whatever happens to 2024. BLS’s January revisions can still move July 2024’s reading of exactly 0.50 (BLS, 2026g).