Key takeaways
- Across the eight 10-year minus 3-month inversions between 1966 and 2019, the gap to the next recession ran from 5 months to 39 months.
- On the daily 10-year minus 2-year spread, the five inversions that were followed by a recession led it by 6, 13, 17, 19 and 24 months.
- The New York Fed's model put a 70.9% probability on a US recession in May 2024. It gave February 2020, the month of the last dated peak, only 24.6%.
- The 2022 inversion ran 537 consecutive trading days on the 10-year minus 2-year spread, the longest stretch in a daily record that starts in 1976.
- The NBER announced its six most recent business cycle peaks 4 to 12 months after they happened, so confirmation has always arrived long after the fact.
An inverted curve has led recessions by anywhere from 5 to 24 months
Someone tells you the yield curve has inverted, and that a recession is coming. The fair next question is: coming when?
On the record since 1966, the honest answer is a range rather than a date. Where a recession followed, the first inverted reading came between 5 and 24 months before it started, depending on which spread you read and which episode you look at. Twice, the next recession was still at least 34 months away, which is no use to anyone. And the inversion that began in 2022 — the longest in the daily record — has no recession attached to it yet.
That dispersion is the finding here, and it's why this signal is far easier to describe than to use. An indicator with a 6-month lead and an indicator with a 24-month lead are not the same instrument, even when they print the same number on the same screen.
A par yield curve, and the two spreads that carry the argument
Start with the object itself. Every business day the US Treasury builds and publishes one curve, and its own description of it is short: "The Treasury's official yield curve is a par yield curve derived using a monotone convex method." Par means each point on the line is the coupon a newly issued security of that maturity would need in order to sell at face value. The inputs are bid-side quotes on the most recently auctioned bills, notes and bonds, gathered by the New York Fed.
On 7 August 2026 that published curve read 3.87% at 3 months, 4.19% at 2 years and 4.65% at 10 years. Plot those and the line slopes upward. Lending for longer pays more, which is the ordinary shape.
A term spread is simply one point on that curve minus another. Two of them do nearly all the work in public commentary. The first is the 10-year yield minus the 2-year yield, published by the St. Louis Fed as the series T10Y2Y, daily from 1 June 1976. The second is the 10-year minus the 3-month bill, published as T10Y3M. When either goes below zero, the curve is described as inverted: you're paid less to lend for a decade than for a quarter or for a couple of years.
Which spread should you read? Michael Bauer and Thomas Mertens of the San Francisco Fed tested exactly that. Using monthly data from January 1972 to July 2018, they scored each candidate on how well it predicted whether the economy would be in recession 12 months later. Their result: "The three long-term spreads and the one short-term spread all have very similar predictive accuracy, with AUCs of 0.85 to 0.89." A score of 0.5 there is a coin toss and 1.0 is perfection, so these are genuinely informative measures. The 10-year minus 3-month spread won, but only just.
Their description of the timing is the sentence worth holding on to. All the spreads they tested "typically drop below zero, indicating an inversion of the yield curve, about a year or two before the onset of a recession." A year or two. That's a 12-month band of uncertainty, volunteered by the researchers themselves — and the raw record is wider still.
Every 10-year minus 3-month inversion since 1966, and how long the recession took
The New York Fed publishes the monthly series behind its own recession-probability model: the 10-year bond rate minus the 3-month bill rate on a bond-equivalent basis, running back to January 1959. Pair that with the NBER's dated business cycle peaks and the record is complete and checkable. A recession begins the month after an NBER peak.
Eight episodes of a negative monthly spread appear between 1966 and 2019, counting January 1966 and that autumn's readings as one episode. The chart plots the months elapsed from the first negative reading of each episode to the next NBER peak.
- September 1966, at -0.3395, to the December 1969 peak: 39 months.
- December 1968 to December 1969: 12 months.
- June 1973 to November 1973: 5 months.
- November 1978 to January 1980: 14 months.
- June 1989 to July 1990: 13 months.
- July 2000 to March 2001: 8 months.
- August 2006 to December 2007: 16 months.
- June 2019 to February 2020: 8 months.
Set the 1966 reading aside for now. The remaining leads are 5, 8, 8, 12, 13, 14 and 16 months. The median is 12. The range spans 11 months, and every one of those readings looked the same on the day it printed — a small negative number on a screen.
One recession is missing from that list, and the omission is worth stating rather than hiding. The spread went negative in November 1978 and stayed negative, with short interruptions, into 1981. The January 1980 recession and the July 1981 recession both began inside that one stretch. There was no fresh signal before the second of them, because the signal had never switched off.
The 10-year minus 2-year record says the same thing with wider numbers
The daily 10-year minus 2-year series begins on 1 June 1976, so it covers one cycle fewer. Measured from the first daily close below zero to the next NBER peak, its episodes run like this.
- 18 August 1978, at -0.04, to the January 1980 peak: 17 months.
- 13 December 1988, at -0.01, to July 1990: 19 months.
- 26 May 1998, at -0.02, to March 2001: 34 months.
- 2 February 2000, at -0.03, to March 2001: 13 months.
- 27 December 2005, at -0.01, to December 2007: 24 months.
- 27 August 2019, at -0.04, to February 2020: 6 months.
Leave 1998 out, for reasons covered below, and the leads are 6, 13, 17, 19 and 24 months. The shortest is 6 and the longest is 24. That is a four-fold difference in how long the wait would have been, from a signal that gave no clue at the time about which case it was.
Note also how small the trigger values are. Every one of those six first closes sat between -0.01 and -0.04 — inside the rounding of a single day's move. The signal is defined by a sign change, not by a magnitude, which makes the exact start date more fragile than the tidy table suggests.
A 6-month lead and a 24-month lead are two different portfolios
Here the dispersion stops being a curiosity. Suppose you treat an inversion as a reason to cut equity exposure and hold cash instead. On the August 2019 signal, the wait to the peak was 6 months. On the December 2005 signal it was 24 months, and the curve itself stayed inverted for 238 trading days of that wait. The midpoint of those two outcomes describes neither of them.
What the wait costs depends on what cash yields against what was sold, and on how long the position is held. Holding 24 months of cash in a portfolio is a decision in its own right, not a pause. That arithmetic is its own subject, and what portfolio cash really costs works through it properly.
There's a second gap that the headline never carries. A recession is not the thing in your portfolio. Equity drawdowns and recessions overlap, but they start and end on different dates and they have their own distribution of durations, which how long bear markets have lasted sets out. Getting the recession call right and the market call wrong is an ordinary outcome, not a freak one.
The three signals any rule would have had to survive
The near-misses matter more than the hits, because they're what a rule has to survive.
The monthly spread was negative in January 1966 at -0.0984, and again from September 1966 into 1967. The next NBER peak was December 1969, 39 months later. Nobody positioned on the 1966 signal was rewarded for it on any reasonable horizon.
In 1998 the 10-year minus 2-year spread closed below zero on 26 May, spent 27 trading days negative, bottomed at -0.07 on 25 June, and then re-steepened. It stayed positive for the next 18 months before inverting again in February 2000. The next peak came 34 months after that first 1998 close. Read as one signal it looks early; read as two, the second one worked and the first didn't.
Then there's the episode still on the board. The 10-year minus 2-year spread first closed negative on 1 April 2022 at -0.05, then inverted continuously from 6 July 2022 to 26 August 2024 — 537 straight trading days, the longest run in the daily record. The 10-year minus 3-month spread ran 534 consecutive trading days negative from 25 October 2022 to 12 December 2024 and reached -1.89 on 4 May 2023, the lowest reading in that series. On the 10-year minus 2-year spread the episode bottomed at -1.08 on 3 July 2023, which is well short of the -2.41 recorded in March 1980.
The NBER's chronology page records that its business cycle data were last updated on 14 March 2023, and the most recent dated peak on it remains February 2020. So 52 months after that first inverted close, the longest inversion in the daily record has no dated recession behind it.
The New York Fed's own model makes the miss concrete. It converts the monthly spread into a probability of recession 12 months later, using parameters estimated on data from January 1959 to December 2009. With the spread at -1.71 in May 2023, the model assigned a 70.9% probability to a recession in May 2024. Compare that with what the same model assigned to February 2020, the month of the last dated peak: 24.6%. December 2007, the peak that began the 2007 to 2009 recession, got 39.4%. The model's most confident recent reading has no recession attached to it, and its readings for the two peaks that did arrive were among its calmer ones.
The NBER confirms a recession late, which removes most of what is left
Even a good signal needs a scorer. The NBER's dating committee is deliberately slow: "There is no fixed timing rule because the committee waits long enough to avoid any doubt about the existence of a peak or trough." Its published table of announcement lags for the six most recent peaks reads 4, 12, 8, 9, 6 and 5 months. For troughs it runs longer still — "The shortest elapsed time was four months (February 2020 peak) and the longest was 21 months (March 1991 trough, announcement in December 1992)."
Stack that on the front half. An inversion, then a wait of 5 to 24 months, then a recession that isn't officially confirmed for another 4 to 12 months. The 2007 to 2009 contraction itself lasted 18 months, and the committee did not announce its start until a year after it happened. Anyone using the curve as a trigger is running a rule whose accuracy can only be scored years later.
The strongest objection: the Fed's own economists disagree about which spread to read
The serious counter-argument doesn't say the curve is useless. It says most people are reading the wrong part of it.
Eric Engstrom and Steven Sharpe of the Federal Reserve Board argued in June 2018 that the popular long-term spreads are "statistically dominated by a more economically intuitive alternative, a 'near-term forward spread'" — the implied 3-month rate 6 quarters ahead, minus today's 3-month bill. That spread is close to a direct read on whether the market expects the Fed to cut. In their probit model on quarterly data from 1972 to 2018, the near-term spread carried a sensitivity of 0.35: a one-standard-deviation fall, "about 80 basis points", raised the probability of recession by 35 percentage points. The 10-year minus 2-year spread, in the same regression, scored 0.06 with a p-value of 0.43 — "economically small and not statistically different from zero".
If they're right, the famous spread is a noisy proxy for the thing that actually predicts, and its inversions are informative only when they coincide with expected rate cuts. Bauer and Mertens tested the claim and partly conceded it: "Engstrom and Sharpe found that their short-term spread statistically dominated the 10y–2y spread, and our findings are consistent with this result." They still concluded that the 10-year minus 3-month spread was the most reliable single summary, and that the differences in forecasting accuracy across spreads were small.
Both papers agree on something more awkward for the popular version of the story. Engstrom and Sharpe: "This does not mean that inversions of the near-term spread cause recessions." Bauer and Mertens: "the predictive relationship of the term spread does not tell us much about the fundamental causes of recessions or even the direction of causation." Causation could run either way, and historically may have run both.
What this record cannot tell you
Sample size comes first, and it's brutal. There are 8 inversion episodes on the monthly 10-year minus 3-month series between 1966 and 2019, and 6 completed episodes on the daily 10-year minus 2-year series before 2022. You cannot estimate a distribution of lead times from that. The "range of 5 to 24 months" is a description of what happened, not a confidence interval.
The episodes aren't independent, either. The 1978 to 1981 stretch produced two recessions from one continuous inversion, and the 1998 and 2000 signals sit inside the same monetary cycle. Treating them as separate observations flatters the count.
Everything above is US data. The NBER chronology is a US chronology, the Treasury curve is a US curve, and nothing here establishes that a UK or euro-area curve carries the same information about a UK or euro-area recession.
The dating is retrospective and, in principle, revisable, though the NBER notes that since 1978 "there have not been any changes to previously-announced business cycle turning points". The 2022 episode is unresolved rather than settled: a committee that has not dated a peak is not the same as a committee that has ruled one out.
The probability model has a shelf-life problem of its own. Its parameters were fitted on data from January 1959 to December 2009 and have not been re-estimated since, so anything the years after 2009 taught about this relationship is not inside the coefficients. The older published version of the same model, from Arturo Estrella and Frederic Mishkin's 1996 paper — estimated on data from the first quarter of 1960 to the first quarter of 1995 — maps a spread of 0.76 percentage points to a 10% chance of recession four quarters out, -0.82 to 50%, -1.46 to 70% and -2.40 to 90%. Those thresholds come from a sample that ended in 1995, before three of the recessions listed above.
Finally, all of it is a backtest. Every lead time on this page is measured against a peak date that was assigned afterwards, by a committee looking at revised data. None of it was available on the day the curve inverted.
What would change the conclusion
Three things would, and they're worth watching for on their own terms.
The first is an NBER peak dated inside the 2022 to 2024 window. The committee has not ruled one out. If it eventually dates a peak in that window, the 2022 signal rejoins the record with a lead measured from April 2022, and the observed range widens rather than tightens. That is the same conclusion, only more so.
The second is a re-estimated model. If the New York Fed re-fits its probit on data through the 2020s, the mapping from spread to probability moves, and the 70.9% reading for May 2024 would be reconsidered rather than defended. Until then, the published probabilities are a 1959-to-2009 relationship applied to 2026 data.
The third is a spread that separates the two stories. If the near-term forward spread keeps diverging from the 10-year minus 2-year spread, as Bauer and Mertens found it did in the years when short-term rates sat at the zero lower bound, the disagreement becomes testable rather than theoretical, and one of the two measures earns the headline.
What none of this supports is a date. On 7 August 2026 the 10-year minus 2-year spread stood at 0.46 and the 10-year minus 3-month spread at 0.78. What those two numbers imply about the next 24 months is not a question the 8 episodes above can answer, and the honest version of this indicator has always been a statement about probability, not about timing.