A prediction market was quoted this morning at 20% odds of a US recession beginning within the next year, and the objection wrote itself: the 10-year Treasury is over 5.30%, so how can the odds be that low?
The premise is that a high 10-year yield means a recession is near. It is testable.
The short answer. Recessions have begun with the 10-year at 14.94% and at 0.87%. Across the whole daily history the 10-year has ranged from 0.52% to 15.84%, so recession starts cover more than nine-tenths of that range. The level does not separate recessions from anything else. The shape of the curve does that far better, and it is not inverted today.
The two most recent recessions both began with the 10-year lower than it is now.
The level does not separate them
Across all 680 eligible months a recession began within twelve months 14.1% of the time. That is the number any conditional figure has to beat.
| 10-year | months | episodes | recession within 12m | distinct recessions | when this band existed |
|---|---|---|---|---|---|
| under 3% | 131 | 5 | 9.2% | 1 | 2010-2022 |
| 3-4% | 63 | 12 | 0.0% | 0 | 1962-2024 |
| 4-5% | 129 | 16 | 8.5% | 2 | 1962-2025 |
| 5-6% | 79 | 15 | 12.7% | 2 | 1966-2007 |
| 6-8% | 154 | 14 | 18.8% | 4 | 1968-2000 |
| 8%+ | 124 | 8 | 27.4% | 3 | 1975-1991 |
Read the fourth column alone and the level looks informative: the rate climbs from 8.5% to 27.4% as yields rise. That reading is the reason this study exists, and the last column is why it does not hold.
The bands are eras. Every month in the 8%+ band falls between 1975 and 1991, a stretch that contained three recessions in sixteen years for reasons that had to do with inflation and the Volcker response rather than with the long end. Every month in the under-3% band falls between 2010 and 2022. The 5-6% band had not occurred since 2007 until this year.
The arithmetic closes the loop. Five of the eight recessions fall between 1970 and 1991, one every 4.2 years. The three since then are one every 11.7 years. That is a 2.8x difference in how often recessions arrived, against a 3.0x difference between the extreme bands (27.4% versus 9.2%). The band gradient is the era frequency restated in yields.
The two middle bands do run the full sixty years, and they are the reason this cuts the way it does: an era-neutral band at 8.5% cannot be read against an era-confined one at 27.4% and called a yield effect.
Two of the six bands carry no usable information whatever: the 3-4% rate of 0.0% rests on zero recessions, and the entire 9.2% in the under-3% band is the 2020 pandemic.
As for the band containing today's 5.3%: at 12.7% against a 14.1% base rate it is indistinguishable from an average month. The gap is 1.4 percentage points built on two recessions, about one month's difference in 79. It is not evidence that today's level is safer, and it is not evidence that it is riskier. It is nothing.
The curve separates them, on the same eight events
Using the 10-year minus the 2-year, which FRED publishes daily from June 1976, there have been six recessions. All six were preceded by at least one day of inversion within the prior twelve months, as the episode table above shows.
That count depends on using daily data. On monthly averages the answer is five of six, because the inversion before 2020 lasted three days - 27, 28 and 29 August 2019, reaching minus four basis points - and a three-day dip disappears inside a monthly average. Daily is the honest basis, and the one most readers will have in mind.
It is also the weakest possible entry in a perfect record. Three days at four basis points did not anticipate a pandemic, and nothing about 2020 should be credited to the curve. A record of six from six that leans on that episode is a lesson in how such records get built.
Inversion is wrong more often than it is right. Counting an episode as a run of months whose average spread was negative, with a new episode starting after any gap, there have been twelve since 1976. Seven were followed by a recession beginning within a year, pointing at five distinct recessions. Five were followed by nothing, including the longest of them: twenty-six months of inversion from July 2022 to August 2024.
Five distinct recessions here against six in the daily count is not a contradiction. The 2019 dip lasted three days, so it registers in a daily test and is far too brief to form a monthly episode. It is the only recession that separates the two counts.
At the month level the same thing looks like this:
| Curve | months | recession began within 12 months |
|---|---|---|
| Inverted | 88 | 44.3% |
| Not inverted | 446 | 7.4% |
Six times the rate, and far wider separation than any level band produces.
And the curve passes the test the level failed. Those twelve inversion episodes run from 1978 to 2024 and fall in five different decades, so unlike the 8%+ yield band they are not one era wearing a disguise. That is the strongest thing this study can say for the curve, and it is worth saying because the same sceptical test had to be applied to both variables.
It still rests on six recessions and five inversions that mattered. It is better than the level. It is not precise.
Where the curve is now: the 10-year minus the 2-year is positive, at about +48 basis points. Not inverted. That one reading is a live quote taken at 15:18 ET on 5 October - 10-year 5.315%, 2-year 4.833% - rather than part of the FRED history above it.
So is 20% too low?
On a curve that is not inverted, a recession began within twelve months 7.4% of the time. A market at 20% is about 2.7 times that.
That comparison needs a caveat, because it is the least rigorous number here. A traded market price is not a historical frequency: it carries a risk premium, it resolves on whatever criteria that specific contract specifies, and that definition may not match a twelve-month NBER-start window. The honest version is that 20% is well above what the better of the two variables has historically implied, not that it is precisely 2.7 times too high.
Either way the direction of the original complaint is backwards. Reasoning from "5.3% is high, therefore recession" uses the variable with the weaker record and arrives at a number the stronger variable does not support.
What this does not say
It does not say a recession will not happen. 7.4% is not zero, and 2020 arrived with three days of inversion behind it and nothing else.
It does not say yields cause or prevent recessions. Everything above is association.
And the sample is small in a way the percentages disguise. Eight recession starts across sixty-four years is eight events, however many months are stacked on top of them. The 10-year also sits inside a band for years at a time, so a band's month count is a handful of multi-year episodes rather than that many separate observations, which is why the episode count sits in the table. Any figure here would move materially if one episode were added or removed.
Educational commentary, not investment advice.