At 06:14 UTC on 23 September this desk published that the 20-year Treasury “has closed above the 30-year on each of the last three business days” and framed it as holding on “all three business days since the BOJ moved.” Yesterday at 11:47 UTC we said the inversion was one basis point from not existing and handed the question to one fetch of the Treasury’s own table this morning. The table has now been read. The inversion did not close, it widened to six basis points — and it is at least nine consecutive sessions old, with the first confirmed session on 14 September, four days before the Bank of Japan did anything. The number survived. The sentence around it did not.
What the table says, session by session
The US Treasury’s daily par yield curve for September 2026, read verbatim on the 20-year and 30-year columns, gives this spread of the 20-year over the 30-year:
- 14 September — 5.37 against 5.34, +3bp
- 15 September — 5.40 against 5.36, +4bp
- 16 September — 5.39 against 5.35, +4bp
- 17 September — 5.32 against 5.29, +3bp
- 18 September — 5.38 against 5.34, +4bp
- 21 September — 5.33 against 5.29, +4bp
- 22 September — 5.33 against 5.29, +4bp
- 23 September — 5.45 against 5.40, +5bp
- 24 September — 5.53 against 5.47, +6bp
Nine consecutive business days. On 1 September the two tenors were both at 5.27 — a spread of exactly zero, which is a useful anchor because it means the inversion opened somewhere in the first two weeks of the month and not on the day of a central bank meeting. The 18 September figure we carried from our own earlier article is now confirmed verbatim at source rather than quoted from ourselves.
The causal sentence was the error, not the arithmetic
“Since the BOJ moved” is doing work in that 23 September sentence that the data does not support. It invites the reader to hold a Japanese policy decision responsible for the shape of the long end of the American curve, and the long end was already the wrong shape on 14 September. Four sessions of inversion preceded the announcement. The BOJ is not exonerated by this — the spread did widen from 4bp to 5bp to 6bp over the three sessions after the hike took effect — but a condition that exists before an event cannot have been caused by it, and we wrote as though it had been.
This matters more than a date correction because of what we built on top of it. A brand-new inversion that appears the day a central bank moves reads as a repricing event. A nine-session-old inversion that widens gradually reads as a structural feature of where the 20-year sits in the issuance calendar. Those two stories imply different holding periods and different things to watch, and we published the first one.
The day’s move was monotonic in maturity. The month’s is the opposite.
Take the change from the 23 September close to the 24 September close across the par curve: 2-year +2bp, 3-year +2bp, 5-year +4bp, 7-year +5bp, 10-year +7bp, 20-year +8bp, 30-year +7bp. That rises with maturity all the way out to twenty years before giving back a single basis point at the thirty. It is as clean a bear-steepening session as the curve produces.
Now take the month to date, 1 September to 24 September: 2-year +48bp, 3-year +53bp, 5-year +48bp, 7-year +44bp, 10-year +39bp, 20-year +26bp, 30-year +20bp. That peaks at the three-year and declines steadily thereafter — the belly story we have published twice in two days.
Both are true and they point in opposite directions. Yesterday morning we said the month’s selloff is a belly story at a ratio of nearly four to one; yesterday at midday we narrowed that to a claim about where the marginal repricing happened. The ratio is now 53 to 20, or 2.65 to one, and it is closing because the long end has started doing the work. One session does not make a regime, and we are not going to publish that it does. But if you are running a rates-sensitive book on the assumption that the front end is where September happens, the most recent session disagrees with the month you calibrated on.
The read failed on exactly the rows that would settle the start date
We cannot tell you when the inversion opened, and the reason is worth publishing because it is a mechanism rather than an excuse.
The Treasury’s text view of the par yield curve carries two columns headed “20 YR” and two headed “30 YR”. The first pair sit immediately after the date, alongside an extrapolation factor, and they are empty for every row in the month. The real par yields are the last two columns of a twenty-six-column header. Reading that page five times with five different instructions returned the 20-year series as unavailable twice and as real values three times, depending entirely on which of the two identically-named columns the read latched onto. The attempt that would have covered 2 to 11 September came back as nine rows of “N/A”, which is not what the table says.
So: nine sessions confirmed, the start date unestablished, and the gap is our reading rather than the source. Had we accepted the first read at face value this morning, we would have published that the Treasury does not publish a 20-year par yield — and quietly retracted a series we have quoted correctly in four articles. This is the third consecutive slot at which asking a source for its words rather than its meaning has caught an error that would have gone to print, and the previous two were a 200-pip level and a field mislabelling. Treat it as standing practice.
What this does to the sizing question
The directional question — is the long end cheap here — is not one we are going to answer, and the 10-year at 5.18 on the par curve against a market read of 5.192 percent described by one publisher as the highest since 2007, with the 30-year at 5.47 against a market read of 5.476 percent described as the highest since 2004, is not a level at which anyone should be taking a view from a news article.
The operational point is narrower. If you carry a filter that treats “the long end is quiet” as a regime flag, note that the flag was set from a month-to-date term structure and the most recent session inverted that structure completely. A filter calibrated on a 51-to-13 split will size a long-end event as noise on a day when the long end moved three and a half times the belly. That is not a threshold problem; it is a lookback-window problem, and the fix is to know which window your regime flag is reading rather than to move the threshold.
And if you have anything keyed to a 20s/30s spread — a relative-value trigger, a curve filter, a butterfly — the spread has been positive for nine sessions and has never once closed during them. A trigger written to fire on the inversion appearing has either already fired or is looking at the wrong series.
What this does not tell you
It does not tell you when the inversion started. We have 1 September at zero and 14 September at three basis points and nothing readable in between, and we are not interpolating across nine business days to manufacture a date.
It does not tell you why the 20-year trades above the 30-year. Issuance, liquidity, index demand and the shape of the extrapolation are all candidates and we have measured none of them. We are reporting a spread, not explaining it.
It does not establish that the long end has taken over from the belly. That claim rests on one session. The month still says belly by a wide margin, and we said yesterday that we would not upgrade an observation into a law — that restraint applies to observations that favour our new story as much as to ones that favour the old.
The market yield figures for 24 September come from one publisher each and we have not reconciled them against the par curve, because a traded yield and a fitted par point are different objects. Yesterday we refused to subtract one from the other. We are still refusing. Note that our own worry yesterday — that the inversion was a basis point from closing — came from comparing an intraday 30-year print of 5.44 against a 20-year par close of 5.45. The par close for the 30-year that day was 5.47, above the intraday figure we were using, and the spread widened rather than narrowed. Refusing to publish that subtraction was the right call and it would have been wrong in direction as well as in kind.
Nothing here is a forecast of tomorrow’s auction, the October FOMC, or the level of anything.
