Yesterday at 23:17 UTC this desk published that the 20-year Treasury has yielded more than the 30-year on every session from 14 to 24 September, called it nine sessions, and said in print that we could not read the 2–11 September window because the Treasury’s own table had defeated us on those rows. We said we would come back. We have. The gap first appeared on 4 September, closed again on the 9th, and reopened on 10 September — the day the 30-year bond was auctioned. It has been open every session since. That is eleven consecutive sessions, not nine, and thirteen positive sessions out of the seventeen September has had.
The answer is two dates, and you need both
A question phrased as “when did the inversion start” assumes it started once. It did not. On 1, 2 and 3 September the 20-year and the 30-year par yields were identical to the basis point — 5.27 and 5.27, 5.27 and 5.27, 5.25 and 5.25. On 4 September the 30-year fell one basis point to 5.24 while the 20-year sat still at 5.25, and the gap existed for the first time. It survived the long weekend at one basis point on the 8th. On 9 September it closed back to exactly zero, both legs at 5.28.
Then on 10 September it reopened at two basis points and has not closed since. Eleven sessions: 10, 11, 14, 15, 16, 17, 18, 21, 22, 23 and 24 September.
So the honest sentence has two clauses. The 20s/30s gap first appeared on 4 September. It has been continuously open since 10 September. If you are trading the persistence, the date you want is the 10th. If you are dating the regime change, the date you want is the 4th. Anyone who gives you one number without saying which question it answers has not read the table.
All seventeen September sessions, from the Treasury’s own file
Par yields, 20-year then 30-year, then the spread in basis points. Every figure read verbatim from the Treasury’s daily par yield curve table.
- 1 Sep · 5.27 / 5.27 · 0
- 2 Sep · 5.27 / 5.27 · 0
- 3 Sep · 5.25 / 5.25 · 0
- 4 Sep · 5.25 / 5.24 · +1
- 8 Sep · 5.26 / 5.25 · +1
- 9 Sep · 5.28 / 5.28 · 0
- 10 Sep · 5.39 / 5.37 · +2
- 11 Sep · 5.38 / 5.35 · +3
- 14 Sep · 5.37 / 5.34 · +3
- 15 Sep · 5.40 / 5.36 · +4
- 16 Sep · 5.39 / 5.35 · +4
- 17 Sep · 5.32 / 5.29 · +3
- 18 Sep · 5.38 / 5.34 · +4
- 21 Sep · 5.33 / 5.29 · +4
- 22 Sep · 5.33 / 5.29 · +4
- 23 Sep · 5.45 / 5.40 · +5
- 24 Sep · 5.53 / 5.47 · +6
The last three sessions read +4, +5, +6. The spread is not merely persisting; it is widening, and it is widening monotonically. Today’s row does not exist yet — the Treasury posts the curve after the New York cash close, which is after this note goes out — so we are not going to tell you what the twelfth session did.
The re-read that worked, and the one rule that made it work
Yesterday’s failure was not the Treasury’s fault and it was not a transient. That table carries twenty-six columns, and two of them are headed “20 YR” and two are headed “30 YR”. The leading pair sits beside an extrapolation factor and is empty for every row in the month. The real par yields are the twenty-fifth and twenty-sixth columns. Ask for “the 20-year” and you have a coin flip about which pair you get back, and yesterday we lost it on exactly the rows we needed.
What worked today was refusing to name the column at all. We asked for the entire header row verbatim first, counted, and then asked for the last two columns of each row by position. Same URL, same month, same afternoon — and the rows that returned nine consecutive N/As yesterday returned clean figures today.
That is worth generalising and we are stating it as a rule: on any table with a repeated column name, a read that names the column is inadmissible. Read the header verbatim, count, and address the data by position. We nearly retracted a correct series on the strength of a read that was doing exactly what we asked it to.
It reopened on auction day — and the obvious explanation does not survive 15 September
The Treasury’s tentative auction schedule puts the September 30-year bond reopening on Thursday 10 September, settling the 15th. That is the session the gap reopened, and it reopened hard: the 20-year rose eleven basis points that day and the 30-year rose nine. The tenor with supply to absorb outperformed the tenor without any.
The tempting sentence writes itself. On 11 September this desk published that the 30-year auction stopped through with dealers taking 22 percent — a strong auction, real demand, the long bond richens. Tidy.
It does not hold. The 20-year bond was auctioned on Tuesday 15 September, and on that session the 20-year rose three basis points against the 30-year’s two, widening the spread to four. On 10 September the 20-year underperformed while carrying no supply. On 15 September the 20-year underperformed while carrying all of it. A concession story has to predict opposite signs on those two days and it predicts the same sign, so it is not a concession story. The 10-year auction on 9 September is the third data point and it goes the other way again: the gap closed to zero that session.
What is left is an observation without a mechanism: the 20-year has underperformed the 30-year on both of their own auction days and on nine other sessions besides. We are not going to wrap a causal clause around it. That is precisely the error we published yesterday — “since the BOJ moved” was four words doing work the data could not support — and we are not making it twice in twelve hours.
Marking our own number
Three claims to settle, in order of how wrong they were.
“Three business days” (23 September). Wrong by nine sessions. Already marked yesterday.
“Nine sessions” (24 September). Wrong by two, in the direction of understatement, and wrong because we published a count bounded below by the rows we could read rather than by the rows that exist. Eleven is the number.
“It predates the BOJ” (24 September). This one holds and gets stronger. The Bank of Japan raised its policy rate on 18 September. The continuous run started six sessions earlier; the first appearance was ten sessions earlier. The clause we were least sure of is the one that survived.
The pattern across the three is not that we were careless with arithmetic. It is that every error came from stating a bound as if it were a value. “At least three” became “three”; “at least nine” became “nine”. Both times the reader got a number that was true and an implication that was not.
What this does not tell you
It does not tell you what the 20s/30s gap is worth. Six basis points between two adjacent long tenors is inside the bid-offer on plenty of ways of expressing it, and a spread that has taken three weeks to travel six basis points is not a trade for anyone paying financing.
It does not tell you why. We have three auction sessions and they do not agree on a mechanism. Anyone selling you a term-premium story or an index-extension story or a Japanese-repatriation story is selling you a story.
It does not tell you about today. The 25 September row publishes after this note, and the market reads circulating this morning are not the par curve — one publisher, at two removes, has the 30-year at 5.502 percent and the 10-year at 5.225 percent as this week’s highs, and the Tokyo wire had the 10-year back near 5.15 percent during the London morning. Those are intraday marks from different clocks. We are not subtracting them from each other and neither should you.
And it does not tell you the Fed path. The one probability figure we have seen this morning with an actual feed named behind it — the CME FedWatch tool, reported by one publisher and reaching us through a second — puts an October hike at 67.5 percent, against 55.4 percent a week ago and 11 percent a month ago. That is the first time in three weeks of writing about these numbers that anybody has told us where one came from, and it is still two removes from the source.
Related
- FX events calendar — the auction and settlement dates behind this piece
- Signals
- EA presets
