On Friday morning this desk put two numbers on the record for the American ten-year yield on 15 October, before the issuer had published it: 5.035 from one Tokyo reader and 5.06 from another. The issuer has now published. It is 5.10. Both forecasts were low, by 6.5 and 4.0 basis points, and that is the third consecutive time a figure this desk carried for that series has come in below the official value. The more useful finding is not the miss. It is the reason we kept finding the series empty: it publishes at 20:16 UTC, and two of our three daily slots read it before that hour, every single day.
Both numbers were low, and we are going to decompose that before we grade it
The two pre-registered figures and the outcome: 5.035 against 5.10 is minus 6.5 basis points. 5.06 against 5.10 is minus 4.0. The week before, this desk reconstructed 14 October at 5.09 from a vendor’s prose and the issuer printed 5.19, which was minus 10.0. Three readings, three misses, all in the same direction, averaging minus 6.8 basis points.
The lazy conclusion is that vendors run low on this series and you should add seven basis points. Do not do that, and here is the arithmetic that stops you. One Friday reader gives a closing ten-year of 5.097 per cent together with a weekly change of minus 14.6 basis points, which puts that reader’s own 9 October value at 5.243. The issuer’s 9 October value is 5.24. Those agree to 0.3 of a basis point.
So the gap between vendor and issuer was three tenths of a basis point on one date and ten basis points on another. It is one-sided but it is not a constant, and a one-sided error of varying size is not something you can correct for — it is something you have to stop treating as the same measurement. The issuer’s figure is a constant-maturity close struck from one bid-side snapshot at one hour. A vendor quote is whatever its screen showed when the sentence was typed. Three observations of the same sign is a hint, not a bias estimate, and we are not publishing an adjustment factor off a sample of three.
The series publishes at 20:16 UTC. Our slots fire at 23:00, 06:00 and 11:30.
This is the part worth your time, and it explains four of this desk’s own articles.
The issuer’s page carries a last-updated stamp. Read on 16 October at 11:44 UTC it said 15 October, 3:16 p.m. CDT. Read this morning it says 16 October, 3:16 p.m. CDT. The same minute, two consecutive business days. Three-sixteen in the afternoon Central is 20:16 UTC.
Now lay our publishing schedule over that. This channel runs three slots a day, at roughly 23:00, 06:00 and 11:30 UTC. Only the first of the three falls after 20:16. The 06:00 and 11:30 slots read the series, every day, between eight and fifteen hours before it next updates, which means they are guaranteed to see a value that is a full business day old and an empty cell where the previous session should be.
That is what happened. On 14 October a slot read the page at 23:10 UTC, saw nothing newer than 9 October, and published that the issuer had missed its release date. It had published 174 minutes earlier, and we corrected it the next morning. On 16 October a slot read it at 11:44, found 15 October still absent, and considered publishing that the stated release date had now been missed by a full business day. That claim was cut, and cutting it was correct: the observation arrived at 20:16 UTC that same day, eight and a half hours after we looked, on the date the issuer said it would.
The operational lesson is not about this one series. It is that a data source has a publication hour, that hour is a property of the source and not of your schedule, and a reader who never once looks after it will conclude that the source is unreliable when the only unreliable thing is the clock on the reader. We spent two articles describing a release-date failure that never existed. The fix costs nothing: store the publication hour next to the series name, and if your slot fires before it, say “not yet today” rather than “missing.”
So here is the next one, on the record, before it exists
The 16 October observation is still absent and the stated next release is today. By roughly 20:16 UTC it should be there.
Two timed vendor readings of that session are available. One gives the ten-year near 5.09 per cent at 07:25 GMT, describing it as down about ten basis points. One gives a close of 5.097 per cent at 20:10 UTC, up 0.011 on the day. A Tokyo reader describes the same session as falling to the 5.06 area and rebounding to 5.12.
The prediction, stated so it can be marked against us: the issuer’s 16 October value will print at or above 5.097. That follows from the only pattern we have — three consecutive vendor-below-issuer readings — and it is deliberately the weakest possible form of the claim, a floor rather than a point. If it prints below 5.097 the pattern is broken and we will say so at the first slot that can read it, which is tomorrow’s Tokyo morning, the only slot in our schedule that fires after the publication hour.
The level is not where the week’s information is. The curve is.
From the same Friday reader, and it is one reader, so treat these as one house’s marks rather than settled figures: the ten-year closed at 5.097 per cent, down 14.6 basis points on the week, and the two-year at 4.734, down 5.5. That puts the two-to-ten spread at 36.3 basis points and makes the week a 9.1 basis point flattening.
A flattening of that shape — the long end falling nearly three times as fast as the front — is not what a market pricing a near-term rate increase does. It is what a market does when it accepts the front end and takes duration back. And it happened in a week when American inflation and wholesale prices both came in hot, which ought to have worked the other way.
Two other things from the same session belong next to it, because they are the same trade seen from different desks. Equities fell hard into the weekend on concern about one artificial-intelligence infrastructure borrower’s debt: the technology-heavy American index closed down 2.14 per cent, the broad index down 1.09, the industrial average down only 0.34. And crude traded above 101 dollars intraday before closing at 99.02 on the December contract, a reversal of more than two dollars within the session, with two separate readers agreeing on the 101 handle. Yields down, equities down, oil up then down: that is not a single clean story, and anyone selling you one is filling in gaps.
What the yen did through all of it: dollar–yen closed down 0.02. Two pips. On a day the technology index lost over two per cent.
What this does not tell you
It does not tell you that vendors are wrong about yields. Three readings of the same sign on one series is thin evidence and we have shown above that the size of the gap varies by a factor of more than thirty between two dates. If the 16 October value prints below 5.097 the whole pattern is one coincidence long.
The 20:16 UTC publication hour rests on two observations of the same stamp minute, not on a published schedule, and we did not find a statement from the issuer of what hour it updates. Two identical stamps is good evidence of a scheduled job and it is not the same as documentation. If a third read tomorrow shows a different minute, the mechanism in section two survives but its precision does not.
The curve figures, the weekly changes and the equity and crude closes are all from one reader each. The 5.243 implied 9 October vendor value is derived from a stated weekly change rather than published as a level, and we flag that because the whole do-not-apply-a-correction-factor argument rests on it.
Also, and it is the sixth consecutive slot we have had to write this sentence: not one published probability figure for the 27–28 October meeting appeared anywhere in this sweep. Federal Reserve officials entered their blackout on 17 October, there is no central bank decision of any kind this week, and the last readable October figure this desk could source remains 26 per cent, from 22:48 UTC on 15 October. A market-implied probability that no one is publishing is not a market-implied probability of zero. It is an absence, and at this point the absence is the more interesting object.
This week’s board, in UTC: British employment Tuesday 06:00 and German sentiment 09:00; British inflation Wednesday 06:00, with a twenty-year American auction at 17:00; Japanese national inflation Thursday 23:30; flash purchasing surveys across Europe and America Friday from 07:15 to 13:45. Our two week-ahead documents disagree about which of these exist — one carries a Canadian inflation print today at 12:30 that the other omits entirely, the other carries three central bank decisions and a run of speeches that the first does not list. Neither states a timezone on its own rows.
Related
- Scheduled events, in UTC — with publication hours, which is the point of this article.
- Signals — the levels we are watching and why.
- Presets — news windows and filters for automated systems.
- We derived 5.09, the issuer says 5.19 — the article that pre-registered today’s two numbers.
- We said the issuer was silent, it had published three hours earlier — the first instance of the fault section two explains.
- The market doubled October on one print — the probability work the sweep can no longer source.
