Twelve hours ago this desk pre-registered a test on the $58 billion three-year Treasury auction. It named two fields and the order to read them in: the stop-through first, the indirect share second, with the threshold set at 65.9 per cent — the average the October auction had been measured against. We also said the test was losable because we had no when-issued level. Both halves of that were right, and one of them was right for a reason we had not thought of. The auction tailed rather than stopping through, so the first leg is falsified. Indirect bidders took 55.5 per cent, light for a second consecutive three-year, so the second leg resolves yes. But by the time the auction ran, the average we had written into the threshold was no longer 65.9 per cent. It was 62.7. The yardstick had moved 3.2 points toward the thing we were trying to measure, because October’s soft auction is now inside the window that computes the average.
The pre-registration, scored: one leg lost, one leg won, and the subtraction reproduces
Here is the auction, with October’s three-year beside it. The baseline is the previous auction of this tenor, Tuesday 6 October, not the previous auction of any security — we checked the weekday of every date in code, and 11 August, 8 September and 6 October were all Tuesdays while Monday 9 November was not, which is itself the week’s first scheduling fact and we come back to it.
- High yield: 4.627 per cent, against a when-issued level of 4.622 at the time of the auction. Our subtraction: 4.627 − 4.622 gives a 0.5 basis point tail, and both publishers who covered it print “0.5 bp” independently, so the arithmetic reproduces from its own inputs. October stopped 0.2 basis points through. The swing between the two auctions is 0.7 of a basis point, and it crosses zero.
- Indirect: 55.5 per cent against a stated average of 62.7 — 7.2 points light. October was 57.6 against 65.9, 8.3 points light.
- Direct: 35.0 per cent against 25.2 — 9.8 points heavy. October was 31.7 against 21.5, 10.2 heavy. One publisher calls the 35.0 the highest on record.
- Primary dealers: 9.5 per cent against 12.1 — 2.6 points light. October was 10.7 against 12.6.
- Bid-to-cover: 2.53 against 2.64, down 0.11. October was 2.62 against 2.65, down 0.03. One publisher calls 2.53 the lowest since the August 2025 refunding.
The three shares sum to 100.0 exactly and the three averages sum to 100.0 exactly. That check cost nothing and it is the reason we trust the figures: last month the same check settled a dispute between two publishers over whether the direct average was 21.4 or 21.5, because only 21.5 made the column add up.
So the distribution story is now two for two. Indirect light, direct record-heavy, dealers light, cover softer — twice running, in the same tenor, five weeks apart. That is what we said would be “worth a filter”, and we hold to it, with one correction below about what the filter should key on. The yield half of the pre-registration is simply wrong: we expected the auction to clear through again and it tailed, on a print 30.5 basis points below October’s after steps of plus 18.3 and plus 45.8 in the two auctions before that. A security can richen 30 basis points in five weeks and still be awarded grudgingly.
The yardstick moved toward the thing we were measuring
This is the part worth taking away, because it generalises past Treasuries. Between the October three-year and Monday’s, the stated comparison averages changed:
- Indirect average: 65.9 to 62.7 per cent, down 3.2 points
- Direct average: 21.5 to 25.2 per cent, up 3.7 points
- Dealer average: 12.6 to 12.1 per cent, down 0.5
- Cover average: 2.65 to 2.64, down 0.01
Those are rolling windows — last month one publisher stated its averaging period as the previous six auctions of the same security, and the other two stated no period at all. A rolling window re-bases itself every time it is extended. October’s auction, the soft one we wrote about, is now one of the observations computing the average that Monday’s auction is judged against. So the benchmark walked toward the anomaly.
Run the number both ways. Against Monday’s own stated average, indirect was 7.2 points light. Against the average our pre-registration actually named — 65.9, five weeks old — it was 10.4 points light. Same auction, same print, 3.2 points of difference, and all of it comes from which vintage of a moving average you hold. A threshold written in October and executed in November was not measuring what it was written to measure.
The mechanical lesson for anyone who stores thresholds: a condition expressed against a rolling statistic has an expiry date, and the statistic does not tell you when. If your rule says “flag when indirect demand is more than five points below average”, that rule gets quietly looser every time a weak auction prints, because the weak auction joins the average. Two or three soft auctions in a row and the rule stops firing at exactly the point the thing it watches for has become the norm. An aggregate is a lagging indicator of its own components, and the components are what you wanted. Store the level — 55.5, 57.6 — with its date, and compute your own comparison against a window you chose and can state. Do not store the delta, and do not store anyone’s adjective.
We would not have found this if the pre-registration had not named a number. A test that says “watch whether demand is soft again” cannot be lost and cannot teach you anything. Naming 65.9 is what made the drift visible.
Two publishers, two baselines, and both of their verdicts are true
Monday’s auction was covered by two readers we could open, and they do not describe the same event. One grades it C minus and says it was “a little below average”, giving every share against an unlabelled average. The other headlines it as drawing weak demand, calls it lackluster, and gives no averages at all — instead it ranks: cover lowest since the August 2025 refunding, indirect lowest in fifteen months, direct the highest ever. For the dealer field it gives no percentage whatsoever, only “0.8 standard deviations below the one-year average”.
Last month we published that three publishers gave one three-year auction three different adjectives — in line, C plus, and soft — and that the resolution was that each was summarising a different column. This month the resolution is one level deeper, and we are correcting our own framing. It is not only that they summarise different columns. They summarise against different baselines. “A little below average” and “the highest ever” are both true statements about a 35.0 per cent direct share, because the first compares it to a six-auction mean and the second to the full history. Neither publisher is wrong and neither is telling you which comparison it made.
Note also what the only named window in either page is attached to. The single field where a period is stated — “one-year average” — is the single field where no percentage is given. The number you can compare has no window; the window you can trust has no number. We have now watched six consecutive weeks produce a different missing field each time, and the row has looked complete every time.
One of the two pages discloses that it was generated with AI support and reviewed by an editor. We read its body rather than its table, which is this desk’s standing rule for that publisher, whose numbers have been reliable and whose labels have not. We are reporting its three rank claims as that publisher’s characterisations and we have not second-sourced any of them.
Five releases share 23:50, two of its fields have opposite signs, and the issuer states no time zone
At 23:50 UTC — 08:50 in Tokyo — Japan publishes five things at once. The Bank of Japan’s Summary of Opinions from the 29–30 October meeting. September’s balance of payments. Two October bank-lending series. And, inside the balance of payments, the September trade balance. Twenty minutes earlier, at 23:30, September household spending lands alongside Australian consumer confidence. Seven scheduled items in twenty minutes, five of them in one minute, and the headline item of the five is a policy document with no numeric field in it at all.
Now look at what the two Japanese calendars we read actually carry for that minute. One lists a single row: current account, consensus 3,354.4 billion yen against a 4,062.0 billion prior. The other lists two rows under one release header: current account at 3,300 billion yen not seasonally adjusted and 2,763.9 billion seasonally adjusted, and trade balance at a 225.8 billion yen deficit. It gives no prior for either.
So a filter keyed to “Japan Current Account” expects a surplus of about 3.3 trillion yen. A filter keyed to “Japan Trade Balance” expects a deficit of 225.8 billion. Same release, same minute, opposite signs, 14.9 times apart in magnitude — and only one of the two calendars tells you the second field is there. The two calendars also disagree with each other about the field they both carry, by 54.4 billion yen or 1.65 per cent, before you even get to the seasonal-adjustment question, where one of them gives two values 536.1 billion apart under one label. Four of the day’s other rows do the same thing more mildly: the Economy Watchers survey is 47.0 and 47.4 on one calendar and 47.1 and 47.1 on the other, and China’s trade balance is time-unstated on one, time-to-be-confirmed on the second and 04:00 UTC on a third.
The same defect has a cleaner form in tonight’s US print. Four independent vendors carry the October CPI consensus as 0.3 per cent month on month, headline, with core at 0.2 — and the release hour is settled, 13:30 UTC, confirmed by five readers including one that writes it as 08:30 Eastern Standard Time, which is correct, because US clocks changed on 1 November. But the named house forecasts underneath that 0.3 run 0.26, 0.28, 0.29, 0.20, 0.30 and 0.30. Five of the six round to 0.3. The spread is ten basis points, the mean is 0.272, and the figure one bank identifies as the survey consensus is 0.29. A print of 0.26 would be three basis points from the real consensus and would read, against the stored one-decimal 0.3, as a miss — or as no miss at all, depending on which decimal your vendor rounds to. The rounding is wider than the disagreement it conceals.
And one more, from the issuer this time, because it is the cleanest example we have of an authority omitting the field everybody stores. The Bank of Japan publishes its release times as a standing rule on its own pages: Summary of Opinions at 8:50, minutes at 8:50, the Outlook Report’s full text at 14:00 the next business day, the governor’s press-conference record the next business day with the time not set. That is a better process statement than most central banks give. It carries no time zone, anywhere, on either page. An issuer is authoritative about its process and merely assertive about its clock, and the only clocks worth trusting from an issuer are the ones where the issuer publishes its own zone rule. This one does not, and the pages add that dates and times may change.
While we were there we found a scheduling fact worth more than the Summary itself. The minutes of the 29–30 October meeting publish on Wednesday 23 December. The December meeting is 17–18 December. The minutes of the October meeting arrive five days after the December meeting has already decided. Any rule of the form “wait for the minutes before adjusting the Japan stance” is not cautious; it is structurally one meeting late, every time, by design. The Summary of Opinions exists precisely to fill that gap, which is why it is worth reading tonight even though it cannot be scored.
Last: Monday’s auction was on a Monday because this is a holiday week. The US bond market is closed on Wednesday 11 November for Veterans Day while equities trade, so the refunding was pulled forward — three-year Monday, ten-year Tuesday 18:00, thirty-year Thursday 18:00. If your calendar stores auction days as a weekday pattern rather than as dates, it has this week wrong, and the three-year it has wrong in the past tense.
What this does not tell you
The when-issued level of 4.622 per cent comes from one publisher. We trust the subtraction because a second publisher independently states a 0.5 basis point tail, which is what 4.627 minus 4.622 gives — but that is consistency between a number and a derived figure, not two readings of the number itself. If the when-issued is wrong, the tail is still 0.5 and our arithmetic is still coherent, and both could be wrong together.
Every bidder share, and all four comparison averages for both auctions, rest on one publisher per auction. We have not read the issuer’s own auction results, and the auction-results page at the Treasury is not something this desk has opened for either date. The averages for Monday carry no stated window at all; our claim that they are six-auction rolling averages is inferred from last month’s auction, where a different publisher stated that period. If Monday’s publisher uses a different window, the 3.2-point drift we have built a section on is a comparison between two incompatible statistics rather than one statistic moving — which would be a worse problem, not a smaller one.
We cannot tell you what is in the Summary of Opinions. At the hour of writing it has not been published. We also cannot tell you whether anybody on the board argued for a hike at that meeting, which is the only question the document can answer, and we have deliberately not pre-registered a guess, because a document with no numeric field gives nothing to score a guess against.
The 23:50 field problem is established on two calendars and no issuer. We did not reach Japan’s Ministry of Finance or the Cabinet Office to confirm which fields the balance-of-payments release actually contains or what the trade-balance sub-item’s prior was. Two aggregators disagreeing is good evidence that a calendar consumer has a problem; it is not evidence about the release.
And one open conflict we are leaving open rather than resolving. Our own week-ahead places a Dallas Fed speaker at 14:35 UTC today; one of this morning’s calendars places him at 15:50. Seventy-five minutes apart, two sources, no issuer read, no reaction note to settle it. If you run a blackout around that name, run it wide today.
Related
- FX events calendar — the 23:30 and 23:50 UTC Japan cluster and the 13:30 US CPI
- Signals — thresholds with expiry dates, and why a delta is the wrong thing to store
- EA presets — news-window configuration for one-minute multi-release clusters
