The United States Treasury sold $25 billion of 30-year bonds at 18:00 UTC yesterday. Within five minutes one publisher had the result on a page graded “in line” with, in its own words, no directional read for bonds. Three hours and nineteen minutes later a second publisher ran it under “weak demand” while its own body text applied no grade at all. A third called demand “weaker”. A fourth called it the weakest since June. The four numbers underneath are the same four numbers. We are not going to adjudicate between those adjectives, because the useful observation is structural: an auction row in your calendar has a time and a size and no consensus field, so there is nothing to score the result against, and whatever your feed puts in the actual column is somebody’s editorial judgement wearing a data field’s clothes. That is one of four events from last night where the row your system stores and the event that actually happened are different objects.
A consensus field is what makes a release sizeable. An auction row does not have one.
Think about what you actually do with a calendar row. For a consumer price print you have a release time, a consensus, a prior, and after the fact an actual — and from the gap between consensus and actual you get a surprise, which is the only quantity that tells you whether the window deserved a blackout. We have published before that vendor dispersion is a poor proxy for that surprise; the ratio in our own measurements has been about ten to one against. But a poor proxy is still a number.
An auction row has no consensus at all. Not a disputed one, not a wide one — none. Here is the complete set of fields three publishers filled in for yesterday’s auction:
- High yield 5.441 per cent. Present at two of the four publishers we read; absent from the one that reported the bidder distribution in most detail.
- When-issued yield 5.437 per cent. Present at exactly one publisher, which gave no tail.
- Tail 0.4 basis points. Present at exactly one publisher, which gave no when-issued level.
- Bid-to-cover 2.36 at two publishers and 2.401 at a third.
- Indirect 68.1 per cent, direct 20.5 per cent, primary dealers 11.4 per cent. The three sum to 100.0.
Nowhere in that set is there a figure anybody published before 18:00 UTC that the result could be measured against. Our own desk wrote a pre-registration for this auction yesterday morning precisely because no public expectation exists, and our sister desk scores it in the companion piece. That is a workaround, not a field. If you are running a news filter that sizes blackouts by expected move, every Treasury auction in your schedule is a window you cannot size, and the honest configuration is a fixed window chosen for some other reason — settlement mechanics, say — rather than a computed one you have quietly fabricated.
The second half of this is worse. Because there is no consensus, the market’s read of the result arrives entirely as prose, and prose is where the four grades came from. One of those disagreements is inside a single publisher: a headline saying weak demand over a body that characterises nothing and merely notes allocations shifting away from indirect bidders and toward dealers. If your pipeline keys on headlines, you have ingested a grade the article itself declined to give.
The sequence your rows cannot recover, and the one ordering that is establishable refutes the story
Dollar-yen touched the 160 handle last night for the first time in roughly two and a half months, then fell away. One publisher’s index carried the touch in a headline stamped 16:56 UTC and had the pair at 159.75 by 17:59 UTC. The auction was at 18:00 UTC.
So the 160 print preceded the auction by at least sixty-four minutes. One of the publishers above ran the headline “weak demand at 30-year Treasury auction sends 10-year yield above 5.10 per cent”, and its body says only that the ten-year reached a session high above 5.10 per cent — no figure, no timestamp, and no evidence offered on timing. We cannot order the yield high against the auction, because nobody published a clock on it. We can order the currency, because the auction’s time is fixed by the Treasury and the touch was reported before it. Whatever took dollar-yen to 160, it was not the result of an auction that had not happened yet.
This is the field problem again and it is the one we keep finding in a different disguise each week. A price level is stored with no observation time; an event is stored with a time and no outcome; and the join your system performs between them is a guess that looks exactly like a measurement. Note what rescued it here: not a better price source, but the fact that a scheduled event has an externally fixed clock. The scheduled thing is the only reliable timestamp in the whole sequence, which is an argument for anchoring your event log on the calendar and treating every price stamp as the softer field — the opposite of how most of these systems are built.
We should also say plainly what we do not have: no publisher we read printed an exact high. The headlines say the 160 level with no decimals. One Japanese morning note says the pair fell about seventy sen from there, which would put the low near 159.30, but it prints neither the high nor the low as a number and attaches no clock to either. We cannot give you last night’s high to two decimal places, and we looked at six article pages trying.
Fourteen minutes is shorter than most filters’ detection latency
At 13:33 UTC one publisher reported that a remark from the US president to the effect that no deal with Iran is needed had briefly lifted the dollar. At 13:47 UTC the same publisher reported dollar-yen around 159.70 with limited reaction to the same remarks.
Fourteen minutes. Read charitably that is not a contradiction at all: the move happened and then decayed, and both headlines are true of their own moment. Read as data it is a problem, because a keyword filter watching for geopolitical headlines on a named country would have fired on an event whose entire effect was inside the window between a wire hitting and your system acting on it. The sizing question this converts into is not “should I trade this” but “what is the shortest effect my pipeline can actually capture”, and if you have never measured that number you do not know which of your news rules are decorative.
It also sharpens something we have been circling. We have written repeatedly that the windows worth keeping in a filter are the earliest and most disaggregated ones. The complement is a floor: an event class whose effect reliably expires faster than your latency does not belong in the schedule at all, however market-moving it looks in a headline archive. Geopolitical remarks by officials, with no document and no release time, are the obvious candidate class — which is the same hole we published on Thursday morning about spoken central-bank guidance, arriving from the other direction.
There is no field for “this disagrees with everything else”
Two things we cut last night, and the cuts are the point.
First, six article pages at one Japanese publisher returned client errors in a single research pass while that publisher’s news index loaded cleanly and carried nineteen timestamped headlines with price levels in them. Everything we have on the overnight session above is index-level for that reason, and it is labelled as such in every line. The operational lesson is one we filed before and have now leaned on hard: when a publisher’s article bodies fail, its index is the better fallback than a neighbouring article ID, because it carries headline-level figures for twenty items at once. Six for six is a bad enough rate that the index is not a fallback any more, it is the route.
Second, and this is the one with teeth. A Japanese morning note we read carries a characterisation of what the market is pricing for the Federal Reserve that is irreconcilable with arithmetic we published yesterday from a named probability feed — a quarter point in December fully discounted plus roughly a ten per cent chance of a further quarter, which reproduces to 27.5 basis points implied against a printed 27.6. The note, read through a translation layer, implies something several times that. We are not publishing the figure, we are not calling the publisher wrong, and we are not resolving it, because a lone source contradicting a reproducible calculation is a translation artefact until proven otherwise and we cannot prove otherwise tonight.
What the systems reader should take is that there is no field for this. Your ingest has a timestamp, a publisher, a language, and a confidence score it made up. It has nowhere to record “this number is inconsistent with a figure I can derive from another source”, which is the single most useful flag available and the only one that requires no new data. The cheapest version is a consistency check between any ingested aggregate and its own components, run at write time rather than at read time. We found the redundancy in those three probability figures that way yesterday; we found this inconsistency the same way tonight. Same check, opposite outcome, and both worth having.
What this does not tell you
- We do not have an exact high for dollar-yen. Every level in this piece from the overnight session is an index headline, not an article body, and the 160 touch is published without decimals by every source we read. If you need the print, you need a source we did not find.
- We cannot time the ten-year’s session high. The only publisher carrying it gives no figure and no clock. The auction sequencing argument rests on the currency, where we have a reporting time, and on the Treasury’s fixed auction hour — not on the yield.
- A reporting timestamp is not an observation timestamp. The 16:56 UTC headline tells us the touch had happened by then, not when it happened. It is an upper bound on the event time, which is the safe direction, and it is still not a measurement.
- The fourteen-minute decay is one observation at one publisher on one remark. It is an anecdote about latency, not a measured latency floor, and we have not measured our own.
- We did not establish the cause of either leg of the dollar-yen move. We looked, and the sources we could read assert no cause for the move to 160 and none for the fall away from it.
- The four auction grades are four publishers’ words. We are not claiming any of them is wrong — the claim is that the field they sit in is editorial, which is true whichever adjective is apt.
- We could not open the issuing Reserve Bank’s own speaking-engagements page last night, so the extension we wanted to run — checking whether a prepared text will exist for each of today’s scheduled speakers at their own institutions — did not happen. It remains the right check and we did not get to make it.
- Nothing here says what dollar-yen does next. Every argument above is about what your system can and cannot record, which is a configuration question and not a directional one.
Related
- FX events calendar · release times in UTC, Tokyo and London windows
- Signals · how we structure entries around scheduled releases
- Presets · news-window and filter configuration
