At 11:30 UTC yesterday this desk published, as a scheduling argument, that September pending home sales would land at 14:00 UTC in the same minute as the New York option cut, and that its consensus could not be stated more precisely than a 70-basis-point band: 0.0 per cent at one publisher and minus 0.7 at another. The release has happened. The National Association of Realtors puts the index at 66.0, down 7.3 per cent on the month and 11.3 per cent on the year. The nearer edge of our band was 660 basis points away from the print. The argument we published was about a quantity roughly a tenth the size of the thing it was supposed to describe.
What the issuer actually published
The primary release is short and clear. The Pending Home Sales Index for September reads 66.0, where 100 is the average level of contract activity in 2001. It is down 7.3 per cent from August and 11.3 per cent from a year earlier. Every region fell: the Northeast 4.2 per cent, the Midwest 6.5 per cent, the South 8.2 per cent and the West 9.2 per cent, with the West down 15.0 per cent on the year. The release attributes the weakness to fast-rising mortgage rates and quantifies no rate.
The hour is confirmed where it matters: the issuer’s own page metadata carries 9:00 a.m. at minus five hours, which is 14:00 UTC. Three secondary readers gave us that hour yesterday and the issuer agrees with all three. The time was the one field on this event that everybody had right.
Two secondary readers describe the number as a record low. The issuer’s release does not say so, and we did not read the historical series, so we are not saying it either.
The consensus was never one number, and one publisher quietly held two of them
Here is the field as it existed across three readings, all from houses this desk uses weekly:
- One Tokyo reader, writing at 08:25 UTC on 20 October, before the release: consensus 0.0 per cent, prior plus 0.3.
- One European reader, writing at 06:18 the same morning: consensus minus 0.7 per cent.
- The same Tokyo reader, writing at 19:08 after the release: consensus minus 0.5 per cent, prior plus 0.3.
The first and third are the same publisher, on the same release, 50 basis points apart, separated by the event itself. We are not alleging that anyone rewrote anything; a desk’s pre-release note and its post-release wrap are produced by different people from different survey snapshots, and consensus figures genuinely move in the final hours. That is precisely the problem. The consensus you traded against is not a property of the release. It is a property of the page you happened to read, and of the hour you happened to read it.
Which means the honest statement of the field is not “0.0 per cent” or “minus 0.5”. It is: somewhere between plus zero and minus 0.7, depending on vendor and timestamp. Our article yesterday said that and treated it as a finding. It was a finding. It was also, measured against the print, irrelevant.
Seventy basis points of disagreement, 680 of surprise — and you cannot size a window with that ratio
Do the arithmetic that matters for a system rather than for a story. The spread between the two pre-release consensus figures was 70 basis points. The distance from the print to the middle reading of the three was 680 basis points. To the most pessimistic of the three it was 660, and to the most optimistic 730.
The ratio is 9.7 at the middle and 9.4 at the far edge. Vendor dispersion under-predicted the realised surprise by a factor of about ten.
That is the transferable lesson and it cuts against a tempting design. If you size a news blackout, or a position reduction, by how far apart the forecasters are — a wide consensus means an uncertain event, a tight one means a safe event — this release says the relationship does not hold. Seventy basis points of disagreement is what a market looks like when every forecaster is anchored on the same recent months and none of them has priced a regime change. Agreement among forecasters is evidence that they share a model, not evidence that the model is right. A filter keyed to consensus dispersion would have treated yesterday’s 14:00 UTC release as a minor event, and the pair fell to 157.82 within the half hour while the largest option notional on the board expired 40 pips out of reach.
The companion release says the same thing from the other side. Housing starts printed 1.221 million against a 1.305 million consensus — 84,000 units, or 6.4 per cent, below. Building permits printed 1.40 million against 1.385 million, essentially in line. Two releases, the same minute class, the same forecasters: one of them nearly exact and the other out by 6.4 per cent, with the contract-signing series out by 680 basis points. A single number describing “how uncertain the housing data is” does not exist.
The second field: the Bank of Japan’s December meeting does not exist at the issuer
For three runs this desk has carried a December Bank of Japan meeting date sourced to one aggregator, which says 17 December, never second-sourced and never published in an article. It has now been taken to the primary source, and the primary source settles it in the only way that was not on the list of expected answers.
The Bank’s own calendar lists exactly one upcoming Monetary Policy Meeting: 29 and 30 October, with the Governor’s press conference on the 30th at 15:30 Tokyo time. Its release calendar runs forward to 13 November and stops. There is no December meeting on it. There is no January meeting on it either.
So the field is not disputed and it is not wrong. It is absent. The institution that will hold the meeting has not published when, and the aggregator that gives you 17 December is giving you something with no issuer behind it. That is worse than a conflict, because a conflict is visible in your data and an aggregator’s confident single value is not. If your calendar carries 17 December as a dated row, it is carrying a forecast formatted as a fact.
This closes the item honestly rather than well. We wanted a date. What we have is a demonstration that the date does not yet exist where dates come from, and a note that the next thing the Bank will publish about its own schedule is more likely to arrive after the October decision than before it.
The third field: a decision time whose only source was retracted
Yesterday we dropped a 175-basis-point Turkish rate-cut figure that had survived six consecutive runs in our own notes while sitting outside every reader’s range. The entry that carried that figure also carried the decision time, 11:00 UTC on Thursday. When the size went, the time lost its source, and none of the three readers we found for the decision gives an hour at all.
It remains unconfirmed this morning. We are repeating that here, in a piece about calendar fields, because the failure mode is the same one in a third register: a row in a schedule has several fields, they are usually sourced together, and when one of them is retracted the others do not automatically survive. A system that validated the number and inherited the timestamp is carrying an unsourced timestamp.
A prior that moved, and why we will not tell you which way
Our pre-release note carried August housing starts at 1.275 million. This morning’s wraps put August at 1.288 million and describe it as revised. The two renderings of that publisher’s text that we read give the revision in opposite directions — one has 1.288 revised from 1.275, the other has 1.288 revised to 1.275 — and both are machine readings of the same Japanese-language source, not two independent publishers.
So the only defensible statement is that the August prior moved by 13,000 units between our note and the print, and that we cannot tell you which way from the readers we have. This desk has a standing rule about trusting a translation layer on a lone source, and the rule applies to us when the lone source is one we like. The useful half of the observation survives anyway: the prior you measure a surprise against is itself a revisable number, and a surprise computed against a stale prior is wrong by the size of the revision before the print is even considered.
What this does not tell you
It does not tell you what the consensus for pending home sales really was. We have three values from two publishers and no survey provider’s own figure, because we did not reach one. The 9.7 ratio in the third section therefore rests on a dispersion measured across vendors rather than across forecasters, and those are not the same thing.
It does not establish that pending home sales caused the dollar’s fall. Crude collapsed on reports of a United States–Iran agreement in the same window, American long yields fell with it, and a denial reversed much of the move. Three candidate drivers inside one hour and no way to separate them from the readings available.
One release does not make a rule about consensus dispersion. It is one observation, on one series, in one month, and the housing-starts figure in the same release block behaved completely differently. Treat the ratio as a reason to stop using dispersion as a risk measure, not as a measurement of anything.
And on the Bank of Japan: the absence of a December date on the issuer’s calendar is the absence of a published date, not evidence that no meeting will occur or that the aggregator’s 17 December is wrong. It may well turn out to be right. It is simply not sourced today.
Related
- FX events calendar — every release in this note, in UTC, with the hour stated.
- Signals — why we size the window rather than the direction.
- EA presets — news-window and filter configuration.
- The $1.1bn at 159.00 Sits 40 Pips Above Today’s High — where the 70-basis-point band was published.
- We Carried a 175 Basis Point Cut for Six Runs — the retraction that left a decision time unsourced.
