We Published 43.7 Per Cent. Seven Minutes Later the Same Feed Read 31.6.

One feed, one meeting, one day, eighteen points

Here is every reading of the October probability we can now account for, all of them attributed to the same futures-implied source, all of them from 14 October or this morning:

• About 22 per cent before the consumer price index printed at 12:30 UTC.
• About 50 per cent at 9:21 Eastern, which is 13:21 UTC, in a reading published by one vendor.
• 43.7 per cent at 13:13 UTC — the figure this desk published, relayed through two intermediaries.
• 49 per cent after the print, in unnamed outlets citing the same tool.
• 33.8 per cent later in the day.
• 31.6 per cent at 7:23 Eastern in the evening, which is 23:23 UTC.
• A prediction-market tracker at 21.5 per cent for a 25-basis-point move, early on 15 October.
• This morning, “climbing toward 40” at 07:44 UTC, and “around 40 per cent or less” at 08:31 UTC from a different research desk.

The readings attributed to the futures tool on 14 October alone span 31.6 to about 50 — a range of 18.4 points on a binary event with one scheduled resolution date. Our 43.7 is a factor of 1.38 above the evening reading of the same tool.

This is not a case of feeds disagreeing with each other, which this desk has documented repeatedly. It is one feed reported at different hours of one day, and the spread is a function of time, not of vendor. Which means the mistake was not in the source. The mistake was treating a probability with a timestamp as a probability without one.

What survives the correction, and what we could not re-test

Last night’s article’s main claim was not the 43.7 at all. It was an arithmetic result taken off the cumulative December buckets published alongside it: 4.3 per cent for no change through December, 55.3 for one increase, 40.4 for two, summing to exactly 100.0. From those, the probability of a December increase conditional on an October increase came to 92.45 per cent, and conditional on an October hold to 92.36 per cent — a difference of 0.086 of a point, which is to say December was priced as statistically indistinguishable from independent of October.

That arithmetic is unaffected by the 43.7 being a snapshot. It is internally consistent and reproducible from the published buckets, and the independence result holds at the precision given whatever the headline October number was doing an hour later.

But we promised to re-run it today, and we cannot. Not one publisher in this morning’s sweep published a bucket set. Every reading available is a headline probability — “toward 40”, “40 or less”, “more than 85 by year-end” — with no distribution behind it. So the independence finding stands untested rather than confirmed, and we are saying so instead of quietly letting it age into received fact. The distribution is the article; the headline is what everybody has; and on most days nobody publishes the distribution at all.

One more gap worth recording. We derived a 95.7 per cent cumulative probability of at least one increase by December. The evening reading gives 89.7 per cent for a December increase, 6.0 points lower, and this morning another publisher says “more than 85 by year-end”. Three numbers, three different objects possibly, three different hours certainly. We are not calling that a contradiction. We are calling it unresolved, which is what it is.

The ten-year is down about twelve basis points and nobody will print a number

The issuer’s constant-maturity series, read this morning, has one more observation than it did last week and still not enough. Its newest value is 13 October at 5.21 per cent. The 12 October slot carries a dot and no value. There is nothing for 14 October and nothing for 15 October. The page’s stated next release date is today.

So for the two trading days in which American inflation surprised to the upside and then yields gave the move back, the most quoted interest rate in the world has no published value at its issuer. What exists instead is description:

• A Tokyo desk note says the yield “briefly rose to the 5.30 area”.
• An Asia-Pacific wrap at 03:22 UTC says yields were 21 basis points below that post-print high in its body text — and 20 basis points in its own page summary. One page, two versions, and no absolute level anywhere on it.
• A European-morning wrap at 07:44 UTC says “around 5.1 per cent”, and separately that the yield was losing nearly 3 per cent on the day, which is a relative figure on a base it does not state.
• A Tokyo close overview yesterday afternoon said “near 5.10 per cent”.

Do the subtraction, because nobody else will: 5.30 less 21 basis points is 5.09, and less 20 basis points is 5.10. Both land on the figure two other publishers characterise without quoting. Against the issuer’s last actual value of 5.21, that is a fall of 11 to 12 basis points.

We are publishing that as a reconstruction and labelling it as one. It is three vendors’ prose plus one subtraction, not a measurement. But a desk that needs to mark a dollar–yen position against the American long end this morning has no better option available, and the operational point is that the arithmetic had to be done by the reader rather than supplied by anybody.

A governor spoke and the only numbers attached to her were the publisher’s

A Federal Reserve governor spoke on 14 October. In the account we read, published at 10:15 UTC, she said the Fed would take time to wait and see how the economy evolves after the last increase before acting again. She named no probability. She named no threshold. Neither the October nor the December meeting appears in the account at all.

That is the ninth consecutive official this desk has checked who has not quantified a position, against exactly one cited probability in the whole file — a roughly-85-per-cent figure from 8 October that was itself attributed to the futures market rather than to the speaker’s own judgement, and which the inflation print has since superseded.

What the account does contain is numbers, and all of them belong to the publisher rather than the governor: a speech-tracker score of 6.2 against a historical average of 5.9, and a proprietary sentiment index at 130.21, down 1.42 points, against a neutral reading of 100. We are not disparaging those constructions; a vendor is entitled to build an index. We are pointing out what happens to a reader who does not notice the difference. The only quantified things in the morning’s coverage of a policymaker who said nothing numeric were manufactured downstream of her, by a party with no vote.

Two more officials speak today, at 14:00 and 20:30 UTC. On the record of the last nine, do not schedule a position around either one in the expectation of a number.

What this changes about how you read a probability

Three rules, and the first one is the one we broke.

A market-implied probability is a quote, not a statistic. Treat it the way you treat a bid: it is good at the moment it was taken and nowhere else. If you are carrying one in a model, carry its timestamp in the same field, and if you cannot, do not carry it.

Second, when a feed does publish a distribution, take the distribution and date it. The conditional structure we derived last night is the kind of thing that survives a day of repricing precisely because it is a relationship between buckets rather than a level. The headline number is the part with the shortest half-life, and it is the only part most coverage reproduces.

Third, when the issuer of a rate has not published it, say that, and keep the vendors’ descriptions and your own arithmetic visibly separate. Twelve basis points is a real move on the American ten-year and it is sitting today in a gap between four publishers’ adjectives.

What this does not tell you

We do not know what the October meeting probability is right now, and nothing above should be read as a claim that 31.6 was righter than 43.7. Both are readings of the same tool at different moments. The 31.6 reached us through a collating article rather than from the tool directly, and we have not read the tool ourselves at any point this week — which is a limitation of our whole treatment of this subject, not just of this paragraph.

Nor do we know that the 5.09 reconstruction is right. If the 5.30 characterisation is loose, or if the 21 and 20 basis points in that one page refer to different reference points rather than being a slip, the subtraction fails and so does the conclusion drawn from it. We have published the inputs so that you can see it fail if it does.

We have no directional view on dollar–yen, on the dollar, or on the long end, and we are not implying one. Five American releases land at 12:30 UTC and we have not tried to guess what they do.

And on the independence result: it is untested today, not vindicated. If the next publisher to print a bucket set shows the two conditionals pulling apart, that finding was an artefact of one afternoon’s repricing, and we will say so here rather than let it stand.

Related

• Event calendar — today’s speakers and the 12:30 block, in UTC.
• Signals — how a dated probability enters a decision, and how it should not.
• Presets — exposure settings for policy-speech windows.

The article being corrected: The Market Doubled October on One Print. December Did Not Notice. See also this morning’s piece on the difference between an issuer’s state and your reader’s view of it.