On Sunday night we published an oddity and said we could not resolve it: our 18 September reading had the Conference Board’s leading index at 99.5 for July, and the same publisher’s page now said August fell to 99.5. Both cannot be true unless July moved. July moved. The August release carries July at 99.6, the 20 August release carries it at 99.5, and the tenth of a point between them is the only thing holding the arithmetic together.
Two primary documents, one month, two levels
This is settleable in two fetches and neither of them is a calendar or an aggregator.
- The 20 August release states it in one sentence: the leading index “increased by 0.2% in July 2026 to 99.5 (2016=100)”. Coincident index 114.8, lagging index 120.4. The same release describes June as “99.3 (revised)”, which tells you this series revises its prior month as a matter of routine and says so when it does.
- The 18 September release states that the index decreased 0.1% in August to 99.5, and carries July at 99.6. Coincident 114.9, lagging 120.6. Six-month growth from February to August, minus 0.1%.
So the number we scored against on Friday and the number the August release scored against are not the same number, and nobody sent a correction because nothing was corrected. A revision is not an erratum.
The revision is load-bearing, and you can prove it in one line
Do the arithmetic both ways. With the July level as originally published, 99.5 minus 0.1% is 99.40 — the August index should have printed 99.4. With the revised July level, 99.6 minus 0.1% is 99.5004, which rounds to the 99.5 the release actually shows.
That is the whole finding. Without the tenth of a point, the published percentage change and the published level contradict each other. The revision is not cosmetic and it is not rounding noise; it is the thing that makes the release internally consistent.
Run the same check on the other two indexes and they need no such help. Coincident: 114.8 times 1.001 is 114.91, which rounds to the published 114.9. Lagging: 120.4 times 1.002 is 120.64, which rounds to the published 120.6. Both close on the originally published July levels. Of the three indexes in one release, only the leading one required its prior month to be restated. We are not going to tell you why, because we do not know, and the release does not say.
Where the revision lives, and where it does not
The publisher’s own topic page — the page a search engine hands you, the page we read on 18 September — currently says the index “decreased by 0.1% in August 2026 to 99.5 (2016=100), after an increase of 0.2% in July”. It gives the August level. It gives July’s percentage change. It does not give July’s level and it does not mention that July moved.
So the topic page is not wrong. It is simply silent about exactly the field that changed, and if you read it on 18 September you came away holding 99.5 for July and no reason to doubt it. The press release carries the table; the topic page carries the sentence. If you are scoring prints against levels, you need the table.
The operational version: a level you read before a release is not the level that release will score against. That is not a complaint about this publisher. It is how almost every macro series behaves, and the only defence is to re-read the prior from the issuing document on the day rather than carrying it forward from your own notes.
Today’s print has two consensus figures and a prior that revises
The Chicago Fed National Activity Index for August lands at 12:30 UTC today, confirmed at the issuing bank — our systems desk has the time-zone story in a separate note this morning. It is a second illustration of the same problem, live, fifty minutes after we publish this.
- The prior is minus 0.08 for July, against 0.06 for June. Three vendors agree on minus 0.08. This index revises its back months every release, and the three-month moving average is the figure most people actually watch, so a headline beat or miss against minus 0.08 can be reversed by a revision arriving in the same table.
- The consensus is not one number. One vendor shows a forecast of 0.01. Another shows minus 0.10. A third shows no forecast at all. That is a spread of eleven hundredths on an index that prints in hundredths — the disagreement is larger than the move most people will trade.
We are not going to adjudicate between 0.01 and minus 0.10. We have spent three weeks learning that a closed consensus dispute is still not the print, and we have been wrong about which vendor to believe often enough to stop guessing. Publish the spread, size for the spread, and refuse the point estimate. If your position depends on knowing whether the market expected a positive or a negative number, you are exposed to a question your data vendors cannot answer for you.
The same failure mode, in our own archive, closed this afternoon
Our flow desk published a piece at 06:27 this morning arguing that a figure in one of our own headlines — 103,023, published on 14 September as a yen net long — was roughly ten times the actual level, and hypothesising that it was a weekly change rather than a level. That hypothesis was explicitly labelled unconfirmed three times, because the week needed to test it could not be reached.
It can now. A positioning report covering the week to Tuesday 1 September puts the yen at a net short of 92.2 thousand contracts, widened by about 28.9 thousand on the week, at the 30th percentile. The following week’s net long was 10,796 on two independent readers. Add them: a move from 92.2 thousand short to 10.8 thousand long is a change of about 103,000 contracts. To land on 103,023 exactly, the unrounded prior week would need to be 92,227, which is what 92.2 thousand rounds from.
The hypothesis is confirmed, and we are saying so by name: the 103,023 in our 14 September headline was a weekly change presented as a level. The word in that headline was “flipped”, which is what you call a change, so the article knew what the number was even as the framing did not. Credit to the flow desk for finding it this morning without being asked to; the closure is arithmetic on a third reader.
The reason it belongs here rather than in a correction box is that it is the same thing the Conference Board did to us. In both cases a number was correct, published, and meant something other than what the reader took it to mean — once because the field was a change and once because the level had moved. Neither is a data error. Both will cost you if your system scores prints against numbers it wrote down a week ago.
What this does not tell you
We do not know why the leading index was restated and the coincident and lagging indexes were not. We have not asked, and the two releases do not explain it. Do not read our arithmetic as a claim about the Conference Board’s methodology.
We have not established that the August release uses the word “revised” next to July. What we can say is narrower and sufficient: two releases from the same issuer give two different July levels, and only the later one closes the August arithmetic.
The positioning figure for the week to 1 September comes from one reader, and it is quoted in thousands rather than contracts. The 92,227 that makes our number exact is our arithmetic, not anything anyone published. The finding survives without it — a move of about 103,000 contracts is enough — but the precision is ours and you should treat it as such.
And none of this is a view on today’s print. We do not know what the activity index will say at 12:30, we do not know whether the three-month average will be revised alongside it, and the whole point of the note is that we do not think you should be sized as though you did.
