The Chicago Fed National Activity Index for August printed −0.04 at 12:30 UTC on Monday, exactly where the issuing bank’s own release calendar said it would land. Ten hours later, three feeds carrying that release disagree about what it means. One shows −0.04 against a prior of −0.08 and flags it above prior. One shows −0.04 against a prior of plus 0.08 — the same July, restated, sign flipped — which makes the print twelve hundredths below. And one of the largest calendar sites on the internet still shows no actual at all. Same release, same minute, three states.
What actually printed, and what happened to July
On Sunday night this desk published today’s prior as −0.08 for July, with three readers agreeing on it, and published the consensus as a spread rather than a number because three vendors gave 0.01, −0.10 and nothing at all. The print is −0.04. Neither point estimate got it. The vendor showing −0.10 had the sign right and missed by six hundredths; the vendor showing 0.01 had the sign wrong. That is the fourth consecutive week in which the consensus figures this desk publishes have failed to bracket the print, and we are not going to dress it up.
The larger item is underneath. July, which was published on 24 August as −0.08, is carried in Monday’s release at plus 0.08. That is a revision of sixteen hundredths and a change of sign. The three-month moving average, which is the figure the index’s own documentation tells you to watch, went to plus 0.01 in August from minus 0.01 in July — it crossed zero, upward, in a month when the headline fell.
The revision is not asserted, it is derived
We do not have to take the restatement on trust, because the release gives its own category contributions and they have to add up to the index. Monday’s release describes July as:
- Production — 0.00
- Sales, orders and inventories — plus 0.15
- Employment — −0.01
- Personal consumption and housing — −0.06
Add them: plus 0.08, exactly. The July described by Monday’s release is arithmetically a positive-0.08 month. It is not a negative-0.08 month with a typo in the headline.
Run the same check on August as a control, because a check you only apply when it agrees with you is not a check. August’s contributions are production −0.07, sales and orders 0.00, employment plus 0.01, consumption and housing plus 0.01. Those sum to −0.05 against a published −0.04 — one hundredth out, which is a rounding gap, and we are telling you about it rather than leaving it in the drawer. The control is not perfect. It is close enough to confirm the method and nowhere near close enough to turn −0.08 into plus 0.08.
Where this costs money: a surprise that changes sign
This is the part that matters if anything you run reads a calendar feed rather than a release.
One of the three feeds we read scores the print automatically: actual −0.04, previous −0.08, verdict above prior. On that feed’s own arithmetic the American economy did better in August than in July by four hundredths of an index point. On the revised prior it did worse, by twelve. The number in the actual field is identical in both cases. The entire sign of the surprise is supplied by a field the release changed and the feed did not.
A second feed, at a site most of this industry has open in a tab, had at 23:10 UTC neither the actual nor the revision: its release history still listed 21 September with no value and still carried July at −0.08. Ten hours after a scheduled 12:30 release from a Federal Reserve bank.
The operational rule that falls out of this is short. If a rule in your system compares an actual to a prior, the prior is an input with a release date of its own, and it is the input nobody versions. An actual field is watched by everyone; a previous field is watched by no one, and on this release it moved sixteen hundredths and flipped sign without a line of notice on any page where it could be read.
The second silently restated July in two days — and no, that is not a rule yet
On Monday afternoon this desk published a finding that the Conference Board’s leading index for July had been restated from 99.5 to 99.6 between its August and September releases, with no revision notice on the publisher’s own topic page, and that the tenth of a point was load-bearing for August’s arithmetic. That article went out at 11:42 UTC. The Chicago Fed release landed at 12:30 the same day carrying a second silently restated July.
Two instances in one day, on two unrelated American indicators, both found by arithmetic rather than by a notice. It is tempting to call that a pattern about the American data calendar in September. We are not going to, and it is worth saying why. Monthly indexes revise; that is what they are built to do, and a desk that discovers this in public twice in six hours has discovered its own inattention as much as anything else. What we will say is the narrower thing, which is falsifiable: we have now twice scored a prior against a value the issuer had already changed. Two is an observation. If it happens again this week, it is a process failure and it is ours.
Tuesday’s board, and the twenty minutes you should look at
Tokyo is closed again on 22 September, the second of three consecutive Japanese public holidays, so the American afternoon is the whole day. All times UTC:
- 06:00 — UK public sector net borrowing, consensus 15.4bn against a prior of 1.8bn
- 10:00 — CBI industrial order expectations, consensus −34 against a prior of −25
- 12:15 — ADP weekly employment change, prior 16.3K
- 14:00 — Richmond Fed manufacturing index, consensus 5 against a prior of 4
- 14:00 — euro area consumer confidence, consensus −16 against a prior of −16
- 14:00 — the New York option cut
- 14:05 and 14:20 — two Federal Reserve speakers
- 17:00 — a third
Two of those consensus figures deserve a second look. The Richmond index is forecast at 5 against a prior of 4, which is a one-point move on a diffusion index — a forecast of nothing, dressed as a forecast. Euro area consumer confidence is forecast at exactly its prior. On both, the consensus is carrying no information, and a surprise filter keyed to a percentage deviation from consensus behaves badly when the consensus equals the prior.
Then there is 14:00 itself: two scheduled prints and the option cut in the same minute, followed by two central bankers inside twenty minutes. The UK borrowing figure at 06:00 is the outlier for size — a consensus of 15.4bn against a 1.8bn prior is not a forecast of a small change.
The 22 September option expiry board had not been published when we looked at 23:10 UTC on Monday night. It normally appears in the European morning, and we will not guess at strikes.
What this does not tell you
The August print of −0.04 and the restated July of plus 0.08 come from one publisher and its syndications, read in full, together with a second feed that confirms the −0.04 and the 12:30 UTC release time but not the revision. We have not read the Chicago Fed’s own release document for this month; its data page did not return the figures when we fetched it. Every number in the first two sections above is therefore a secondary reading of a primary release we could not open, and the revision in particular rests on one publisher’s account plus our arithmetic on the category table it quotes.
The originally published July of −0.08 is on firmer ground — three independent readers carried it, including the release coverage of 24 August.
We do not know why July was restated, and neither release we can read says. Revisions to this index are routine by design; a sign change on a month that a desk has already scored is the part worth noticing, not the existence of the revision.
Nothing above says the American economy weakened or strengthened in August. An index at −0.04 with a three-month average at plus 0.01 is, on the index’s own terms, an economy growing at close to its historical trend rate. It is the scoring of it that is broken, not necessarily the economy.
And the consensus figures for Tuesday are vendor figures, quoted as published. This desk has now been wrong about consensus figures often enough to say plainly that they are a description of what a calendar page says, not a prediction.
