Euro-area flash inflation for September printed at 3.8 per cent at 09:00 UTC. This desk published the consensus as 3.7 per cent yesterday morning and flagged that one publisher had it at 3.6, and we logged the tenth of a point between those two figures as a live gap with our own number on one side of it. The print came in a tenth above ours and two tenths above theirs. The thing we were tracking was exactly the size of the smaller of the two misses, and half the size of the larger. Meanwhile the core rate, which we published as 2.5 per cent from 2.4 and then said almost nothing more about, printed at 2.5 per cent. The field we argued over was wrong twice. The field we got right went unmentioned.
Both figures were low, and the one nobody disputed was exact
The numbers, from a European reader of the Eurostat flash: headline 3.8 per cent against 3.2 in August and 2.9 in July. Core, excluding energy, food, alcohol and tobacco, 2.5 per cent against 2.4. All-items excluding energy only, 2.3 against 2.1. Energy 18.8 per cent against 14.3. Services 3.2 against 3.0. Food, alcohol and tobacco 1.4 against 1.1. Non-energy industrial goods 1.1, which is the single component that went down.
So the headline accelerated 0.6 of a point. Our consensus implied 0.5 and the other publisher’s implied 0.4. Ours captured 83.3 per cent of the move and theirs 66.7 per cent, and the distance between the two of them — the quantity this desk carried as an open item — was 0.1. You cannot settle a dispute whose entire span is smaller than the error either side is carrying. We spent a slot on the wrong question, and the way to have noticed is that neither figure was ever corroborated against the other’s component assumptions.
Which field your filter reads decides whether anything happened today
This is the part that belongs in a system rather than in a commentary. A news filter keyed on the headline surprise saw 3.8 against 3.7 or 3.6 and will have scored today as a beat, sized somewhere between one and two tenths. A filter keyed on core saw 2.5 against 2.5 and scored nothing at all. A filter keyed on energy saw a 4.5-point jump in a component that is one of five. All three readings are correct about the field they read, and they disagree about whether a 3.8 per cent inflation print is an event.
The mechanism point is that a surprise threshold is only meaningful relative to the field it is applied to, and almost nobody writes the field into the configuration. If your filter says “euro-area CPI, suppress trading if the surprise exceeds 0.2 points”, today either tripped it or did not depending on a choice you probably made by accepting a default. Go and look at which series your calendar feed is actually putting in the surprise column, because on this release the answer changes the sign of the decision.
And the 0.2-point threshold is itself the problem in miniature. The headline moved 0.6 points month on month while core moved 0.1. A threshold calibrated on core-sized moves will trip on almost every headline print; one calibrated on headline-sized moves will never trip on core. One number cannot do both jobs.
A three-year high bought twenty-one basis points of euro
Having beaten both published consensus figures on the headline and printed the highest euro-area inflation rate since 2023, the euro was at 1.1265 against the dollar and up 0.21 per cent on the day when an ECB Governing Council member spoke a little under an hour later. Olli Rehn said elevated energy prices bring the central bank closer to its adverse scenario for inflation, that the projections are subject to very high and pervasive uncertainty, and that higher long-term rates will moderate growth and dampen the transmission of the energy shock into prices and wages.
Read that last clause carefully, because it is the reaction function and it cuts against the obvious trade. A policymaker describing an inflation overshoot and then pointing at long-term rates as the thing already doing the work is not a policymaker in a hurry. Twenty-one basis points of currency on a three-year high in the target variable is consistent with a market that had already priced the energy pass-through and heard nothing new about the response to it. We are not going to tell you what the ECB does next; we will say that the move tells you the print was not information, and the speech was not either.
One release this afternoon, five consensus values — and a prior that changed sign
Non-farm payrolls at 12:30 UTC, confirmed at the issuer: the Bureau of Labor Statistics’ August release names Friday 2 October, 8:30 a.m. Eastern. The consensus, depending on where you look, is 90,000, 89,000, around 100,000, around 60,000 from one named bank, or the 98,000 two readers carried last week. That is five values and this desk has stopped trying to pick one.
The metadata problem underneath it is worse and is checkable in a single table. One calendar’s own history of this release shows July 2026 at minus 23,000 as printed. The Bureau’s August release states that July was revised from minus 23,000 to plus 21,000 — a 44,000 revision that changed the sign of the month. June appears at 57,000 in the same table as originally printed, and the Bureau describes it as revised from plus 20,000 to plus 31,000. So June has three published values across three vintages: 57,000, 20,000 and 31,000, a 37,000 spread on one month of one series. June and July combined are 55,000 higher than first reported.
If you have a rule that reacts to a contraction in payrolls, it fired in August on a month that did not contract. The figure it acted on no longer exists. This is not a vendor error — every number above is internally consistent and the calendar matches the issuer at each vintage. It is the ordinary behaviour of the series, and the system implication is that any logic conditioning on the sign or the level of a payrolls print needs to know which vintage it is holding. Most backtests silently use the final one.
Forty-seven hours and forty-three minutes
The ADP national employment report for September was released at 12:15 UTC on 30 September. Read at 11:58 UTC today, one calendar’s row for it still has both the actual and the forecast cells blank, with the date correct, the time correct at 12:15, and the previous value correct at 38,000. That is 47 hours and 43 minutes. Every earlier month on the same page is fully populated: August 38,000 against 47,000, July 44,000 against 68,000, June 98,000 against 118,000, May 122,000 against 118,000.
This desk has now watched that cell at 11 hours, 18 hours 10 minutes, 23 hours 32 minutes, 35 hours 15 minutes, 42 hours and now 47 hours 43 minutes. The hours are no longer the measurement. The measurement is that our automated reader of that page concluded, again and without being asked, that the release has not yet occurred. Two days after it did. A trading system reading that page does not see a gap; it sees a future event, and a future event is something you wait for.
One more date on the same theme, because it is coming and it is cheap to prepare for: the United States leaves daylight time on 1 November. Four commitments this desk has published land there, including whether a feed we have been tracking corrects a one-hour offset by itself and whether the Institute for Supply Management’s own “EST” label becomes accidentally correct. If any of your release windows are stored in local time rather than UTC, that is the weekend they move.
What this does not tell you
The Eurostat flash figures here are a secondary reading. Eurostat’s own release page returned a 429 rate-limit error on two attempts from this desk and we could not open it, so every euro-area number above is one European publisher’s transcription of the issuer rather than the issuer. We have no second reader for the component breakdown. The house rule says to say so in the footer when the primary source is unreachable, and we are saying it here as well, because the whole first half of this article is an argument about precision.
We did not obtain a market-implied probability for the ECB’s next move, before or after the print, and so the claim that the print was not information rests on a single currency pair’s intraday change at one timestamp. That is weak evidence and should be read as such. We also have no reading on whether the one-hour offset we have been tracking on a session-preview page appeared a third time today — the page was not found in the time available, which is not the same as the page being correct.
Nothing above is a view on what payrolls prints in half an hour, or on what the euro or the dollar should do about it. The five consensus figures and the sign-flipped prior are an argument about how confidently you should act on the first number you see, not a direction.
Related
- FX event calendar — release windows in UTC
- Signals — how we treat scheduled-event windows
- EA presets — news-filter and window configurations
