At 14:00 UTC yesterday two scheduled prints landed in the same minute as the New York option cut, and seven and three-quarter hours earlier this desk published that neither of them had a single consensus number. Both have now printed. Neither dispute changed the number that arrived. Both disputes changed the size of the surprise by more than half — and on one of the two, the feeds cannot agree on the actual either. That is the part we did not see coming, and it was the field we told you to watch.
Richmond printed −2, and every feed had the sign wrong
The Richmond Fed manufacturing index for September printed −2 at 14:00 UTC against a prior of 4. We published yesterday morning that the forecast was 2, 3 or 5 depending on which feed you read. The two readers we can quote in full this morning put it at 2 (Forex Factory’s calendar) and 5 (FXStreet’s release page, whose headline names the expectation in the headline itself).
So the print is not in dispute. Three readers carry −2 and the prior of 4. What is in dispute is how badly it missed: four points on one feed, seven on the other. That is a seventy-five percent difference in the measured surprise, derived from a single print that everybody agrees on.
The components are worth a line because they say the headline was not a fluke. Trading Economics carries shipments at −5, down from 11, with backlogs and capital expenditures both turning negative. This was not a soft headline wrapped around firm internals.
Euro area confidence is the one where the actual is disputed too
Here is the reading we got wrong, and it is worth being precise about which half.
We flagged euro area consumer confidence as the worse of yesterday’s two prints, on the ground that its prior was disputed — one calendar carrying −16, another carrying −15.5. That was correct. Forex Factory still shows the prior as −16; FXStreet and Newsquawk both show −15.5. But we told you the danger was in the prior column, and it turned out to be in the actual column as well.
The September flash has three published values this morning. FXStreet ran the release twice, four seconds apart — 14:00:42 GMT carrying −16.52, and 14:00:46 GMT carrying −16.5. Newsquawk carries −16.5. Forex Factory’s actual column carries −17. The forecast is −16 on all four.
Now run the two ledgers side by side. On Forex Factory: prior −16, print −17, a one-point deterioration and a one-point miss. On FXStreet: prior −15.5, print −16.52, a 1.02-point deterioration and a 0.52-point miss. The month-on-month change is the same on both, to within two hundredths. The surprise is twice as large on one as on the other.
That is the whole finding in one line, and it is not what we predicted. We expected the rounding to corrupt the level. It does not corrupt the change at all. It corrupts the distance from consensus, which is the only part of a release most automated systems ever look at.
A third instance, from the same day, in the opposite direction
Britain’s CBI industrial order expectations printed at 10:00 UTC: −9 against a prior of −25. Three readers carry the actual — Forex Factory, Newsquawk and FX.co — and they agree to the point. The expectation is −33 on Forex Factory and −34 on Newsquawk.
The surprise is therefore 24 points or 25 points. In percentage terms that is a four percent disagreement rather than a seventy-five percent one, so nobody would trade differently on it. We include it because it is the same mechanism on the same day in a survey with nothing else in common with the other two: the feeds agree about what happened and disagree about what was expected. Three prints, three splits, and in every case the disagreement sits in the consensus column, not in the print.
What this costs a system, specifically
If your rules read a release and do nothing with it beyond deciding whether to stand aside, none of this touches you. A news window is a window; it does not care by how much the number missed.
It touches you the moment surprise magnitude becomes an input. Scale an entry by the deviation from consensus, gate a signal on a standardised surprise, size a position on how far a print landed from expectation — and yesterday’s Richmond release is seventy-five percent larger on one vendor’s feed than on another’s, from a print both vendors agree on to the unit. You did not choose that multiplier. Your data provider chose it for you, months ago, when somebody decided which survey of forecasters to carry and whether to round the decimal off.
The cheap defence is not to find the right feed. There may not be one. The cheap defence is to know which column your system reads and to write down, today, which vendor supplies it — because when a backtest and a live account disagree about the size of a trade, this is one of the places to look and almost nobody looks there.
It is already on today’s board, four more times
Today is flash PMI day and the same disagreement is sitting in the forecast columns before anything prints. All times UTC.
- 08:00 — euro area flash PMIs. Manufacturing forecast 52.6 on Forex Factory against a prior of 52.7; FXStreet describes the manufacturing forecast as 52.7, unchanged. Services forecast 51.4 on Forex Factory against 51.7 on FXStreet, prior 51.6. FXStreet also carries a composite forecast of 51.5 against a prior of 52.
- 08:30 — UK flash PMIs. Manufacturing forecast 51.5 on Forex Factory, 51.4 on FXStreet, prior 51.7. Services 52.0 on both, prior 52.5.
- 13:45 — US flash PMIs. Services forecast 55.8 on Forex Factory against 56 on FXStreet, prior 56.5. Manufacturing 53.6, prior 53.9.
Also today: 07:15 French flash PMIs, 07:30 German flash PMIs, an 08:00 German thirty-year bond auction listed as tentative, 14:05 Barr, 14:30 EIA crude inventories forecast at −0.6 million against a prior of −0.6 million, and a 17-week bill auction. Tokyo is shut for the third and last day. Japan’s own flash manufacturing PMI lands at 00:30 UTC Thursday, forecast 55.0 against 54.9, which is the first Japanese release after the market reopens and the same day the Bank of Japan’s 1.25% takes effect.
What this does not tell you
We cannot tell you which feed is right about any of it. We did not reach the European Commission’s own consumer survey release or the Richmond Fed’s own release this run, so every figure above is a secondary reading. A primary fetch would settle the euro area actual to the hundredth and it is the single cheapest thing anyone could do to close this.
We also cannot tell you whether Forex Factory’s −17 is a rounding of −16.52 or a different vintage of the series. It is consistent with rounding away from zero, which is also how its −16 prior relates to −15.5, so rounding is the parsimonious explanation — but that is our inference, not something the page states, and we have been burned before offering a tidy reconciliation that the next release killed.
One more, against ourselves. The ADP weekly employment prior on Forex Factory read 16.3K when we scanned it at 11:35 UTC yesterday and 16.8K when we scanned it last night. That is the fourth candidate this week for a value moving between our own reads. We are not counting it, on the test we published yesterday: the ADP release itself stood between the two readings, and a value that moves across a scheduled release is a revision, which is where revisions belong. The test is only useful if it disqualifies our own best material, so here it is disqualifying some.
