At 14:00 UTC today two scheduled prints land in the same minute as the New York option cut. Seven and three-quarter hours out, neither print has a single consensus number. The Richmond Fed manufacturing index is forecast at 2, 3 or 5 depending on which feed you read. Euro area consumer confidence is worse than that: its prior is disputed. We published both of these yesterday as carrying no information, and we were reading one feed.
Richmond is 2, 3 or 5, and the prior is the clean part
Four readers, this morning, on a release three and a half hours away:
- Forex Factory: forecast 2, previous 4.
- Trading Economics: consensus 5, its own house forecast 3, previous 4.
- FXStreet’s economic-calendar event page: consensus 5, previous 4, release stamped “Sep 22, 14:00 UTC.”
- Investing.com, writing yesterday: forecast 5, previous 4.
The prior is the settleable half, and it settles cleanly. All four say 4, and Trading Economics adds the provenance: the index printed 4 in August, released 25 August, down from 5 in July. That is the rule this desk has been applying since Canada in the middle of the month — when vendors disagree about a forecast, check the prior at the issuing end first, because the prior is a fact and the forecast never is. Here the prior holds and the disagreement is entirely in the forecast column.
What the disagreement probably is: Trading Economics is the only one of the four that separates the two fields on the page, showing “Consensus 5” and “Forecast 3” as different things. Forex Factory has a single forecast column. If Forex Factory’s 2 is a house model rather than a survey median, it is not disagreeing with anyone — it is answering a different question in a column labelled as though it were the same one. We cannot establish that from the page, so we are offering it as the likeliest reconciliation and not as a finding.
Euro area confidence is the harder case, because the prior moves too
Same minute, same two calendars, and here the rule that saved us on Richmond does not apply, because the thing it tells you to check is itself contested:
- Forex Factory: forecast −16, previous −16.
- FXStreet: consensus −16.5, previous −15.5, release 22 September 14:00 UTC.
- Trading Economics: consensus −16.5, house forecast −17.0, August actual −15.5, July −15.9.
Three readers put August at −15.5, and Trading Economics adds that it was “in line with the preliminary estimate,” which is about as settled as a flash series gets. Forex Factory has −16. That is not a different number so much as the same number with the decimal taken off, and the rounding is the whole problem: this index is published to one decimal place and one of the two calendars a system trader is most likely to be scraping does not carry the decimal.
Follow it through. On Forex Factory, forecast −16 against a prior of −16 is a flat expectation — no change at all. On the other three, a consensus of −16.5 against a prior of −15.5 is a full point of expected deterioration. Those are not two readings of one expectation. They are opposite descriptions of what the market thinks September looked like, produced by rounding twice in the same direction.
We published both of these yesterday, and both claims rested on one feed
Our 22 September Tokyo piece on the Chicago Fed index carried a section on today’s 14:00 cluster, and it said this: Richmond’s forecast of 5 against a prior of 4 is one point on a diffusion index, and euro area consumer confidence’s forecast of −16 against an identical prior of −16 is nothing at all — so a percentage-deviation surprise filter has almost no denominator to work with. We then drew the operational consequence for anyone running one.
The Richmond half of that claim survives on three feeds out of four and dies on the fourth: a forecast of 2 against a prior of 4 is a two-point expected decline, which is a real expectation and not a rounding artefact. The euro area half dies on three feeds out of four: a full point of expected deterioration is not “identical” by any reading.
The instructive part is not that we were wrong. It is that we were reading one calendar, and one calendar is exactly what a scraper reads. The claim was not careless in the ordinary sense — the numbers were copied correctly from a page that is right about the release times and right about the priors more often than not. It was wrong because we treated a single aggregator’s forecast column as the consensus, which is a thing this desk has written a hard rule against and then done anyway inside a week.
There is a second, smaller version of the same fault. At 23:10 UTC last night Forex Factory showed UK public sector net borrowing forecast at 15.4 billion, CBI industrial order expectations at −34, and Richmond at 5. At 06:15 this morning the same page shows 15.2 billion, −33 and 2. Three forecasts changed overnight on one page. If your scan runs the evening before and your filter runs at the release, you are trading a denominator the page has since replaced.
The other direction: the auction one calendar has and the other does not
Aggregator error is not only additive, and today gives a clean example running the other way. Investing.com lists a 2-year Treasury note auction at 13:00 New York time, 17:00 UTC, with a previous clearing yield of 4.315%. Forex Factory lists nothing in that minute except Barkin speaking.
The issuing department settles it. Treasury’s own tentative auction schedule has a 2-year note auctioning Tuesday 22 September 2026, announced 17 September, settling 30 September. The auction is real and one of the two calendars does not have it. The schedule also announces three bills today — the 17-week for auction tomorrow, the 4-week and 8-week for Thursday, all settling 29 September.
We cannot tell you why it is absent, and the benign explanation is available: a calendar with an impact filter may simply not surface auctions at the default setting. That is a configuration question rather than a data error, and it does not change the operational fact, which is that a system trading off one calendar has a 17:00 UTC supply event and a Fed speaker landing in the same minute with only the speaker on its schedule. The auction time itself is single-sourced to Investing.com, because Treasury’s schedule publishes dates and not times.
What to actually do about a feed-dependent denominator
Three changes, in order of how cheaply they pay.
First, stop computing surprise as a percentage deviation on diffusion indices. Richmond and euro area confidence are both bounded indices that cross zero. A percentage deviation on a consensus of 2 is a different animal from the same deviation on a consensus of 5, and on a consensus that could be zero it is undefined. Absolute deviation, scaled by the series’ own recent standard deviation, does not care which feed you read for anything except the actual.
Second, read the prior from a second source and the consensus from at least two. The prior is the cheap one — it settled in one fetch for Richmond and in two for euro area confidence — and it is also the one that determines whether a print is a beat. If two feeds disagree about the prior, as they do today on a European release, your surprise sign is feed-dependent before you have seen the number.
Third, timestamp your consensus snapshot. Not the release time, which everyone stores, but the moment you read the forecast. Three of Forex Factory’s figures moved in seven hours last night. A consensus read at 23:10 and applied at 14:00 the next day is a stale denominator, and nothing in a normal event record tells you that.
And one thing not to do: do not resolve the disagreement by picking the feed that makes the cluster look most tradeable. The correct response to a consensus that is 2, 3 or 5 is a wider error bar on the surprise, not a decision about which vendor to believe.
What this does not tell you
It does not tell you which forecast is the survey median. We have offered a reconciliation for Forex Factory’s 2 — that it may be a house model in a column labelled “forecast” — and we cannot confirm it from the page.
It does not tell you what either index prints. Both land at 14:00 UTC, after this article.
The Forex Factory rounding argument is an inference from three readers carrying a decimal and one not. It is a strong inference and it is still an inference; we have not seen the calendar state its own rounding rule.
The 17:00 UTC auction time rests on one publisher. Treasury confirms the auction exists on that date and does not publish the hour on the schedule we read.
And we have not established that yesterday’s Forex Factory figures were revised rather than differently sourced overnight. What we can say is that the values on one URL are not the values we recorded from it seven hours earlier.
Related
- FX events calendar — the 14:00 UTC minute and the 17:00 supply event
- Signals
- EA presets — news-window and surprise-filter configuration
