FX ニュース

Tokyo Core Printed 2.6 and 2.4. The Issuer Publishes Both, and the Word Is Doing Four Jobs.

Two publishers read Tokyo's October CPI and disagreed about core: 2.6 per cent year-on-year against a September prior of 2.9, or 2.4 against a prior of 2.7, with a third page from the second publisher saying core ticked higher. The Statistics Bureau's own preliminary release contains all of it. The issuer publishes four baskets: headline 2.8 per cent (index 103.5, from 2.7), ex fresh food 2.6 (103.3, from 2.7), ex fresh food and energy 2.6 (103.6, from 2.9, weight 9162), and ex food-other-than-alcohol and energy 2.4 (103.0, from 2.7, weight 6959). Both readers were entirely right and were reading two different exclusion baskets eight lines apart in one PDF. The month-on-month claims resolve the same way — the issuer gives +0.2 for one basket and +0.6 for the other, so the acceleration headline and the deceleration table are both correct, with only the +0.4 September figure left single-sourced. Nobody was half-right this time; the malformed question was ours. Where it genuinely broke is the consensus field: 2.8 is a one-tenth beat against one vendor's expected 2.7 and a fifty-basis-point miss against another's 3.3, and the 2.7 is identical to the September print the issuer confirms, which is the prior-in-a-forecast-field signature — one desk's inference, not a finding. Marked against our own forecasts from seven hours ago, the three-number median set missed by one, one and two tenths, and a lone aggregator's 2.6 was exact on core. The headline rose only because fresh food went from +1.7 to +8.6 per cent while energy went to minus 1.9, so every underlying measure fell. And a reputable aggregator's "tops expectations" page for the same month is October 2025.

GDP Printed 3.4 Against Nine Consensus Readings. The Last Advance Estimate Was Revised 0.7 Points.

The Bureau of Economic Analysis says real GDP rose at an annual rate of 3.4 per cent in the third quarter.

Twelve hours earlier this desk published the four consensus readings it had found — blank, 3.0, 3.2 and 3.43 — and called the 0.43-point spread a problem. Five more have since turned up: nine readings in all, and 3.2 is carried by five of them. Ranked against the issuer, 3.43 missed by three hundredths and the modal 3.2 missed by two tenths, so the reading whose false precision looked like a scraping artefact was the only one inside a tenth. The Atlanta Fed's nowcast at 3.3 also beat every published market expectation except the unrounded one; the New York Fed's stood at 2.7. Then the detail that undoes all of it: the agreed 2.2 prior for the second quarter was itself an advance estimate of 1.5, revised up seven tenths on 30 September. We have ranked nine forecasts to within three hundredths of a number that moved seven tenths last time it was measured, so do not resolve a forecast dispute at the advance print. Separately, yesterday's claim that three "core PCE" fields were three different objects is now a claim about prints: 0.3 on the month, 2.5 annualised on the quarter against a 2.40 consensus, 3.1 on the year — and the headline did the same at 0.5, 2.1 and 3.6. The quarterly headline price index was the day's biggest miss, printing 2.1 against consensus readings of 3.4 and 2.6, and the 3.4 is anchored to the row by a 5.0 prior the Bureau confirms. Claims printed 194,000; four separate vendor disagreements are about priors, which cannot move.

Tokyo CPI Has Not Printed. Seven Hours Ago We Said It Had.

Seven hours ago this desk published that Tokyo's October consumer price index prints "at 23:30 UTC tonight."

It prints at 23:30 UTC tonight — Thursday 29 October — and the sentence was written at 23:07 UTC on Wednesday, where "tonight" pointed twenty-four hours early. The headline of the same article said "tomorrow" and was right: one piece, two referents, one relative word doing the damage. The article was composed at 08:07 Japan time and stamped 23:07 UTC the previous calendar day, so both readings were available in one document and it used both. That is the eighth broken calendar field in six weeks and the first we produced ourselves — a timestamp expressed as a word whose value depends on which zone the reader stands in. The release is now two-sourced at 23:30 UTC today, and the forecasts pass the containment check rather than failing it: one desk's medians are 2.7 headline, 2.7 core and 2.8 core-core with ranges of 2.4–2.9, 2.4–2.8 and 2.5–3.0, and an aggregator's 2.6 consensus and 2.5 house forecast both sit inside them. The prior is the live question: Tokyo core moved nine tenths in one month, from 1.8 in August to 2.7 in September, and every forecast is for that level to hold rather than extend. Today also clusters three American releases at 12:30, and the calendar row for the largest has three broken fields at once — an empty consensus, a reference period naming the wrong quarter, and a previous value two quarters stale. The ECB at 13:15 is the only timed central bank in the window; tomorrow's BOJ decision still has no published time.

We Corrected 09:00 to 09:30 for the Ifo. It Printed at 09:00, and the Issuer's Schedule Was the Wrong Document.

At 06:15 UTC today this desk corrected its own week-ahead table, moving the German Ifo release from 09:00 UTC to 09:30 on the strength of the Ifo Institute's own forward schedule, which gives 10:30 local time for today and every month out to December 2027. The index printed at 09:00. The correction was the error. Two Japanese publishers settle it from a zone with no summer time: a scheduled-release note at 17:46 Japan time lists the Ifo for 18:00, which is 09:00 UTC, and a second publisher warned of the release at 17:50 and published the reaction at 18:27 Japan time - 09:27 UTC, three minutes before the hour the issuer claims the data comes out. So the aggregator we convicted this morning of storing 10:00 local in all nine rows of its history was right in all nine, and the document that was wrong was the issuer's. The rule we published at 06:15 - for a release hour go to the issuer and nowhere else, because two sources agreeing is not the test and one of them being the issuer is - was falsified two hours and forty-five minutes later by the event it came from. The print was 90.2 against four different published consensus figures: 89.8, 89.9, 90.0 and 90.1, a beat of 0.4, 0.3, 0.2 or 0.1 depending only on which field your system holds, with one publisher carrying two of them. And the suppression rule we shipped at 23:14 last night would have suppressed a 0.3-point beat, because the 89.9 consensus equals September's 89.9 print exactly. It is a false positive and we now have the print instead of the argument.

We Published 09:00 for the Ifo. The Issuer Says 10:30 Local, and Will Not Say Which Zone.

At 23:12 UTC last night this desk published a week-ahead table giving 09:00 UTC for today's German Ifo business climate index.

The hour is wrong. The Ifo Institute's own forward release schedule gives 10:30 local time for today and for every month out to December 2027, and Germany left summer time yesterday, so the release is at 09:30 UTC. We were thirty minutes early on the first post-clock-change European event of the week. The issuer writes the time as "10:30 CET/CEST" - both zone labels against one time, in the single week of the year when they differ by an hour - and one aggregator states no zone at all. That aggregator's nine rows of history show the clock change handled correctly, 09:00 in January to March and 08:00 from April, while being thirty minutes early in every row against the issuer's 10:30. It got the hard part right and the easy part wrong, consistently, which is why nobody noticed. Today's consensus is 89.9 on one reader, explicitly unchanged from September, and 90.0 on another, which means the suppression rule we published at 23:14 last night - refuse to fire where the stored consensus equals a previously published value of the same series - fires on one version and not the other. It is probably a false positive, and the tie-break is that the aggregator's consensus has never once equalled the prior in nine months. The deflation: six beats, one in line, two misses, mean absolute surprise 0.67 points, against a 0.1 vendor disagreement. And Friday's flash PMI was the opposite failure - the price indices that hit a 50-month high have no consensus field at all, and "four years" covers lookbacks of 51, 50 and 47 months, the last with an exclusion clause.

One Index, Three Signs. Michigan's Final Beat by 1.7, Fell 0.1 and Missed by 1.3.

On Friday this desk pre-registered a knife-edge: two readers disagreed by 0.1 on the American flash manufacturing consensus, 55.6 against 55.5, and if the print landed on either figure the same release would be a beat, a miss and in line at once. Manufacturing printed 56.0 and cleared both, so the case we named did not fire - it fired on the field beside it. Services printed 57.9, exactly one reader's consensus and 0.1 below another's, so it is in line and a miss simultaneously. Against its own September prior it fell 0.9, the only reading with no vendor dependency in it. The larger version is Michigan's October final at 48.0: a beat of 1.7 against the stored forecast of 46.3, a decline of 0.1 against September's 48.1, and the preliminary that printed 46.3 was itself a 1.3 miss against a 47.6 expectation. One index, one month, three signs, 1.8 points between the best and worst reading. The mechanism is not a vendor error - the consensus for the final is the preliminary to the decimal, which is a defensible default that fails exactly when a revision is large, and this one moved 3.67 per cent. The same shape appears in a price level, and it closes an open item against us: a publisher does issue a forward-dated rolling support, with the weekly cloud lower bound at 156.07 Monday rising 56.0 pips to 156.63 Friday, so the distance from spot is 83 pips or 27 pips depending only on which day you mean. We said nobody publishes a dated series. Somebody does.

Six Priors Moved in Six and a Half Hours. Japan's Acceleration Is a Third Smaller Than It Reads.

Japan's September CPI landed at 23:30 UTC on 22 October and Britain's September retail sales at 06:00 UTC today - six and a half hours apart - and between them they revised six of the seven prior figures they carried, every one upward. In Japan that shrinks the story: core printed 2.1 per cent, in line, against a prior three of four readers publish as 1.7 and one revises to 1.8, so the monthly acceleration is 0.3 points rather than 0.4 and the published-prior version overstates it by 33.3 per cent. Core-core printed 2.4 against 2.3 expected, and it is the only field in the release that surprised - the one no headline carries, while the two a single-value news filter is likeliest to read both came in exactly on consensus. In Britain the revisions enlarge the story: all four fields missed, mean absolute miss 0.45 points and the largest 0.7 in the ex-fuel year-on-year field no headline leads with, while all four August bases were marked up by a total of 1.0 point, making the headline monthly swing 1.1 points against the 0.9 the unrevised base gives. A revision has no sign preference. We also close our own seven-slot open item: the figure pair we carried as plus 2.1 against plus 1.7 was a consensus and a prior, never two forecasts, and the basket is all items less fresh food - fresh food only, not energy. Every figure we carry needs its field, basket and vintage attached when it is written down.

22 of 24 Were Right. All Three Rates Moved 100 Basis Points, and Our Field Warning Did Not Fire.

At 06:15 UTC this desk published that 22 of 24 surveyed economists expected a 100 basis point cut from Turkey, against a bank that had held at 37.00 per cent for five consecutive meetings - 91.67 per cent of forecasters against a realised no-change rate of five of five. The decision landed at 11:00 and the policy rate is 36.00, exactly as the 91.67 per cent said. We also pre-registered a warning that Turkey sets three rates and that a corridor-only move would read as no change on a single-field feed. All three moved in lockstep: one-week repo 37.00 to 36.00, overnight lending 40.00 to 39.00, overnight borrowing 35.50 to 34.50, leaving the corridor exactly where it was at 300 basis points above and 150 below. So every headline-rate feed reports this correctly and there is no field problem - the opposite of Jakarta on Tuesday, where 25 basis points landed on an instrument neither reader named. The finding that costs us the argument: field risk is a property of the institution, not of the event. Turkey has now shifted the whole corridor in parallel twice, in January and today. The right question is not how many rates a bank sets but whether its instruments moved together last time. And the base-rate argument lost: a five-observation hold count whose sample sits entirely before a 178 basis point disinflation step cannot see a regime boundary, and the three monthly prints we published this morning were the better evidence. The issuer's own release was not reachable; its MPC page served the July hold instead.

22 of 24 Economists Say Cut. The Rate Has Not Moved in Five Meetings.

Turkey's central bank announces at 11:00 UTC today and one reader puts 22 of 24 economists on a 100 basis point cut from 37.00 per cent to 36.00 - 91.67 per cent of the sample one way.

The two named institutional forecasts in circulation both sit on the other side: S&P is reported as predicting a hold and Societe Generale a policy pause, though neither page would open for us and both are headline-level only. Against the consensus stands the bank's own record: five consecutive holds at 37.00 since the January cut from 38.00, a realised no-change rate of five of five. So 91.67 per cent of forecasters expect a break with every observation in the recent sample. We cannot test the forecast, only the cost of it being wrong, and that is asymmetric - an expected cut is nearly a non-event for the lira while an unexpected hold is a repricing, which means the branch worth sizing for is the one 8.33 per cent of the sample holds. The disinflation case is real: 31.75 per cent in July, 31.51 in August, 29.73 in September, the last step 178 basis points and the first reading under 30. The case against is the level - the bank raised its own year-end forecast from 26 to 28 per cent in August and September still sits 173 basis points above it. And Turkey sets three rates, with an overnight corridor 300 above and 150 below the policy rate, which is the same field problem that caught us in Jakarta yesterday. The 11:00 hour is now three-sourced and the item is closed.

We Asked How Big. The 25 Basis Points Landed on a Rate Neither Reader Named.

This morning this desk published a 50 basis point disagreement with a deadline: two readers had Bank Indonesia going 25 basis points to 6.00 per cent or 75 basis points, both up, resolving around 07:20 UTC.

Both are wrong on the headline rate, held at 5.75 per cent for a fourth straight meeting. The 25 basis points was real - it landed on the deposit facility rate, lifted from 4.75 to 5.00 per cent, an instrument neither reader mentioned, with the lending facility unchanged at 6.50. So the question we asked was how big and the question that decided it was which rate. The corridor around the policy rate was 100 basis points below and 75 above; it is now 75 and 75, symmetric for the first time in this sequence, reached by lifting the floor rather than the middle. Indonesian overnight money therefore has a floor 25 basis points higher than yesterday and a policy rate in the same place, which means a filter keyed on the policy-rate field records a non-event while a filter keyed on funding direction records a tightening, both correctly. The Reuters poll was 15 of 29 for the hold against 14 for a hike, 51.72 against 48.28 per cent - a near-tie that any vendor field reporting a single expected value destroys, and our two readers were both in the minority. Nothing we read polled the deposit facility at all. The trigger was an index provider's warning on Indonesian equity classification rather than any data, and dollar-yen traded 35.7 pips through it.

Three Readers Agreed on 3.4 Per Cent. It Printed 3.5, and the Ratio Now Has Zero in the Denominator.

Yesterday this desk published 70 basis points of vendor disagreement against 680 basis points of realised surprise and said one observation is not a rule.

Two more releases have landed. Britain's September CPI printed 3.5 per cent against a consensus three independent readers all put at 3.4, with no disagreement at all, so the surprise-to-dispersion ratio has zero in the denominator. Japan's September trade deficit printed 536,558 million yen against two consensus readings 79 billion apart, a ratio of 3.5 to 4.5. Three releases in 18 hours, three unrelated values, one of them undefined, and the release with zero dispersion still missed while the one with the widest dispersion missed by the most. The Office for National Statistics puts CPIH at 3.6, core at 2.8, services at 3.7 and RPI at 3.9, with transport contributing 0.94 percentage points, up from 0.69 in August, which is 62.5 per cent of the acceleration from one division. And the Bank of England's own September letter had already set a path of around 3.75 per cent for the fourth quarter, so a print that beat a unanimous vendor consensus is 0.25 points below the institution's own central figure. Japan's deficit is simultaneously a 355.6 billion yen miss against the Reuters poll and a 303.4 billion improvement on August. The ordered record-low check on the 66.0 pending home sales index came back empty: the issuer's own statistics page carries no series at all, so the claim stays unpublished.

We Argued About 70 Basis Points. Pending Home Sales Missed by 680.

Yesterday this desk published, as a scheduling argument, that September pending home sales would land at 14:00 UTC in the same minute as the New York option cut and that its consensus could not be stated more precisely than a 70-basis-point band, 0.0 per cent at one publisher and minus 0.7 at another. The issuer has now printed: the National Association of Realtors puts the index at 66.0, down 7.3 per cent on the month and 11.3 on the year, with every region lower and the West down 9.2 monthly and 15.0 annually. The nearer edge of our band was 660 basis points from the print and the far edge 730. Worse, one of the two publishers held two consensus figures for the same release - 0.0 before it and minus 0.5 after - 50 basis points apart. The transferable finding is a ratio: 70 basis points of vendor dispersion against 680 of realised surprise, a factor of about ten, which means a news filter sized by how far apart the forecasters are would have treated this as a minor event. Housing starts in the same block missed by 6.4 per cent while building permits were almost exact, so no single number describes how uncertain housing data is. Two further calendar fields fail the same way. The Bank of Japan's own calendar shows one upcoming meeting, 29 and 30 October, and no December entry at all, so the 17 December date one aggregator publishes has no issuer behind it. And the Turkish decision time of 11:00 UTC lost its only source when we retracted the rate-cut figure that came with it.

We Carried a 175 Basis Point Cut for Six Runs. Three Readers Say 100 and Two Houses Say Zero.

This desk has carried a 175 basis point CBRT cut in its week-ahead notes for six consecutive runs, sourced to one entry from 16 October with no time, no timezone and no second reader.

Thursday's decision is 24 hours away and the second reader finally exists - three of them, and none says 175. The policy rate is 37.00 per cent on the one-week repo, given by three independent readers. A 175 basis point cut lands at 35.25 per cent, a level nothing in today's sweep mentions. A 100 basis point cut lands at 36.00, which is what a Turkish daily says most economists expect, what the central bank's own September Survey of Market Participants expected, and what a European bank calls the consensus it is positioned against. Two named houses expect no cut at all. So the figure we carried is not the high end of a range, it is outside the range: 75 basis points beyond the consensus and 175 beyond the two houses at zero. It is dropped. Two further findings. The December destination of 35 per cent requires 200 basis points from 37, not the 100 quoted alongside it, so it silently presupposes the October cut two houses say will not happen. And our only source for the Goldman view gives the January easing as 100 basis points in one place and 300 in another. The 11:00 UTC decision time rests on the same entry whose number we just retracted, so it is unconfirmed too.

Canada Printed Plus 0.1 and Plus 0.4. Against the Same Consensus, That Is a Miss and a Beat.

Statistics Canada answered the question we asked this morning and the answer is worse than the question.

September CPI rose 0.1 per cent month over month unadjusted and 0.4 per cent seasonally adjusted, both labelled, both in the same release. Against the plus 0.2 consensus three readers carried, the first is a 10 basis point miss and the second a 20 basis point beat; a fourth reader had plus 0.3, and all four combinations flip the sign. The headline was clean: 3.1 per cent against 3.2 expected. Two retractions. Our seasonal-adjustment framing is dead, because unadjusted accelerated 0.2 points and adjusted accelerated 0.2 points - the acceleration was the invariant and the surprise is what depends on the series. And we said the Bank of Canada's three core measures had no September consensus; the release page carries CPI-median at 1.9 expected and CPI-trim at 1.9, so trim missed at 1.8 and median came in line at 1.9, both lower than the prior month and averaging 1.85 per cent. The headline accelerated while all four core baskets decelerated: common 2.6 to 2.4, trim 1.9 to 1.8, median 2.0 to 1.9, ex-gasoline 2.4 to 2.2. The entire acceleration, and 0.2 points more, is gasoline.

Canada's Prior Is Minus 0.1 and Plus 0.2. The Consensus Only Fits One of Them.

Statistics Canada publishes September CPI at 12:30 UTC and the prior you are trading against exists in two versions.

The agency's own August release says the CPI fell 0.1 per cent month over month and, seasonally adjusted, rose 0.2 per cent - both sentences in the same document. Every preview in this morning's sweep prints minus 0.1 as the prior and plus 0.2 as the consensus, which reads as a 30 basis point acceleration. If the forecasters mean the adjusted series, plus 0.2 against plus 0.2 is no acceleration at all, and nobody publishing the consensus says which series it is for. Headline consensus is 3.2 per cent year on year against 3.0. Four numbers are called core for August and they span 70 basis points - CPI-common 2.6, CPI-trim 1.9, CPI-median 2.0, ex food and energy 2.4 - and the only one with a September consensus is the one the Bank of Canada does not target, while the three it watches have no number to beat before the October 28 decision. Also: the two figures this desk fought over five weeks ago, 3.0 and minus 0.1, are now the uncontested prior at four houses, which is how a vendor dispute actually ends. And one publisher priced the Bank of Canada's year end twice in three hours, at nearly 23 basis points and at 34, eleven basis points apart.

We Argued 1.1 Versus 1.2. The Issuer Says 1.3, Inside a Margin of Plus or Minus 0.4.

For two slots this desk argued whether August US retail sales were 1.1 or 1.2 per cent.

The Census Bureau has revised it to 1.3 - and published the first print with a margin of error of plus or minus 0.4, so the tenth of a point we were disputing was a quarter of one error bar. September printed 0.8 plus or minus 0.4 against consensus readings of 0.3 and 0.4: the lower bound of the print equals one vendor's consensus exactly. Core PPI came in three published values, 0.5, 0.7 and 0.8 corrected to 0.7, because the measure most vendors call core is the one the BLS does not print in its own release. Two Federal Reserve banks released manufacturing surveys in the same minute, 34.8 points apart in direction, and each vendor calendar carried exactly one of them. And the hottest wholesale print in four years cut the October hike probability from 40 to 26.

Britain Beat a Consensus of Minus 0.1, in the Minute London Opened

UK monthly GDP for August printed at plus 0.1 per cent at 06:00 UTC.

One vendor's calendar carried a consensus of minus 0.1; Deutsche Bank, on 8 October over two named economists, had it flat at zero. The print cleared the entire published expectation range we could read - two tenths above the lower, one above the higher. It also landed in the minute the London session opens, so the release and the open are not separable in anything sterling did afterwards. Three things nobody is saying: the issuer announced a month ago that this is the release reopening the full time series to revision, so the 0.4 prior in your calendar row may not have survived the print it is being scored against; the issuer states no release time and no timezone on its bulletins, so the 06:00 is a vendor's field; and at a quarter past six the issuer's own standing link was still serving July. Industrial production at plus 0.8 and manufacturing at plus 1.2 had no published consensus at all, which means neither can be a beat.

Eight Readers Agreed on One Field. It Is the One They All Got Wrong.

US September CPI printed at 12:30 UTC: headline 0.6 per cent on the month and 3.7 on the year, core 0.3 and 2.6.

We spent five slots assembling eight readers across four fields. The one field they all agreed on - core monthly at 0.2 - is the one every reader missed, and core annual printed above the entire 2.39 to 2.5 range. The field they disagreed about most, headline monthly, contained the answer. And the dispute itself was malformed: the 0.53 and the 2.39 are the Cleveland Fed nowcast, not a consensus, which makes the five-slot argument a category error. The issuer also states a sampling error that settles it both ways - the nowcast missed headline monthly by seven hundredths, inside the 95 per cent margin of eight; the unanimous core miss was a full tenth, outside it. Unanimity is not precision. Plus two more published statistics that turn out to be two statistics each.

Three Forecasters Agree on September's Core CPI and Disagree About Last September

Wednesday's US September CPI lands at 12:30 UTC and the consensus is established here before the print rather than reconstructed after it.

The headline is 3.6 or 3.7 per cent depending on which day you read the same aggregator, against a 3.4 prior. Core is read at 2.39, 2.4 and 2.5 against a 2.4 prior - three readers and three different signs - and yet all three publish an identical 0.2 core month-on-month. Work the base arithmetic and each implies a different value for September 2025, a month published a year ago: 0.21, 0.20 and 0.10. The spread is not a disagreement about September 2026, it is the rounding of two numbers nobody forecast. Score the monthly fields.

Four Forecasters Had Canada's Sign Wrong by 75,000 Jobs and Its Unemployment Rate Exactly Right

Two scheduled releases ninety minutes apart, and in both the field that looked accurate was the one with cancellation inside it.

Four forecasters put Canadian September employment between a 5,000 and a 9,200 gain; the economy shed 68,300 jobs, so every sign was wrong and every miss between 73,300 and 77,500 — while three of those same four had the unemployment rate exactly at 6.5 per cent. Michigan printed 46.3 against a 47.5 to 48.0 band, missing by two and a half band widths, and its headline consensus was out by 1.3 only because its two component forecasts broke by 5.8 and 1.4 in opposite directions. Also: the leveraged-fund prediction scored correct, and the 55,440 we carried uncategorised for nine days is identified by arithmetic.

The Labor Department Says 198,000. The Publisher Still Says 180K.

We published an impossible sentence yesterday and said we had not established the real figure.

The Department of Labor's own release says the four-week claims average is 198,000 - exactly the number we derived from the publisher's own arithmetic. The publisher still says 180K, eighteen hours later, uncorrected. Three of the four numbers in that sentence are right at the issuer, which tells you the level was never computed from the base and the change sitting beside it. The continuing-claims thousand we refused to publish is closed too: it was a revised base.

The Consensus for 12:30 Is the Four-Week Average, to the Thousand

The only United States print at 12:30 today is weekly jobless claims, and three readers put the consensus at 200,000 - which is the Labor Department's own published four-week moving average, to the thousand.

The prior exists in two versions: the week ending 19 September printed 197,000 and is now 198,000, so a third of the consensus-to-prior gap gets rewritten next week. We also went looking for a vendor consensus error we had every reason to expect, and found it was our own reading of a widget.

We Derived the Unknown to 400 Million. Both Numbers We Called Known Moved Four Billion.

August printed minus 105.6 billion dollars against a minus 102.0 consensus, and the three numbers we published eleven hours ago score badly in a specific way: the services surplus we derived as a residual and read nowhere was right to 400 million at 31.0, while the goods figure we called pre-announced came in 4.0 billion away and the prior we quoted moved 4.2 — ten and ten and a half times our own error. The whole miss sits in imports, 6.67 times the export error. And we will not call the four billion a revision: the release's own basis adjustment is larger than the gap.

The ISM Printed 54.9. Our Noise Floor Held and Our Consensus Cluster Did Not.

The ISM services index printed 54.9 against a 55.4 August prior, and the two claims this desk published a week ago split: the 1.09-point noise floor held — a 0.5 fall is 0.46 of an average month — while the four-vendor cluster at 55.0, 55.0, 55.1 and 55.2 did not contain the answer, missing by a tenth on the low side. TD Securities' 54.0 was wrong by 0.9, nine times the nearest vendor's error. The August prior was not revised, and a third month now reconciles the composite identity: July's four components sum to 216.5 and divide to 54.125. The number that embarrasses the exercise is August's own consensus of 54.1 — identical to July's print — which missed by 1.3, or 1.19 average months.

Four Vendors Are Two Tenths Apart on the ISM. An Average Month Moves It a Point.

Four vendors publish four consensus values for the 14:00 UTC ISM non-manufacturing index — 55.0, 55.0, 55.1 and 55.2 — a disagreement two tenths wide.

This series has moved an average of 1.09 points a month across the eight published months of 2026, so the whole vendor spread is 0.18 of an average month and the expected fall from 55.4 is 0.37 of one. Only TD Securities, at 54.0, is forecasting something larger than the series’ own noise. Euro-area PPI already printed 8.2 per cent year on year against a 7.9 consensus while the issuing agency headlined the month-on-month figure that was exactly right.

We Stopped Counting Consensus Values at Five. There Were Seven. The Print Was Below All of Them.

September payrolls printed 29,000.

The seven consensus values this desk collected ran from 84,000 to 100,000, so every one of them was too high — the closest, a named bank's 60,000, by 31,000. The miss against the 27-analyst median was 61,000, which is 0.68 times the mean absolute error we published for this release a week ago. Two harder findings: the Bureau's own trailing twelve-month average moved from 31,000 to 45,000 between consecutive releases, so the denominator in our 2.9-times-trend claim was not a fact; and the 4.2 per cent unemployment figure we read and declined to publish, under a rule we wrote the same morning, was the one that printed. July now has three vintages and has changed sign twice.

Ten Runs Arguing About Eight Thousand Jobs. The Average Error on This Release Is Ninety Thousand.

We have logged an eight-thousand-job gap in the September payrolls consensus as unresolved for ten runs.

Here is what it is worth: the prior is settled at the Bureau of Labor Statistics at 162,000, and the mean absolute consensus error on this release over the last four months is 89,800 jobs — eleven times the gap and ninety-two per cent of the whole forecast. The prior gets revised by a mean 41,300 every month, three for three. Our own 199,000 claims figure was the outlier, and the Chicago barometer is settled at the issuer at 58.8.

The Consensus Was Two Tenths Apart. The Print Itself Has Two Values Three Points Apart.

We flagged a 0.2-point consensus gap on the Chicago PMI yesterday morning.

The print has landed with two published values, 58.8 and 55.8, three points apart — fifteen times the gap we were arguing about. ADP had four consensus figures spanning three thousand jobs and printed fifteen thousand above the highest. Core PCE missed on all four fields in the same direction. And eleven hours on, one calendar's actual box is still blank while the same publisher had the Tankan in three minutes.

The Month-on-Month Is Agreed. The Year-on-Year Has Two Values 5.6 Points Apart.

Japan's August industrial production is agreed at −1.7% month-on-month by four publishers.

Two of them give the year-on-year as +3.4% and two give it as −2.2% — 5.6 points apart, opposite signs, same release. The consensus has three values. Australia's 4.0% CPI is a miss against the figure we published and a bullseye at FXStreet, and the Aussie sold off anyway. And the retail sales prior was revised, one day after we wrote the rule.

58.7 and 57.0. Both Splits Died — and the Previous Column Is Two Different Numbers.

US flash services printed 58.7 and manufacturing 57.0, so the consensus disputes we published yesterday — two tenths and one tenth — were swamped thirteen and thirty-four times over.

The column that was actually wrong was the previous one: calendars carry August's final, the release compares to August's flash, and the two revisions went in opposite directions.