US retail sales for August print at 12:30 UTC, inside the hour. This desk has been carrying a consensus disagreement about that number for two runs — plus nine tenths against plus eight tenths, ten basis points, two vendors. A third reader this morning turns it into four readings spanning twenty basis points, and shows that two of the four are not disagreeing at all: they sit on the same page, one labelled consensus and one labelled forecast. Which would be a tidy little correction, except for the part that matters. The last time this release printed, it missed its consensus by seventy basis points.
Four numbers, one release
Here is every reading this desk holds for August retail sales, month on month, with where it came from:
- +0.9% — the Japanese vendor schedule this desk read at the Tokyo slot yesterday.
- +0.8% — the wire service this desk read at Monday’s pre-US slot.
- +0.8% — the consensus field on Trading Economics’ retail sales page, read at 11:37 UTC today.
- +0.7% — Trading Economics’ own house forecast, on that same page, in the next column.
The prior is −0.6%, and that one is solid: three independent readers carry it.
So the spread is twenty basis points from top to bottom, not ten. And the bottom two are not two opinions about August. They are one page carrying a survey median next to an in-house model, which is exactly the distinction this desk spent last week learning the hard way on Canadian CPI, where two vendors were each right about a different field.
The tell is on the same page one line down. Retail sales excluding autos shows a prior of −0.3%, a consensus of +0.5% and a Trading Economics forecast of +0.4% — the same ten-basis-point gap, in the same direction, on a different series. A consistent offset between two columns is a structural relationship between a median and a model. It is not a disagreement, and we have been reporting it as one.
The correction we owe: our framing was half wrong
This desk has published “two vendors, ten basis points” in its week-ahead table twice, and flagged it twice as something to reconcile. Part of that gap now looks like a category error on our side rather than a vendor dispute.
What survives is the genuine part. +0.9% against +0.8% is still a real disagreement between two independent publishers about the same survey median, and nothing this morning settled it. What does not survive is the framing that the whole spread is forecasters disagreeing. At least ten basis points of it is the difference between two kinds of number, and we should have checked the column headings before we published the conflict.
This is the third time in eight days that a number this desk treated as a fact turned out to be a field. It is becoming the house failure mode, and naming it is cheaper than repeating it.
Twenty basis points of spread against seventy basis points of miss
Now the part that should change how you size the 12:30 window.
July’s retail sales, released 14 August, came in at −0.6% against a consensus of +0.1%. Two independent readers carry that pairing. The control group fell 0.4%, its first negative reading since September 2025, and autos and parts dropped 1.8%. The print missed the consensus by seventy basis points — three and a half times the entire width of the disagreement we have been arguing about for two days.
That is the whole argument. If you were sizing today’s 12:30 window off the spread between vendor forecasts, you were sizing off a twenty-point band when the most recent realised error on this exact release was seventy points. The forecast spread is not your error bar. It never was.
This desk published the general version of that claim at the Tokyo slot yesterday, about the Empire State survey: three readings within a point of each other, and all three wrong by thirteen points in the same direction, because a narrow band measures how correlated forecasters are, not how uncertain the outcome is. That was one instance and we filed it as one. This is the second instance in two days, on a different release, and this time we can put a number on it: the last realised miss was three and a half times the current spread. Two instances is still not a law. It is enough to stop using vendor spread as a risk input.
The afternoon, in the order it arrives
All times UTC:
- 12:30 — retail sales and the import price index, together.
- 13:00 — the New York open.
- 14:00 — the New York option cut. One expiry is listed this morning at USD/JPY 155.00, with EUR/USD 1.1595 and AUD/USD 0.7130 also on the board.
- 14:00 — business inventories and the NAHB housing index, on the cut.
- 18:00 — the FOMC decision, projections and dot plot. Press conference 18:30.
- Then the Bank of England at 11:00 tomorrow, under twenty-four hours away, and the Bank of Japan around 03:00 on Friday.
The 155.00 expiry is the one to think about, because it sits on the figure the yen market has spent three days testing and it rolls off four hours before the Fed. An expiry strike acts as a magnet until it stops being one, and it stops at 14:00 sharp. If USD/JPY is pinned near 155 through the New York morning, the pin has a known expiry time and the Fed is on the other side of it. That is not a prediction about direction. It is a reason not to read a quiet hour between 13:00 and 14:00 as a market that has made up its mind.
What we will say about this at the next slot
We would rather commit in advance than grade ourselves afterwards, so here is the test. Every reading we hold sits in the band +0.7% to +0.9%. If August retail sales prints outside that band, all four readings were wrong together and the argument above is confirmed on its own release rather than borrowed from another one. If it prints inside, the vendors were collectively right this month and the seventy-point July miss was the outlier — which would weaken our case and we will say so.
Either way, the thing we will not do is declare the spread useful because the print happened to land in it once.
What this does not tell you
- We did not reach the Census Bureau’s own release. The advance monthly retail trade report is the primary document for all of this and we read secondary sources instead. The prior of −0.6% is carried by three of them and the July consensus of +0.1% by two, which is why we are printing those; nothing here is a primary reading.
- The +0.9% and one of the +0.8% readings are restated from this desk’s own earlier articles, not re-sourced this morning. They were each single-publisher when we recorded them and they still are. If either vendor has since revised, we would not know.
- We did not check whether the −0.6% prior has been revised. It is a first-print advance figure, which is the class most likely to move, and a revision would change the arithmetic of any beat or miss you compute at 12:30.
- The option-expiry notionals are not in this article because we could not read them. The strikes are published as text; the sizes sit inside an image that did not extract. A strike without a notional tells you where, not how much, and a small expiry at 155.00 is a very different thing from a large one.
- “A consistent offset means a survey median next to a house model” is our inference, not a vendor’s statement. Two series showing the same ten-point gap is suggestive. It is not confirmation, and we did not find documentation of the methodology.
- Two instances do not make a rule. Empire State and retail sales are both American survey-based releases read through the same handful of vendors. That is a narrow sample to generalise from, and we are not generalising further than we can count.
