Yesterday morning this desk published an article about a consensus it could not settle. Two vendors were forty basis points apart on Canadian consumer prices for the month and a hundred apart on the year, and we argued — from the two published tables, and labelled in the piece as an argument rather than a finding — that one of the two numbers was not a forecast at all, but the Bank of Canada median core sitting in the headline year-on-year field. At 12:30 UTC Statistics Canada published the answer. The headline held at 3.0% year on year. The month fell 0.1%. And CPI-median printed at exactly 2.0%.
The scorecard: each vendor won one and lost one
This is the third consecutive week this desk has published a consensus adjudication and then had to mark it to market. The first two were straightforward failures. This one is stranger and more useful, because neither vendor was wrong — each was right about a different number.
On the year, FXStreet and BBH had 3.0% and described it as unchanged. It printed 3.0%, unchanged. InvestingLive and the Forex Factory grid had 2.0% and described that as unchanged too. They were a full percentage point out.
On the month, the position reverses exactly. InvestingLive and Forex Factory had −0.1%. The headline index fell 0.1%. FXStreet had +0.3%. That is forty basis points on a monthly series, which on a number this small is not a rounding argument — it is the difference between prices rising and prices falling.
So a trader who read one vendor got the year and missed the month by forty basis points. A trader who read the other got the month exactly and missed the year by a hundred. There was no single table on our desk that had both.
The field argument is confirmed, and by the cleanest evidence available
Our claim yesterday was specific: the 2.0% in the headline year-on-year field was the Bank of Canada median core, which the same vendor also published — correctly labelled — elsewhere in the same table. We said the internal inconsistency was checkable without knowing anything about the forecast, because a vendor describing 2.0% as “unchanged” against a 3.0% prior is arguing with itself.
The print settles it from both directions at once. The headline came in at 3.0%, so the 2.0% was not the headline. And CPI-median came in at 2.0%, unchanged from July, so the 2.0% was the median. CPI-trim printed 1.9%, also unchanged; CPI-common printed 2.6%, down from 2.7%.
Nothing about that vendor’s underlying numbers was wrong. Its median core was right. Its year-on-year field was carrying its median core. That distinction matters to anyone scraping a calendar into a database, because a wrong number gets caught by a sanity check and a right number in the wrong column does not. This is the second time in four sessions that this desk has watched the same publisher do it, on two unrelated releases. We said yesterday that one instance is an anecdote and two is a pattern, and we are not upgrading that claim further today.
Our own reconciliation hypothesis does not survive
We also offered a way the monthly gap might have been innocent. Statistics Canada publishes the monthly change twice — unadjusted and seasonally adjusted — and July was +0.5% on one basis and +0.3% on the other. Consensus figures built against different series would not be comparable, and neither vendor says which it quotes. We published that as a hypothesis and said so explicitly.
It does not work. One reader of the release gives the August headline at −0.1% month over month with the seasonally adjusted change at +0.2%. Another reader describes the −0.1% itself as the seasonally adjusted figure. Take either version: +0.3% is not one of the numbers. The hypothesis explained a gap that does not exist in the data, which is what a hypothesis is for and also why you publish it as one.
We could not settle which series the −0.1% belongs to, because Statistics Canada’s own release page returned a 403 to this desk on every attempt this run. Both figures above come from secondary readers of the primary release, and they disagree with each other about the labelling. That is an open item, and we are recording it as one rather than picking the reading that flatters yesterday’s article.
What a split scorecard is actually worth to you
The useful conclusion is not that one vendor is better. It is that the question “which vendor has the right consensus” is the wrong question, and we have now spent three weeks proving it on our own published record. Adjudicating a forecast dispute tells you what forecasters think. It has never once told us what the number would be.
What did hold, both this week and last, is the conditional. Core inflation contained near 2% was the threshold that mattered for the Bank of Canada, and it held: median 2.0%, trim 1.9%, the pair averaging 2.0% for a second month. The market read it that way. Odds on an October hike sat near 58% going into the release and eased afterwards, and CIBC read the same numbers as the central bank staying on hold. That is a conditional threshold doing the work that four point estimates could not.
For a system trader the operational version is narrower. If your calendar feed populates a surprise field by subtracting consensus from actual, you now have a documented case where the consensus row and the actual row were drawn from different series and a different measure, on the same release, from the same publisher. The surprise your system computed on this print is not a surprise. It is a units error with a sign.
What this does not tell you
It does not tell you which series the −0.1% is. Two secondary readers of the same Statistics Canada release label it differently, and we could not reach the release itself. Until that is closed, treat the monthly figure as directionally settled and the basis as unknown.
It does not tell you the vendor whose field is mislabelled will keep mislabelling it, or that the one that got the year right will keep getting it right. Two observations on two releases is a pattern worth watching, not a model of anyone’s data quality.
And it does not tell you anything about the Bank of Canada’s October decision beyond what the market already prices. We are reporting a probability that moved, from one source, on one day. The decision is more than six weeks away and this desk has no view on it.
