The Australian Bureau of Statistics says the monthly consumer price index rose 4.3 per cent in the twelve months to September, up from 4.0 per cent in August. Two publishers put that same row at 3.9 per cent against a consensus of 4.2, with the prior correctly at 4.0. If your system read either of them, it recorded a 0.3-point miss on the headline. The issuer’s own number is a 0.1-point beat. The magnitude is small and the sign is opposite, and that is the whole problem with storing an economic release as one row.
What the issuer actually published, at 00:30 UTC
The release went out at 11:30am AEDT, which the Bureau states on its own page as Canberra time, so 00:30 UTC — the one conversion this desk did not have to guess at, and we said so yesterday. What came out of it is four separate annual inflation rates in a single minute, under a single calendar row most systems hold as “Australian CPI”.
From the Bureau: the monthly indicator rose 0.8 per cent on the month and 4.3 per cent on the year, from 4.0 per cent in August. The monthly trimmed mean ran 3.6 per cent annual, up from a revised 3.5 per cent. The quarterly series, which the Bureau now carries in appendix tables rather than as the headline, rose 1.3 per cent in the September quarter, with the quarterly trimmed mean up 1.0 per cent in the quarter. The next release is 25 November.
Four numbers, and the Bureau does not print all of them in one place. The quarterly annual figure is not in its key statistics at all, which matters in a moment.
The row that is wrong in both vendors, and it flips the sign
Both publishers we could read carry a row reading, in substance, monthly CPI year-on-year 3.9 per cent, expected 4.2, prior 4.0. The prior is right, which tells you the row is genuinely pointed at the monthly series. The print is not: the issuer says 4.3.
And 3.9 is not a random error. It is exactly the quarterly annual figure, which both of the same tables also carry, correctly, one row away. So the actual field was populated from the wrong series while the expected field and the prior field stayed with the right one. That is the failure this desk has now seen in some form for six consecutive weeks, and it is the most expensive variant yet, because it does not look broken. A row reading 3.9 against 4.2 from 4.0 is internally coherent. It tells a story — inflation accelerating less than feared — and the story is backwards.
A system that scores surprises by subtracting consensus from actual got minus 0.3 where the issuer gives plus 0.1. Nothing in the row would have raised a flag. There is no consistency check available inside the row that catches this, because the row is consistent.
The field the central bank targets is 3.6 or 3.7 depending on who you read
This is the one we said yesterday would show no symptom, and we were wrong within twenty minutes.
At 23:12 UTC last night this desk published that the Australian consensus was 3.9 and 4.2 at once, that neither was wrong because the Bureau releases the quarter and the month in the same minute under the same name, and that the one field where the two agreed numerically — trimmed mean at 3.6 — was the field the Reserve Bank of Australia actually targets, so a mislabelled series would produce no visible symptom. The print produced three.
One publisher headlined the result as quarterly trimmed mean 3.6 per cent against an expected 3.6, prior 3.5 — dead in line. The other headlined trimmed mean rising to 3.7, above forecasts, with the Reserve Bank hike case firming. Same release, same minute, opposite verdicts on the only inflation measure the central bank formally targets. The Bureau gives 3.6 for the monthly trimmed mean against a revised 3.5, which matches the first publisher’s quarterly row exactly and neither publisher’s monthly row, both of which say 3.7 from 3.6.
We are not going to pick. The defensible statement is that the trimmed mean annual rate for September is 3.6 or 3.7, that the issuer supports 3.6 for the monthly series, that both publishers support 3.7 for the same series, and that we cannot establish from the Bureau’s key statistics what the quarterly annual figure is. If an argument you are about to make depends on that tenth of a point, it is weaker than it reads. The spread is your error bar, not ours.
And the prior was revised, and nobody carried the revision
The Bureau says August’s monthly trimmed mean annual rate was revised to 3.5 per cent. Both vendor tables carry the prior as 3.6.
This is the cheapest of the four errors to fix and the one most likely to be sitting in your own database right now, because priors are written once and rarely rewritten. It also compounds the others: on the issuer’s numbers the monthly trimmed mean accelerated by a tenth, from 3.5 to 3.6. On the vendors’ numbers it accelerated by a tenth, from 3.6 to 3.7. Identical momentum, different level, and a threshold rule set at 3.65 fires on one and not the other.
One publisher has a fifth problem we will name without building on it: its quarterly trimmed mean quarterly row reads 0.3 per cent against an expected 1.0, where the Bureau gives 1.0 per cent for that series and 0.3 per cent for the monthly change. That one is seventy basis points and is demonstrably the monthly figure in the quarterly row. We mention it because it sits two rows from that publisher’s own “in line” headline.
The error was invisible in the price, which is the uncomfortable part
None of this moved anything in the direction the errors point. The Australian dollar rose sharply after the print, overnight index swaps went above 80 per cent for a November hike on one reader, Governor Bullock told a Senate committee that inflation is likely to run above the August forecasts, and Deputy Governor Hauser called the data too high. The Australian dollar then stalled once the buying faded, and by 13:49 Japan time one Tokyo desk had it going nowhere.
So the market read the release as hot, which the issuer’s numbers support, while two published tables said the headline missed. Whatever the price is responding to, it is not those tables. That is reassuring about the market and not at all reassuring about your system, because a discretionary trader reads the Bureau or the wire and an automated one reads the row.
The fix is not a better vendor. We have now watched the issuer get a time zone right and a publisher get the hour wrong, a publisher get the hour right and the issuer get it wrong, and today a publisher and an issuer disagree on the print itself. The fix is to stop storing one number per event. Store the series name next to the figure, store the release’s own time zone convention, and store the prior with a revision flag — and where two readers disagree inside a tenth, size as though you do not have the number, because you do not.
What this does not tell you
We could not read the issuer’s full quarterly tables. The Bureau’s key statistics give the quarterly change and not the quarterly annual rate, so where this article says a vendor’s quarterly row disagrees with another vendor’s, that disagreement is between two secondary readers and is not settled by the issuer. It is possible that one publisher’s 3.7 is a correctly-labelled quarterly annual figure and that the two series genuinely differ by a tenth; we cannot distinguish that from a swap, and the distinction matters.
The 4.3 per cent headline is firm: the Bureau publishes it in its own media release and the figure appears in a second publisher’s headline against the same 4.2 consensus. The 80 per cent November pricing rests on one reader, and we have no second source for the Bullock or Hauser remarks beyond the paraphrases we read. We have not been able to open an option-expiry board for today at either of our usual publishers, so nothing in this article knows where the 14:00 UTC cut sits. And we have no figure at all for how much Australian inflation the Reserve Bank of Australia will say it expects when it meets next week.
