Yesterday this desk published a sentence from a major FX publisher that could not be true: that the four-week moving average of US initial jobless claims “went down by 2.5K to 180K vs. the previous week’s revised prints (200.5K)”. Two hundred thousand five hundred minus two thousand five hundred is one hundred and ninety-eight thousand, not one hundred and eighty thousand. We derived 198.0 thousand, labelled it as our own arithmetic, and said in terms that we had not established what the Labor Department actually published because we had not fetched the issuer. We have now fetched the issuer. The figure is 198,000. The publisher’s page, eighteen hours later, still says 180K.
The issuer settles it, and it settles it to the digit
The Department of Labor’s weekly release for the week ending 3 October, embargoed to 08:30 Eastern on Thursday 8 October, says the seasonally adjusted advance figure for initial claims was 197,000, that the previous week’s level was revised up by 2,000 from 197,000 to 199,000, and that the four-week moving average was 198,000, a decrease of 2,500 from the previous week’s revised average.
That is the whole answer. 198,000 at the issuer against our derived 198.0 thousand, and 198,000 plus 2,500 is 200,500, which reconciles the base the publisher quoted as well. The gap between the issuer and the publisher is eighteen thousand claims on a single field.
We are stating this plainly because we were careful yesterday not to: an internally impossible sentence tells you one of its numbers is wrong, and it does not tell you which. It could have been the change. It could have been the base. We said so and we went no further. One fetch at the issuing agency, which is the standing rule on this desk for anything on a fixed weekly cadence, closed it in under a minute.
Three numbers right, one wrong, and that is diagnostic
Hold the two published sentences against the release. The 2,500 decrease: correct. The 200,500 previous revised average: correct. The 197,000 revised to 199,000: correct, and worded almost identically to the issuer. The 180,000 level: wrong by eighteen thousand.
That pattern is more informative than a single wrong number usually is. If the level had been computed from the base and the change — the two figures sitting in the same clause — it would have come out at 198,000, because those two figures are right. It did not. So the level was fetched or entered separately from the arithmetic printed around it, which is exactly the failure mode that a sentence like this is supposed to make impossible.
If you are consuming a vendor feed, the operational point is that internal consistency is a test the feed can fail while still being mostly correct. Three of four numbers were clean. A reasonableness check on each field independently passes on all four, because 180,000 is a perfectly plausible claims average. Only the relationship between them fails. The check that catches this is not a range check on the value — it is recomputing the identity the sentence asserts, and rejecting the record when the identity does not close.
For completeness: we re-read the page this morning, cache busted, asking only for verbatim quotation of every sentence containing “moving average” or “revised”. The sentence is unchanged and there is no correction notice. We are not naming the publisher in this body. The footer cites the URL, which has now been read three times across two days, and anyone can check it.
The thousand claims we refused to publish yesterday, closed
A smaller open item from the same release. Continuing claims came in at 1,716,000 for the week ending 26 September, described as an increase of seventeen thousand. The forecast carried by the readers we used was 1.70 million, and 1,716 minus 1,700 is sixteen thousand, not seventeen. We declined to publish the one-thousand discrepancy yesterday on the grounds that we had not established the base, and that accusing a publisher on an unestablished base would have repeated the exact mistake the companion article was about.
The issuer’s wording settles it: an increase of seventeen thousand from the previous week’s revised level. That makes the revised prior 1,699,000. Our sixteen thousand was computed off the forecast, which happened to sit one thousand above the revised print. Nothing was wrong with anybody’s number; the two figures were differencing against different bases.
This is the dull version of the same lesson as the section above, and it is the one that costs people money more often. A change field and a level field in the same row are frequently measured against different things, and a revision silently moves one of them. If your system stores the change rather than recomputing it from two stored levels, a prior-week revision will leave you holding a number that no longer reconciles to anything.
The consensus matched the four-week average and still missed
One more mark to score. Two days ago we noted that the 200,000 consensus for this print was identical to the published four-week moving average, and asked out loud whether a consensus equalling the trailing average is normal practice or a coincidence of one week.
The print was 197,000. The consensus was three thousand high. So whatever that identity is, it did not predict this week, and the observation stays an observation rather than becoming a method. One data point, pointing the unhelpful way.
The revision is the more interesting half. The prior week went from 197,000 to 199,000, two thousand, which is two thirds of the three-thousand gap between the consensus and the prior as it stood when the consensus was set. We published one third yesterday from the figures then in hand. Two thirds is the corrected number and it strengthens rather than weakens the original point: a meaningful share of the apparent surprise in a weekly claims print is the previous week’s number moving underneath it. If you trade the surprise, you are trading a difference in which both terms are provisional.
And there is a second, cleaner confirmation sitting in the same release, which we only noticed because we had published the first number ourselves. Yesterday we put the previously published four-week moving average at 200,000, sourced to the Labor Department. This week’s release calls the previous week’s revised average 200,500, and so does the publisher whose level we are disputing. Those two figures differ by five hundred. A four-week average moves by exactly one quarter of any revision to one of its four weeks, and the revision reported in this same release is two thousand. Two thousand divided by four is five hundred. The 200,000 we published, the 200,500 the issuer now carries and the 2,000 revision all close on each other to the exact thousand, which is why we are comfortable saying the base in that impossible sentence was never the problem.
What this does not tell you
We have established what the Labor Department published and what one publisher published. We have not established how the publisher’s 180,000 arose, and the reconstruction in the second section is an inference from which numbers were right, not a finding about anyone’s pipeline. A transposition, a stale cache, a mis-mapped field and a typo all produce the same evidence we have.
We have read exactly one publisher on this field. We have not surveyed others, so we cannot say whether 180,000 propagated anywhere, and a reader whose vendor sources this field downstream should check their own rather than assume ours is representative.
None of this is a view on the labour market. 197,000 against 200,000 expected, with a four-week average at 198,000 and continuing claims up seventeen thousand, is a set of figures consistent with a labour market that is not deteriorating quickly, and that reading would be unchanged if every number here had been printed correctly everywhere. The argument is about the plumbing. The plumbing happens to be what an automated system trades on, which is why it gets an article and the labour market does not.
Finally, the two-thirds figure in the previous section rests on the consensus having been set against the 197,000 prior rather than the 199,000 revision. That is the natural reading of the sequence and we have not sourced it to anyone who says so explicitly.
