The United States has exactly one scheduled print at 12:30 today: weekly initial jobless claims for the week ending 3 October. Three readers put the consensus at 200,000. The Labor Department’s own release from last Thursday puts the four-week moving average at 200,000. Those are the same number. The market’s expectation for this week’s claims figure is, to the thousand, the arithmetic mean of the last four claims figures — which tells you precisely what a forecaster does with a series that has no information in it, and should tell you what to do with the 12:30 window.
The consensus is the four-week average, to the thousand
The issuer’s release of 1 October, covering the week ending 26 September, gives seasonally adjusted initial claims of 197,000 and a four-week moving average of 200,000. Insured unemployment was 1,701,000 for the week ending 19 September. The unadjusted count was 156,738, which is 79.6 per cent of the adjusted figure — a reminder that the number everyone trades is four-fifths a seasonal factor by construction.
For today, one vendor calendar gives a forecast of 200,000 against a previous of 197,000; a second gives a consensus of 200,000 and its own house forecast at 200,000 as well; and the analysis desk of a third publisher writes that claims are “expected to rise by 3,000 to 200,000”. Three independent surfaces, one figure, and that figure is the published four-week average unchanged.
We are not calling that lazy. It is the correct forecast for a series whose last four prints were, on our reconstruction, 206, 199, 198 and 197 thousand — a reconstruction, because we have the published average and two of the four weeks firmly and inferred the rest to make 800. A mean-reverting forecast on a mean-reverting series is the right answer. The consequence for a trading schedule is the part nobody writes down: when consensus equals the trailing average, a print at consensus carries no information at all, and the only outcomes that move anything are the tails. You are not buying exposure to a number. You are buying exposure to a residual.
The prior is two numbers, and one of them stopped existing after a week
The week ending 19 September originally printed 197,000. In last Thursday’s release it is 198,000, revised up by 1,000. So the issuer’s own release for the week ending 26 September contains two instances of 197,000: the new advance figure, and the superseded original for the week before it.
That matters for anyone whose event window or filter keys on a surprise threshold. The consensus-to-prior gap today is 3,000 claims, or 1.52 per cent of the prior. The routine weekly revision to the figure the gap is measured against is 1,000 — a third of it. Put plainly: a third of the distance between what the market expects and what it last saw will be quietly rewritten next week, and your post-trade analysis will be measuring a surprise against a number that no longer exists. We have made this mistake in print ourselves on a Japanese wage series, where five fields of the prior month were restated under the current month’s headline and three superlatives were left attached to numbers that had been withdrawn. This is the same shape, at a weekly cadence instead of a monthly one, and the correct response is to store the vintage alongside the value.
We went looking for a vendor fault and found our own reading of a widget
This is worth recording because the alternative was publishing something false. Our first fetch of a major calendar’s claims page came back reporting the consensus as 197,000 — which is exactly last week’s actual print, and therefore exactly the prior-in-the-forecast-field error this desk has documented in four other vendor tables this month. It would also have been a same-publisher contradiction, because that publisher’s own analysis desk wrote 200,000 in a piece filed at 06:16 today. Two surfaces, one house, a 3,000-claim disagreement. We wanted that article.
So we re-read the page with the cache defeated, and asked only one question: what is the exact label-and-value string? It is “200KConsensus197KPrevious”. The values precede their labels. The consensus is 200,000 and the previous is 197,000, which is correct, agrees with both other readers, and agrees with the publisher’s own analysis desk. The error was entirely ours — a summariser reading a widget’s element order and guessing the pairing.
The operational rule, and it is the second time this week it has saved us: when a fetch hands you a vendor error you were already expecting, that is the moment the page gets read again, verbatim, with the cache busted. A source’s reputation in your own notes is a prior, and a prior that confirms your story is the one most likely to be your own thumb on the scale. We had a named publisher ready to accuse. We have no publisher to accuse.
Where today’s exposure actually sits, and it is not 12:30
The day’s schedule, all times UTC: the European Central Bank’s September minutes at 11:30; initial and continuing claims at 12:30, consensus 200,000 and 1.7 million against 197,000 and 1.701 million; revised wholesale inventories at 14:00, which is also the New York option cut; natural-gas storage at 14:30; a four-week bill at 15:30; and a 22-billion-dollar thirty-year bond reopening at 17:00. Governor Waller already spoke, before European lunch, and said more increases are likely with flexibility on the pace. There is no tier-one American data release today.
If you run a news filter that blocks around every scheduled United States release, today it blocks 12:30 and probably not 17:00, because auctions rarely carry a consensus and therefore rarely carry a surprise field for a filter to key on. That is backwards. A claims print whose consensus equals its own trailing average, published alongside a continuing-claims consensus one thousandth away from the prior, is the lowest-information American event of the week. A thirty-year reopening into a ten-year at 5.341 per cent has confirmed supply, no consensus and, last month, an allocation that left primary dealers with 2.2 per cent of the paper. Our companion piece this slot takes that apart.
The pair was 158.28 at 10:00, 70.0 pips up from the Tokyo low near 157.58 and 28.0 pips above the 158.00 strike expiring at the 14:00 cut, whose size the publisher that lists it does not disclose. Unknown size at 28 pips is not a magnet; it is an unknown. The honest schedule for today puts the small window at 12:30 and the large one at 17:00, and most calendars will tell you the opposite.
What this does not tell you
It does not tell you the four individual weeks behind the 200,000 average. We have the average from the issuer and the 197,000 and 198,000 firmly; the 206 and 199 are our reconstruction to make the arithmetic close, built partly on a vendor’s loosely dated weekly figures. Treat the decomposition as ours and the average as the issuer’s.
It does not tell you what today’s print will be, and the whole point of the piece is that a 200,000 print would tell you nothing either. Nor does it establish how large a miss has to be to move the dollar; we have not measured that and are not asserting a threshold.
It does not establish that a consensus equalling the four-week average is normal rather than today’s coincidence. One week is one observation. We have not checked how often the two coincide, and until we do, “the consensus is the trailing mean” is a statement about 8 October, not about weekly claims.
And it does not tell you that the 17:00 auction will move the yen more than 12:30 will. We are making a claim about information content and about which window your filter notices, not a forecast about realised volatility. A claims number at 215,000 would move more than any auction result; we are saying the market does not expect one, and that your schedule should reflect where the unknowns are rather than where the surprise fields are.
