Three Clocks Sit Behind One Short Interest Field Open a US equity on almost any broker page and you will find a line labeled short interest: a share count, often a percentage of float beside it, sometimes a days-to-cover figure. The field sits next to the last sale and the session volume, both of which update in seconds, and it quietly inherits their air of currency. It should not. The short interest field is a photograph of a settlement date that has already passed, developed and released on a calendar FINRA publishes a year in advance. On September 22, 2026, the most recent FINRA short interest figure a US screen can be showing comes from the August 31 reporting settlement date. Member firms filed it by 6:00 p.m. Eastern on September 2. FINRA released it on September 10. The next figure, capturing September 15, does not reach the public until September 24. So for twelve calendar days, the field labeled short interest has been describing the last Monday in August. That is not a ...
Inside a 206,000 Jobless Claims Print, the States Counted 170,626
On September 3, 2026 the Employment and Training Administration published a seasonally adjusted initial claims figure of 206,000 for the week ending August 29. State unemployment offices had counted 170,626 actual claims that week. The difference is neither an error nor a revision. It is a division: 170,626 divided by 0.829, the seasonal factor assigned to that calendar week, gives 205,821, and the release prints that to the nearest thousand.
The divisor is not a mystery number. The Bureau of Labor Statistics publishes the whole schedule ahead of time, in a file titled “Current seasonal factors for adjusting weekly initial and continued claims for unemployment insurance, January 2021–March 2027.” Every remaining 2026 divisor is already published. Across the 52 week-ending dates in 2026 the initial claims factor runs from 0.777 to 1.644, a spread of 2.12 to one, and 33 of the 52 sit below 1.00.
Two things follow, both arithmetic. The factor for any past week can be recovered from the two levels the release already prints, and the week-over-week change in the headline can be rebuilt from two other figures on the same page. ETA and BLS publish the inputs; the factor comparisons, the identity and the residuals below are this article’s calculation.
Two Agencies Share One Number
The weekly report is a joint product, and the division of labour is stated plainly. BLS writes that “ETA collects the weekly UI claims reported by each state’s unemployment insurance program offices and publishes a weekly news release.” On the adjustment itself: “ETA uses the set of seasonal factors BLS provides annually and applies them to the unadjusted data from the regular UI program during that year.”
So the agency that prints the headline does not build the divisor. For 2026 the adjustment is multiplicative: BLS states that “Before and after the highly volatile portion of the pandemic, the series are adjusted multiplicatively and in scope for regular revision.” Multiplicative means the relationship is a quotient — the adjusted level is the unadjusted count divided by the week’s factor. Everything below rests on that form, and the exception to it is set out near the end.
The Divisor Is Published a Year Ahead
The shape of the 2026 factor curve is not subtle. It peaks at 1.644 in the week ending January 10. It reaches 1.176 in the week ending July 11. Its low is 0.777 in the week ending September 12. The trough matters as much as the peak, because a factor below 1.00 scales a raw count up on the way to the headline.
Send the same raw count through the two extremes. Two hundred thousand actual claims filed in the week ending September 12 would print near 257,400; the identical 200,000 filed in the week ending January 10 would print near 121,700. The gap is about 135,700 claims, produced entirely by the calendar.
A live version of the same point: the week ending August 29 carried a factor of 0.829 and produced 206,000 from 170,626 raw claims. The week ending September 12 carries 0.777. If exactly 170,626 people filed in that week too, the headline would land near 219,600 against the 205,821 behind that 206,000 print — a rise of about 13,800 on identical filings.
Recovering the Factor From a Release
Because the adjustment is multiplicative, the factor is the ratio of the two levels every release prints. Divide the unadjusted advance count by the seasonally adjusted advance figure and the divisor falls out. Applied to twelve 2026 release weeks, the recovered values sit almost exactly on the published curve.
In the nine sampled weeks from March 14 onward, the recovered factor differs from the BLS published factor by at most 0.0019, and by 0.0011 on average. That is about what the rounding allows: the adjusted level is printed to the nearest thousand, so on a 200,000 base the ratio carries roughly a quarter of a percent of slack either way.
Two of the three January weeks — those ending January 3 and January 31 — sit outside that range, and the reason is on the record. The week ending January 31 recovers a factor of 1.089 against a published 1.100, a gap of 0.0106, more than five times the worst post-March miss. Those releases predate the annual revision. BLS states that “The current factors most recently were revised from 2021 forward in conjunction with the news release issued on March 19, 2026,” and the March 19 release says that “The seasonal adjustment factors used for the UI Weekly Claims data from 2021 forward, along with the resulting seasonally adjusted values for initial claims and continuing claims, have been revised.” The January headlines were built with the previous vintage of factors, which is not the vintage in the current file.
Two Published Figures Rebuild the Headline Change
Every release carries a pair of sentences that are easy to skip. The first gives the change in the unadjusted count from the previous week. The second, for the week ending August 29, reads: “The seasonal factors had expected a decrease of 1,226 (or -0.7 percent) from the previous week.” Call the first figure A and the second E.
The release does not define E. Treating it as the change that would have left the adjusted level flat is this article’s reading, and it has a testable consequence: the change in the headline should equal the amount by which the raw data missed the expectation, divided by the week’s factor. That is, (A minus E) divided by f.
Nine of the twelve are listed in full below, so the reading can be checked rather than asserted. Surprise, identity result, published change, gap:
Week ending January 3 — surprise +11,891, identity +8,221, published +8,000, gap 221.
Week ending January 17 — surprise +2,048, identity +1,578, published +1,000, gap 578.
Week ending March 14 — surprise −8,131, identity −8,756, published −8,000, gap 756.
Week ending March 28 — surprise −8,904, identity −9,730, published −9,000, gap 730.
Week ending April 25 — surprise −24,944, identity −26,225, published −26,000, gap 225.
Week ending June 20 — surprise −11,917, identity −12,370, published −12,000, gap 370.
Week ending July 11 — surprise −9,740, identity −8,275, published −8,000, gap 275.
Week ending August 8 — surprise +8,027, identity +8,976, published +9,000, gap 24.
Week ending August 29 — surprise +1,256, identity +1,516, published +2,000, gap 484.
These nine gaps run from 24 to 756 and average 407; the three sampled weeks not listed land at 168, 631 and 214. None of the twelve reaches 1,000, which is the width of the band rounding alone can produce: the published change is the difference of two levels each rounded to the nearest thousand. Twelve for twelve inside that band is the evidence for the reading of E, and also the reason to stop there. The identity reproduces the published change to the precision the release itself carries, and no further.
The Week Raw Claims Fell 21.6 Percent and the Headline Rose
The week ending January 17 shows why the decomposition is worth doing. The release of January 22 states that “The advance number of actual initial claims under state programs, unadjusted, totaled 259,552 in the week ending January 17, a decrease of 71,411 (or -21.6 percent) from the previous week.”
The same page states that “The seasonal factors had expected a decrease of 73,459 (or -22.2 percent) from the previous week.” The model had penciled in a slightly larger collapse, so the surprise is positive: the raw data fell 2,048 short of the expected fall. Dividing by the recovered factor of 1.298 gives about +1,578, and the release duly reports that “In the week ending January 17, the advance figure for seasonally adjusted initial claims was 200,000, an increase of 1,000 from the previous week’s revised level.” Raw claims down 21.6 percent, headline up, both from one page.
What This Changes at 8:30 on Thursday
Three shifts follow, none requiring a view on the labour market. The surprise against the seasonal factors is a different quantity from the surprise against consensus: consensus forecasts the adjusted number, while A and E are facts about the unadjusted one. A print can beat consensus while the raw data ran hot against the model, or the reverse.
The four-week moving average inherits every factor inside it. The 207,250 average reported alongside the 206,000 print averages four separately scaled numbers, not four comparable counts.
And the divisor for the next print is already known. Converting a raw expectation into a headline expectation is a lookup, not an edge, but it is the difference between reading a data point and reading a quotient.
What Would Invalidate This
The multiplicative form is not universal. BLS states that “For the most volatile portion of the pandemic for these series, covering the week ending March 21, 2020, through the week ending June 19, 2021, the series are additively adjusted.”
Vintages do not mix. Factors are revised once a year from 2021 forward except inside the additive window, and the current file runs only through March 2027. A factor recovered from a January 2026 release and one read from the current file describe the same week with different numbers, as the 0.0106 gap shows.
Precision stops at the rounding. Adjusted levels are printed to the nearest thousand. The recovered factor is good to roughly a quarter of a percent and the identity to within the thousand-claim band.
Advance figures are not final. Both the adjusted and the unadjusted advance numbers are revised in the following week’s release, so anything computed from an advance pair is provisional by construction.
Continued claims are a separate series. The insured unemployment figures carry their own factors and their own revision history, and nothing here transfers to them without redoing the work.
A factor is not a forecast of layoffs. A high divisor week is not a bad week. The factor describes a regular annual pattern, and says nothing about whether this year’s will hold.
Concrete Framework
A checklist for reading the Thursday release as a quotient rather than a count.
Look up the factor before the release. It sits in the current BLS factor file, keyed by week-ending date, and does not change unless an annual revision intervenes.
Convert the raw expectation, not the headline one. Divide the raw count you expect by that factor; that is the number the release will print.
Read A and E before reading the headline. The actual change and the expected change are two sentences apart, and their difference is the whole of the week-over-week move.
Check the direction of both. The January 17 case is the test: raw claims can fall hard while the headline rises, and the reverse holds equally.
Recover the factor as a check. Unadjusted divided by adjusted should land within about 0.002 of the published factor; a wider gap points to a vintage boundary.
Mark the annual revision. The 2026 revision landed with the March 19 release and rewrote the adjusted series from 2021 forward, so stored history from before that date is a different series.
This article is for information only, is not investment advice, and does not recommend buying or selling any security. Figures are drawn from the Bureau of Labor Statistics current seasonal factor file for weekly UI claims and its explanatory page, and from Employment and Training Administration UI Weekly Claims news releases for 2026, as available on September 6, 2026.
Comments
Post a Comment