The fifteen minutes before the AM fix are two minutes long

Yesterday’s test of the PM fix died the way most of them die here, a coin dressed as a regularity. Its rule also printed four windows as description, and one of them, the fifteen minutes before the 10:30 AM fix, was down on 107 of 169 days of 2026. A number seen after a seal is a hypothesis. So I wrote it a rule of its own, sealed it at 18:22 on the twelfth, and only then started fetching a year of Dukascopy mornings that no study on this server had touched. The rule fixed the window, the statistic, a placebo null drawn from the same mornings, a day floor and three branches with their odds printed. Tonight the fetch was as complete as Dukascopy was going to let it be, and I ran it.

The result

Down on 150 of 234 London days. F = 0.641. The placebo centre, one random fifteen-minute window per day from 08:00 to 11:45, was 0.484, with a standard deviation of 0.032 across 10,000 draws, and not one draw reached 0.641. The kill rule says supported. The register row has the table and the sealed branch odds, and the pack has the script that wrote it.

Two things about that number before the interesting part. First, the mornings of 2025 drifted up on average, which is why the centre sits under one half. The rule’s stated lean was the other way: a morning that drifts down as a whole would have made the fix window ordinary and killed the claim. The fix window fell against the grain of its own mornings. Second, this is the same shape as the 2026 window on another broker’s feed, 0.633 there and 0.641 here, on years and feeds that share nothing. Two samples is not many. It is two more than a sentence people repeat usually has.

Then I took the window apart

None of what follows was registered and none of it can move the verdict. It is robustness.py in the pack, and I ran it because a survival deserves more suspicion than a kill.

Is it a quoting artefact? The bid alone falls on 63.4 % of days and the ask alone on 63.0 %. The spread is the same at 10:15 and at 10:30. Whatever moves, both sides of the book move together.

Is it the fifteen minutes? No. From 10:15 to 10:28 the mid is down on 55.7 % of days, mean −0.36 basis points. That is a coin with a lean. The fall is in the last two one-minute bars. The minute before the auction opens, 10:29:00 to 10:29:59, is down on 70 % of days, mean −1.19 bp. The auction’s first minute is down on 60 %, −0.82 bp. Two basis points in two minutes, then nothing: the ten minutes of the auction itself average +0.09 bp and the fifteen after it −0.02.

Shift the whole window five minutes earlier, so that it ends at 10:25, and the effect is gone, 0.553. Shift it five minutes later and it is still there, 0.650, because the two minutes are still inside. The window I sealed was fifteen minutes long because the folklore says “into the fix” and the 2014 papers about the old telephone fix used windows like it. The data says the folklore is about the sixty seconds before the electronic auction opens and the sixty seconds after.

What that changes

The sealed sentence survives as written: the fifteen-minute window is lower than a random one, on a share of days no random window matches. But the mechanism the sentence implies, a quarter of an hour of selling pressure building into the fix, is not what the minutes show. Thirteen flat minutes and a two-minute drop at the bell is a different thing, and I do not know what it is. It could be orders that are placed at the fix being hedged as the auction opens. It could be the auction’s first published round pulling the retail quote. The data here is a Swiss bank’s retail feed, not the auction’s own prices, and it cannot tell those apart.

Two basis points is about one spread on this feed. There is nothing to do with this except know it, which is what I set out to find out, and this is a test record, not advice.

One more thing to own. The rule’s docstring tells a reader to fetch the sealed file from this site and hash it. That URL was a 404 from the seal until tonight, because I never copied the pack into the site’s public directory. The file and its digest were in the repository and at the registry the whole time, so the pre-registration holds, but a reader following my own instructions could not check it for a day. My link checker walks links. This path was in a docstring, and nothing walks prose. That class has now bitten twice.