The PM fix is a coin

Between 2014 and 2016, “gold gets sold into the fix” stopped being a thing traders muttered and became a thing with regressions behind it. Papers on the old London gold fixing, when five banks settled the price by telephone twice a day, found the minutes before the 3 p.m. fix moving down far more often than up. There was a class action. In March 2015 the fix became an electronic auction, and the sentence went on being said about the new one, as a reason to be short at ten to three.

That is a claim with a sign and a clock. It can be tested on any year of minute bars. I have eight months of 2026 gold from one broker, so I tested it, and the test was sealed before I read a single return.

What the sentence asserts, and what it does not

“Sold into the fix” does not mean “gold goes down in the afternoon”. It means the fix window is different from the rest of the day. So the null is not “zero” but “any other fifteen minutes”: for each of the 169 days, pick a start minute at random from 08:00 to 16:15 London, take the fifteen-minute return, and ask on what share of days it is negative. Ten thousand such draws give the distribution of that share when nothing about the fix is special, on these exact days, with this year’s drift and each day’s own volatility kept. The fix window’s share is then read against it.

The statistic is a sign, not a mean, and that was a decision the rule forced on me before the data. The first draft used the mean move with a floor of three basis points, the smallest move I would call worth the sentence. Then I ran the draft’s own branch table, the one that says, for each assumed truth, how often the rule reaches each verdict. At any volatility above about twelve basis points per fifteen minutes the killed branch was dead: a null result would have come out “inconclusive” by construction, whatever the data said. The observed volatility was thirty-one. That draft would have run to completion, printed “inconclusive”, and looked like modesty. It was a rule that could not lose, and a rule that cannot lose is a silence dressed as a test. A fraction of down days has no scale, so its floor can be set as a rule of thumb (two days in three) and the branch table shows every verdict reachable. That is the one I sealed.

What happened

A coin. The test could have seen a regularity that held on two days in three and it saw eighty-eight of a hundred and sixty-nine. The window is busier than a random one, twice the median absolute move, which the half-hour study found last week, and busy is not the claim.

The rule also listed four descriptive windows to print, with the sentence “none of these can move the verdict”, and one of them is going to tempt people. The AM fix’s run-in, 10:15 to 10:30 London, closed lower on 107 days of 169, a fraction of 0.633 against a centre of 0.498, with a p-value of 0.0002 on its own draws and a mean of minus four basis points. That is the folklore’s exact shape, in the window I did not register. I am printing it because the rule said to, and I am not calling it anything, because it was found after the seal, in one of four windows I chose to look at, on one broker, in one year. Its job is to be the next rule: sealed, on a second price feed, before a word.

The pack, the rule, the seal and the per-day table are at /research/gold-pm-fix/, and the register row is on the experiments page.

This is a test record, not advice.