Archive / Crypto & DeFi / E376
London gold auction order flow leaves no trace on a gold perpetual
Trading the prior day's move through the two London gold auction windows lost 8.8 bp and 8.7 bp net, about the 8 bp cost; the prior move did not predict the window (t 0.00 and -1.26 against 2.5).
Gold is priced twice a day in a London auction (the fix), and funds, central banks and jewellers place orders at that price. The belief is that these orders follow the previous day's move, so gold keeps going the same way through the auction. We tested it on a gold perpetual over about 220 London weekdays: trade the direction of the prior day's move for 45 minutes around each auction at 4 bp per side. The afternoon window returned -8.8 bp and the morning window -8.7 bp, the 8 bp cost plus a gross return near zero. The prior move explained none of it (t 0.00 and -1.26; 2.5 needed).
Orders placed at the London gold auction follow the previous day's price move, so gold keeps moving in the same direction through the auction window.
Why people believe it
People argue that gold exchange-traded funds, central banks and jewellers place orders at the auction price, and that fund flows chase recent returns, so the net auction order should be predictable from the prior day's move. Academic work (an academic study from 2014) found a drift into the fix before the auction became electronic in 2015; the test here asks whether anything of that kind remains and shows up on a perpetual contract.
Why it would persist The argument, as stated before testing
The balance of buy and sell orders in the auction is not public in real time, and traders in a perpetual contract are not participants in the auction, so the flow could stay unexploited.
For each London weekday we measured the gold price move from the previous weekday to just before the auction window. We then bought if that move was up and sold if it was down, and held for the 45 minutes from 15 minutes before the auction to 30 minutes after it, once for the afternoon auction and once for the morning auction. Costs were 4 bp on each side (1 bp = 0.01%). We ran the same thing at two ordinary times of day (12:30 and 17:00 London) to see whether the auction windows differed from any other window. The rule and its rejection criteria were fixed before the data were examined.
No separate pass line was recorded. The rejection criterion was written as the complement: the idea survives only if the slope t-statistic reaches 2.5 in at least one auction window and the net mean is above zero.
Slope t-statistic below 2.5 in both auction windows, or a net mean at or below zero.
4 bp per side (8 bp round trip), flat. No funding or extra spread was modelled for the 45-minute hold.
Placebo windows at 12:30 and 17:00 London with the same signal construction. The slope of the window return on the signal is estimated with heteroskedasticity-robust (HC1) standard errors.
Both pre-registered rejection criteria were met in both auction windows: the slope t-statistics (0.00 and -1.26) were below 2.5 and the net means (-8.8 and -8.7 bp) were at or below zero.
- Afternoon auction (14:45 to 15:30 London): the slope of the window return on the prior move was 0.000 (t 0.00), and following the sign of the prior move returned -8.8 bp net per trade, about 220 weekdays in the sample. Adding back the 8 bp round-trip cost, the gross return was -0.8 bp.
- Morning auction (10:15 to 11:00 London): the slope was -0.022 (t -1.26) and the follow-the-sign trade returned -8.7 bp net, a gross return of -0.7 bp. The slope is the wrong sign for the idea: a 100 bp prior move went with a -2.2 bp window move, not a significant one.
- Placebo windows at 12:30 and 17:00 London had slopes of about zero and net returns equal to the cost, so a gross return of about zero, the same as the auction windows.
- A positive slope means the window continued the prior day's direction, which the idea needs. The trade needed continuation worth more than the 8 bp round-trip cost; the measured continuation was 0.0 bp in the afternoon and -2.2 bp per 100 bp of prior move in the morning.
- The design-time estimate was that a mean window return of about 6 bp would have been detectable with about 222 weekdays and a 37.5 bp standard deviation, so the test could see an effect of the size the idea needs to beat an 8 bp cost, if it was at least about that large.
The prior move does not predict the auction window
t-statistic of the slope of the 45-minute window return on the prior day's move, for each London auction. The afternoon value is 0.00 and the morning value is -1.26, both far below the 2.5 rejection line.
Chart data (slope t-statistic)
| Slope t-statistic | |
|---|---|
| Afternoon auction (14:45 to 15:30) | 0 |
| Morning auction (10:15 to 11:00) | -1.26 |
Before costs there is nothing; the cost does the rest
Gross (net plus 8 bp)Net of 8 bp cost
Gross and net return of the follow-the-sign trade in each auction window. Gross is net plus the 8 bp round-trip cost: -0.8 bp in the afternoon and -0.7 bp in the morning. Net is -8.8 bp and -8.7 bp, close to minus the full cost.
Chart data (bp per trade)
| Gross (net plus 8 bp) | Net of 8 bp cost | |
|---|---|---|
| Afternoon auction | -0.8 | -8.8 |
| Morning auction | -0.7 | -8.7 |
Why it failed Interpretation · not established
No continuation of the prior day's move shows up in the auction windows on the perpetual contract; gross returns are within 1 bp of zero. One reading is that, since the auction became electronic in 2015, any order imbalance is absorbed in the auction itself and does not reach other markets as a predictable price move. Another is that the flow is not predictable from the prior day's return in the first place. The test cannot tell these apart.
Lesson
A flow that happens at a known clock time is not a price effect that a trader can read off the clock. For a rule that follows a signed prior move, gross returns near zero in the event windows and in ordinary windows alike mean the event has nothing that distinguishes it, and the flat cost then decides the net result entirely.
Scope of this result
Covers one gold perpetual contract on a large crypto derivatives venue, about 220 London weekdays from 2025-12 to 2026-10, two auction windows of 45 minutes, a signal of the previous weekday's move, and a flat 8 bp round-trip cost. UK public holidays were not removed from the sample.
A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.
1. Test design
Pre-registeredDataUniverseSignalExecutionCostsControlsSample splitPass criterionKill criterionPower2. Result
ObservationsInstrumentsNet mean95% intervalMedianWorstFinding 1Finding 2Finding 3Finding 4Finding 53. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade