When does gold actually move?

This is a test record, not advice. Do not trade on it.

Search for “best time to trade gold” and every page says the same two things: the market is most active when London and New York are both open, roughly 08:00 to 12:00 New York time, and quiet during the Asian session; and spreads widen around the daily rollover and in Asian hours, so those hours cost more. Never with a number. I have eight months of tick data for one broker’s gold contract on this server, so I wrote the claims down as testable statements, decided in advance what would count as confirming them, and then ran the numbers. The claims, the definitions and the verdict rules are in research/gold-hours/README.md; the scripts and the full tables are next to it.

The data: 232,215 one-minute bid/ask bars from 2026-01-01 to 2026-08-27, 170 weekdays, built from the broker’s own tick export. One broker’s CFD, not the futures market, and one year that happened to include a crash. The clock is New York’s, because the events that matter most are anchored there and London’s open stays at 03:00 New York for all but three weeks of the year.

The overlap claim, mostly right

The four hours from 08:00 to 12:00 New York are one sixth of the day. Averaged over all days, they carry 31.2% of the day’s variance, nearly twice their share of the clock, and the busiest half hour of the day lies inside them. My rule said “confirmed” needed more than 33%, so by the rule I wrote before looking, the verdict is partly, and I am leaving it there rather than moving the bar after the fact. Measured on the median day instead of the mean, the same window carries 35%. Either way the shape is not in doubt: the chart has one large hump and it is in that window.

What the guides get wrong is the peak. It is not the data release at 08:30. On the median day the busiest half hour is 09:30, the New York equities open and the first half hour of the main COMEX session, at about 16 USD/oz of movement, followed by 10:00 at 14. The 08:30 half hour is fourth on the mean and seventh on the median day.

The 08:30 minute is another story. It is the largest single minute of the day by root-mean-square movement, 10.6 USD, but its median range is only 4.6 USD. Most days nothing much happens at 08:30; on payroll and inflation days the price moves 50 or 60 dollars in sixty seconds. The minute with the largest median range is 10:00, at 6.0 USD, the London afternoon fix, which moves the price every single day and rarely by a lot. Those are two different kinds of busy, and a strategy has to know which one it is exposed to.

The Asian claim, half right

The quietest half hours on the median day are 16:00 and 16:30, the hour before the broker’s break, at about 5 USD, then 18:30 just after it, then 23:30 and 00:00. So the quiet hours exist, and the late evening in New York, which is midday in Asia, is one of them. That half of the claim holds under the rule I set.

But “the Asian session is quiet” hides a second hump. The half hour at 21:00 New York, which is 09:00 in Shanghai when the Shanghai Gold Exchange opens, is the fourth busiest of the whole day on the median day, at 11.8 USD, and its first minute has the second-largest median range of any minute in the day, 5.9 USD, just behind the London fix. The Tokyo open at 20:00 is a smaller bump. Both beat every half hour of the European morning: the London open at 03:00 New York, which the guides describe as the start of the real day, is 8.3 USD, less than the Shanghai open, and no more than the hours around it. On this data the price of gold is set in New York first and Shanghai second, and London’s morning is a long, even middle.

The spread claim, wrong for this broker

I expected this to be the easy one and it turned into the clearest result. This broker’s median spread is 0.09 USD in every one of the 48 half hours. The 90th percentile is 0.11 or 0.12 in every half hour but the one before the close. Asian hours, the overlap, the European morning, the dead hour before the close: the same number. The claim that Asian hours are expensive is refuted for this broker as plainly as anything can be, because there is nothing to compare.

The rollover half of the claim survives in a much smaller form than it is told. The spread goes to 0.50 in exactly one minute, 16:59, the last minute before the break, every day, and in the first two minutes after the reopen at 18:00 on about one day in ten. Not the hour around rollover. Three minutes at most.

My pre-registered rule had a flaw here that I want on the record: I defined “the rollover hour” as 17:00 to 18:00 New York, and on this broker that hour is the break, present in the data only during the three weeks in March when the American and European clocks disagreed. The rule as written returned a ratio of 1.00 on fifteen days, which is a measurement of nothing. The minute-level check is the amendment, added after seeing that, and marked as such in the results file.

What the flat spread means in practice is that the cost of trading this contract varies by hour only because the movement does. One spread is 2% of a typical minute’s range at 09:30 and 10:00, and 7 or 8% in the hour before the close. The Asian hours are not expensive; they are slow, and a slow hour is expensive only for a rule that needs the price to go somewhere.

Two things about the measuring

The whole-day figure is 87.5 USD/oz on the average weekday and 63.3 on the median one. The gap between those two numbers is 2026: on 29 January the price fell 125 dollars in a single minute at 10:27 New York, and that one minute holds 87% of all the variance the 10:27 minute has accumulated in eight months. Rank the half hours by mean variance and the 10:00 bin comes first, with a third of its total from that one day; rank them by the median day and 09:30 comes first, and the crash is invisible. I ran both and the entry uses the median where it says “typical”, because that is the day you actually get. The mean is in the tables for anyone who wants the day you fear.

And the weekday table, which I computed because everybody asks: Thursday 97, Monday 91, Wednesday 85, Tuesday 82, Friday 80 USD/oz on average. Thirty-four days each, single days moving the mean by ten dollars. I would not call that an ordering.

Not measured: whether any of this is stable across years, brokers or regimes. It is one broker, one contract, one year that ran from 4,330 to 4,600 with a crash in the middle. The 21:00 hump and the flat spread are the two things I would check first on another dataset, because they are the two things that contradict what is repeated.

Addendum, the same afternoon: a second source

The server has been slowly downloading Dukascopy’s public tick feed for the same instrument, a different quote source, and by this afternoon it covered January to mid-June, 119 weekdays. I ran the same script on it unchanged; the table is dukascopy-bins.csv next to the others.

The movement profile is the same picture. Busiest half hours on the median day: 09:30, 10:00, 10:30, then 21:00; quietest: 16:30, 23:30, 16:00. The overlap share is 31.9%. The Shanghai hump is not this broker’s artefact.

The spread is not the same picture, and this is the correction. Dukascopy’s median spread is about seven times the broker’s, 0.62 to 0.87 USD depending on the half hour, and it does move with the clock: narrowest in the European morning around 04:30 to 06:30 at 0.62, about 0.67 to 0.71 through New York, and widest in the hour after the reopen, 0.87 at 18:00 and 0.83 at 19:00, where the 90th percentile reaches 2.10 against 0.85 to 1.00 for most of the day. So on the raw feed the second half of the spread claim is partly true: the Asian evening and the reopen are more expensive, by about a third at the median and by double in the tail. Not “blown out”, but real. What the first source showed was not the market’s spread at all; it was a fixed spread the broker sets, which erases the pattern and replaces it with one 0.50 minute at the close. Whether that fixed nine cents comes with a commission I cannot see in the data. The lesson I take is the one I should have stated before writing the section: a broker’s spread is a price list, and a claim about “spreads” needs a source that is not a price list.

Scripts and tables: research/gold-hours/ in the repository, profile.py for the numbers, chart.py for the figure, bins.csv for all 48 half hours and minutes.csv for all 1,440 minutes.