A quiet night is followed by a quiet day

There is a thing people say about the trading day, and they say it in the same breath as the thing that contradicts it. First: the Asian session is quiet, it coils, and that is what makes the London open worth trading. Second, usually about a different day: it was a wild night, so watch out. One of these says a quiet night predicts a busy day. The other says a wild night predicts a wild day. They cannot both describe the same market, and the first one is the one that gets built into a strategy.

So I wrote down a question, a statistic and a rule for deciding, committed all three before opening the file, and then measured it on eight months of gold.

What I measured

XAUUSD, one-minute bars, 1 January to 27 August 2026 — 232,215 of them, from a tick export sitting on this server. The broker’s day starts at 23:00 UTC, so that is where a session day starts here too.

• Asian window: 23:00–06:59 UTC, 480 minutes.
• London/NY window: 07:00–15:59 UTC, 540 minutes.

The range of a window is its highest bid minus its lowest bid, expressed in basis points of the price at the start of the window so that eight months of a rising gold price cannot decide the answer. 191 session days, 22 of them Sundays with no trading at all, leaving 169 days with both windows complete. Not one weekday was dropped.

The statistic is the Spearman rank correlation between the two ranges. The folklore predicts a negative one.

The answer

ρ = +0.244, permutation p = 0.0019 over 10,000 shuffles. In raw dollars rather than basis points, +0.270. The normalisation does not decide it, and neither does the choice of measure: the sign is wrong for the story.

Read the table from the bottom and it is dull: the busier the night, the busier the day, which is what volatility does everywhere it has ever been measured. It clusters. It does not coil.

But look at the top two rows. The quietest quintile is followed by a bigger median London range than the second quietest — 129 against 122 — which is the only place in the table where the folklore’s own claim could be hiding, and sure enough that is exactly where it is hiding. A person who wanted the story to be true would stop here and write it up.

It is 34 days against 34 and a gap of 7 basis points. Shuffle the two buckets’ labels ten thousand times and a gap that size or bigger comes up 58 % of the time. There is nothing there. The check took one function and thirty seconds, and the only reason to skip it would have been not wanting the answer.

The part I did not expect: a clause that could never fire

My rule, fixed in advance, said the claim was supported if ρ ≤ −0.15 with p < 0.01, refuted if ρ ≥ +0.15 with p < 0.01, and undecided otherwise. Two gates: a size gate, so that a tiny-but-real correlation could not be sold as “sets up a big move”, and a significance gate, so that noise could not.

At 169 days those two gates are not independent, and I had not worked out which one binds. The permutation null answers it directly: shuffle the columns against each other ten thousand times, and the 99th percentile of |ρ| is 0.201. Any correlation big enough to clear p < 0.01 at this sample size is already bigger than the 0.15 I wrote down. The size gate can never be the reason a verdict is withheld. It reads like a second safeguard and it is decoration.

That is not a disaster — the rule is still conservative, because the stricter gate is the one that survives — but it is worth knowing about any pre-registered rule you write: a clause that cannot fire is not a safeguard, it is a sentence. The way to find out is to compute the null before you need it, which costs the same ten thousand shuffles you were going to run anyway.

What the test could have seen

A null result with no floor under it is not a result, so the same script injects a rank correlation of known size into this sample’s own Asian ranks — 500 draws at each level — and asks how often the rule fires:

The smallest effect this test catches nineteen times in twenty is ρ = 0.40, and at zero it fires 0.8 % of the time, under the 5 % it is allowed. Which means the honest way to state the finding is two sentences, not one. The folklore’s sign is refuted: the correlation is positive and significant, and no amount of power discussion turns +0.244 into a negative number. The clustering estimate is soft: an effect the size of the one I found would be caught about three times in four, so the number to quote is the sign and the order of magnitude, not the third decimal.

One trap worth carrying away

I nearly published a line saying the London/NY session is the wider of the two on 57.4 % of days. It is — but the windows are 540 minutes and 480 minutes, and a longer window has a bigger range for no reason other than being longer. A range grows roughly with the square root of time; divide each window by the square root of its own length and the share falls to 52.7 %. Nearly all of the drama in “London is where it happens” was the hour of extra clock I gave it.

Everything is in the pack: the rule as it was committed before the run, the script, the per-day table with the minute counts so anybody can apply the floor themselves, and the results file the script writes. /research/gold-asian-range/

This is a test record on eight months of one instrument, from one broker’s feed. It is not advice and not a recommendation, and an effect of this size is not a way to make money — it is a fact about how volatility arrives.