Do round numbers matter in forex?

On the evidence I have, no — and this one is worth reading carefully, because the original finding was mine, it was strong, and I spent months treating it as the most promising structural feature on the board before it dissolved.

The test that killed it is one anybody can run, and it is not a significance test. I call the failure mode Window Luck: an effect that looks real on the stretch of history where it was found, and scatters around zero when the same test is run on other stretches of the same data.

The finding I had

Round-50 levels swept measurably more often than other levels. Measured over a thirteen-month window: z = −2.59.

That is a conventional, respectable, publishable result. It was also a-priori — you can identify a round number before anything happens at it, which is exactly what a tradable structural feature requires and what almost nothing else I had managed. When my engine’s other candidates were being eliminated one by one, this was the survivor I recorded as still live.

The test that killed it

Run the identical test across twelve roughly sixteen-month windows of the same sixteen-year history.

• Mean z across windows: +0.08
• Windows reaching significance: 0 of 12
• Individual windows: some as high as +2.5

The last line is the one that settles it. Plus 2.5 is the opposite direction — round levels sweeping less than others. If the effect were real and merely noisy, windows would disagree about magnitude while agreeing about sign. These disagree about sign.

That is what noise looks like when you slice it twelve ways. The original z of −2.59 was one of those twelve slices, found first and believed because it was found first.

Why this is not a sample-size objection

The natural response is that the windows were too short.

They were longer. Roughly sixteen months each, against the thirteen months that produced the original result. Every replication had strictly more data than the finding it was checking.

And the failure mode is diagnostic. A power problem gives you wide confidence intervals around a consistent centre — the effect is there, you just cannot pin it down. This gave a scattered centre. No quantity of additional history repairs that, because the disagreement is between windows rather than within them.

The general lesson, which is worth more than the finding

Almost every trading idea is discovered on a particular stretch of data. That is not a flaw; it is how discovery works. The flaw is stopping there.

A significance test tells you whether the effect differs from zero on the data you found it in. It cannot tell you whether you would have found it had you looked somewhere else, and that is the question that decides whether an idea survives contact with next year.

The check costs almost nothing: split your history into windows, run the same test on each, and look at the sign before you look at the p-values. If the sign scatters, you have a property of your sample. Nothing about the original result — not its magnitude, not its p-value, not how sensible the mechanism sounds — distinguishes the two cases in advance. Only replication does.

Limits

One instrument, sixteen years, bar data, round-50 levels specifically. A single-instrument result establishes nothing about any other, and I have not tested whether round levels behave differently elsewhere.

The scope is also narrower than “levels do not matter”. This tests a claim about the price scale — that certain numeric values are special before anything happens at them. It does not test whether levels with recent history behave differently, and that separate effect is real: a level’s most recent outcome predicts its next one with an odds ratio near 3.5, out of sample.

Structure exists. It is sequential, not positional, and it is a general property of price series rather than evidence for anything about where round numbers sit.