Do round numbers work as gold support and resistance?

The claim being tested

Look at any “XAUUSD support and resistance” page and you will find a list of prices. They are stated as fact, they expire within a day, and they cluster heavily on multiples of $50 and $100. The implied claim is that these prices are special, and that price is more likely to turn there than elsewhere.

That claim is testable, and as far as we can find nobody has published the test. This is it.

How it was measured

70,952 five-minute XAUUSD bars from a live MetaTrader 5 feed, 2025-07-31 to 2026-08-04, spanning a price range of $3,281.59 to $5,595.35.

An approach is the first bar where price trades into a band around a level, having been outside it. From there we look forward one hour and ask a single question: did price reject the level, ending back on the side it came from, or break it, closing through and staying through. Approaches within 24 bars of the previous one are ignored, so price oscillating around a level does not contribute dozens of correlated readings.

The part that decides everything is the control. For every round level we also test two deliberately unremarkable ones, offset 30% and 70% into the same increment — prices nobody draws on a chart. If roundness carries information, round levels must reject more often than these. If they do not, the levels everyone publishes are decoration.

The result

Round levels rejected price 56.6% of the time. Levels chosen specifically because they are unremarkable rejected it 56.1% of the time.

The difference is +0.45 percentage points, with a 95% confidence interval running from −0.65 to +1.55. That interval contains zero, on a sample of 35,002 decided approaches. Taken individually, none of the four level sizes produced an edge distinguishable from chance either.

This is not a case of the sample being too small to tell. It is large enough that a real effect of even one percentage point would have shown up. The honest reading is that roundness carries no measurable information about where gold turns.

One number in that table is misleading, and it is ours

Rejection climbs steadily with level size — 54.2% at $10 up to 69.8% at $100. It is tempting to read that as bigger levels being stronger. It is not. It is an artefact of our own method.

The band scales with the increment, so at $100 price must travel $5 to count as a break, against $0.50 at $10. Wider band, more moves that fail to clear it, higher rejection rate. Round and control share the same band at each size, so the comparison down each row is sound — but the trend across rows measures our band width, not the market. We are pointing it out because someone would otherwise quote it as a finding, and it is not one.

Why round numbers feel like they work

Because price reverses at most levels. A 56% rejection rate at an arbitrary price is not a special property of that price — it is what an oscillating market does over an hour. Every level you draw will appear to work slightly more often than it fails.

What makes round numbers feel different is recall. A reversal at $2,700 is memorable and gets annotated. An identical reversal at $2,683.50 is not, because nobody drew a line there. The reversals are equally common; only the noticing is selective.

That is worth separating from a stronger claim we did not test. Levels taken from actual market structure — a prior swing high, a session extreme, the edge of an overnight range — are a different proposition and were not measured here. We have tested session ranges separately and found follow-through of 51.7%, which is also close to a coin flip. What this study rules out specifically is choosing a level because it ends in two zeros.

What to do instead

If the level is not what decides the trade, the stop distance is. We measured that separately: a $2 stop on gold is hit 81% of the time within an hour by ordinary movement, and gold’s median daily range is around $76. Placing a stop just beyond a round number, on the theory that the number will hold, puts it exactly where noise reaches — see the cost of scalping gold.

The practical replacement for a level list is a size calculation. Decide the stop from volatility, then size the position so that stop is affordable. The stop loss calculator does it against your own account, and support and resistance covers what the concept means when it is drawn from structure rather than from arithmetic.

Method, limits, and the data

One symbol, one year, one broker feed, one horizon. A different holding period would produce different absolute rejection rates — a longer horizon lets more levels break and pushes every number down — but it would apply equally to round and control levels, which is the comparison the conclusion rests on.

The band is the judgement call. Too tight and genuine touches are missed; too wide and unrelated price action counts as an approach. We used 5% of the increment either side and applied it identically to both groups. The script is research-round-numbers.py, it runs on the same published bar archive as our other studies, and the full output including break counts is in the JSON behind this page.

Underlying data, method and licence: the XAUUSD archive, published under CC BY 4.0. This study has its own record too — DOI 10.5281/zenodo.22637578, mirrored on Hugging Face — so the counts behind every figure above are downloadable. If you disagree with the band, the horizon or the control choice, the inputs are public and the result is yours to overturn.

Common questions