MEASURED · 21 SEPTEMBER 2026

Should a gold stop be fixed, or scaled to volatility?

Everyone is told to size stops with ATR. Almost nobody is shown what it buys. We ran 55,678 trades with both rules risking the same average distance. The adaptive stop won on every measure — by less than you would expect, and not in the place you would expect.

Why the question exists

Gold does not move at a constant speed. On 15-minute bars, ATR(14) had a median of $2.08 per ounce across twelve years — but the calmest tenth of the time it sat near $0.90 and the fastest tenth above $7.04.

That is roughly an eight-to-one range. A stop that gives a trade sensible room in a quiet market is sitting inside the noise in a fast one. Whether that actually costs you anything is the measurable question.

Why most comparisons of this are rigged

A wider stop wins more often — it simply has more room before it is hit. So comparing “1.5 × ATR” against “$3” tells you which is wider, not which is better. Almost every comparison you will read has this problem.

Here the fixed stop is calibrated: set to the mean ATR-scaled distance across the same entries, so both rules risk the same average amount. In the headline run that came to $4.78. Whatever difference survives is adaptivity, not width.

Entries are deliberately worthless — one every two hours, long and short simultaneously so any trend cancels. There is no signal here. The stop rule is the only thing that varies, which is the only way to attribute the result to it. Method checks came first: the simulation returns exactly 50.0% on a random walk with a 1:1 target, and hand-built cases confirmed the long, short and losing paths resolve correctly.

The result at 1.5 × ATR

Win rate by volatility decile. Decile 1 is the calmest market the sample contains, decile 10 the fastest.

For eight of the ten deciles the two rules are indistinguishable — both sit at roughly 49.8%. The fixed stop is not generally worse. Then in the fastest tenth it drops to 44.5% while the ATR stop holds at 50.0%.

That is the finding in one line: a fixed stop is fine almost all of the time, and fails precisely when the market is moving fastest — which is when it matters most and when you are least able to react.

The mechanism is visible in the data

There is a column that shows why. An ambiguous trade is one where a single 15-minute bar covers both the stop and the target, so which was hit first is unknowable. We count those as losses in both regimes. Their frequency is a direct measure of how small the stop is relative to how far price travels in one bar.

A fixed stop produces two to four times as many of these, in every configuration. That is the fixed stop being swallowed by single candles in fast markets, and it is the same phenomenon the decile table shows from a different angle.

What holds everywhere

Three configurations were tested. On the headline measures, the adaptive stop won all three times:

Note the size of it. The win-rate gain is between 0.4 and 1.2 percentage points. Real, consistent, and nothing like the transformation ATR is usually sold as.

Note also the sign. Every configuration lost money. The entries are random and every trade pays the spread. ATR sizing lost less; it did not win. A stop rule changes the shape of your results, not their sign — the same conclusion our risk-reward study reached about target placement.

The result that does not hold

At a 2:1 target, the consistency advantage reverses. The fixed stop varied 1.5 points across the deciles against 2.6 for ATR — the opposite of the headline result, on the same data.

The obvious explanation would be trades running out of time before resolving, which is exactly what distorted the high-ratio rows in our risk-reward study. It is not that: essentially every trade resolved in all three runs, so truncation is ruled out.

We do not have a confirmed mechanism. Reporting it as “ATR is more consistent” and leaving this run out would have produced a cleaner article and a false one. The robust claim is the narrower one: ATR sizing gives a better win rate, better expectancy and far fewer single-bar wipeouts in every configuration tested. The dramatic consistency gain is real at a 1:1 target and does not generalise.

What we do with it

Our own system sizes every stop at 1.5 × ATR and rejects any signal whose stop would exceed 3.5 × ATR. That is the configuration in the first row of the tables above, chosen before this study existed and left unchanged by it.

The practical translation, if you prefer a number: at gold’s median 15-minute ATR of $2.08, 1.5 × ATR is about $3.11 per ounce. In the fastest tenth of conditions the same rule gives you roughly $10.56. A fixed stop cannot be both, and the difference between them is where the 5.4-point drop in the table comes from. You can work the lot size for either with the lot size calculator.

Limits

One instrument, one broker, 2014-01-14 to 2026-09-21, on 15-minute bars. Ambiguous bars are resolved as losses, which is pessimistic but applied identically to both rules, so it cannot favour either. Entries are random by construction: this isolates the stop rule cleanly but says nothing about how either rule interacts with an entry that has a genuine edge — plausibly the more useful question and a much harder one.

Only two multipliers and two target ratios were tested, and one of the four combinations contradicted the consistency result. Treat the win-rate and expectancy findings as the durable ones. The full decile tables for all three runs are published with our other studies.

Common questions