MEASURED · 17 SEPTEMBER 2026

Do Fibonacci retracements work?

We found 32,592 gold pullbacks by a fixed rule and measured exactly where each one stopped. If Fibonacci levels mattered, pullbacks would stop at them more often than at the levels next to them. The golden ratio scored 1.00 — exactly as popular as its neighbours.

What a Fibonacci retracement is

Take a move from a low to a high. Draw horizontal lines at 23.6%, 38.2%, 50%, 61.8% and 78.6% of the way back down. The claim is that a pullback tends to stop near one of those lines before the original move resumes — most often 61.8%, the “golden ratio”. The percentages come from ratios between numbers in the Fibonacci sequence; 50% is not one of them but is included by convention.

The usual justification is self-fulfilling: so many traders watch these levels that their orders make price react there. That is a reasonable mechanism. It is also testable, because if it were true it would leave a mark in where pullbacks actually end.

Why the usual demonstration proves nothing

Fibonacci is normally shown working on a chart where someone chose the swing, drew the levels, and pointed at a reversal near one of them. That process cannot fail. Choose a different swing and a different level lines up. Most pullbacks pass through several Fibonacci levels on the way down, so one of them is always near wherever price turned.

So the test here removes the human from both choices. Swings are identified by a mechanical zigzag rule, and every level is scored the same way, whether it is a Fibonacci level or not.

How it was measured

1. Find swings mechanically. A swing point is confirmed only once price reverses by a fixed percentage. No discretion, no hindsight.
2. Measure where each pullback stopped. For three consecutive swing points, the depth is how far the second leg travelled as a percentage of the first.
3. Compare each level to its neighbours. Count pullbacks ending within 2 points of a level, and divide by the average count in the bands just above and below it. A ratio of 1.00 means the level is exactly as popular as the ones beside it.
4. Rank it against every other level. The same ratio is computed for all 131 levels from 20% to 85%. A real Fibonacci effect would put those levels near the top.

Before it was run on gold, the method was tested on data where the answer was known. On a random walk it flagged nothing. With extra pullbacks deliberately planted at 61.8%, it flagged 61.8% and nothing else. A test that cannot find a planted effect cannot be trusted when it finds nothing, so this step is the reason the result below means something.

The result

The largest sample: 18,304 pullbacks on the 15-minute chart, 2014-01-14 to 2026-09-17. “Most it could be” is the top of the 99% interval — the largest excess over its neighbours the data still allows.

None of the five stands out. The golden ratio sits at 1.00, ranked 70 of 131 — the middle of the pack. And because the sample is large, this is not just an absence of evidence: the data rules out 61.8% being more than about 12% more popular than the levels either side of it.

It is not one lucky timeframe

A real effect should survive a change in how swings are defined, so the whole test was repeated on four more settings.

0 of 25 tests found a Fibonacci level that pullbacks preferred. Look at the last column: the most popular level on each run was a different number every time, and in four of five runs it was not a Fibonacci level at all. That is what noise looks like — some level always comes top, and it is never the same one.

Only one test out of 25 produced an interval that excluded 1.00, and it went the wrong way: 23.6% on the hourly chart was less popular than its neighbours. With 25 tests at this strictness, about a quarter of one such result is expected by chance, so it is not evidence against Fibonacci either.

The result that would have fooled us

On the 4-hour chart, 23.6% ranked first out of 131 levels, with a ratio of 2.33 — more than twice as popular as its neighbours. That is exactly the kind of number that gets screenshotted.

It rests on 7 pullbacks. The 99% interval runs from 0.40 to 16.03, which means the data is equally consistent with that level being unpopular or sixteen times as popular as its neighbours. Seven observations cannot tell the difference, and the same number ranked nowhere on the charts with thousands of pullbacks behind them. The general lesson is the one in how many trades a result needs: a striking number from a small sample is more likely to be luck than a law.

So why does it feel like it works?

Because it is almost impossible for it not to look right. Five levels spread across a move, each given a small zone either side, cover a large share of everywhere a pullback can end. We measured how large:

The two numbers match. With zones three points wide, nearly half of all pullbacks “respect a Fibonacci level” — because nearly half of the space where pullbacks can end is inside a Fibonacci zone. Replace the five levels with any five others spaced the same way and they would be respected exactly as often. Widen the zones, or add 88.6% and 127.2% as many charting tools do, and almost every pullback touches one.

This is also why the self-fulfilling argument does not rescue it. If enough traders acted at these levels to move price, pullbacks would end there more often than the area covered predicts. They end there almost exactly as often as geometry alone would put them. It is the same finding as our study of whether round numbers act as support and resistance on gold, arrived at from a different direction.

What this does and does not say

It says that on gold, pullbacks do not stop at Fibonacci levels more often than at the levels next to them. That was the specific claim, and it does not hold.

It does not say a trader using Fibonacci loses money. Fibonacci can still be a consistent way to decide where to place an order or a stop, and a rule followed consistently can be better than no rule. What it should not be trusted to do is predict where price will turn. Our own trading system uses no Fibonacci levels — verified in its source before this page said so.

Limits

One instrument, one broker, 2014-01-14 to 2026-09-17. Swings are defined by a percentage zigzag, and a human picking swings by eye would choose differently — though a method that only works when a person chooses the swing after the fact is exactly the problem described above. The zone was fixed at two points either side; a much tighter precision effect could in principle hide inside that, but the neighbouring bands are the same width, so a genuine spike at 61.8% would still raise its ratio.

The daily chart has only 407 pullbacks in 12.7 years and cannot support any conclusion on its own. The script and the raw numbers for all five runs are published with our other studies, including the intervals that make each result checkable.

Common questions