Do Fibonacci retracements work?
A Fibonacci tool marks three depths on a pullback — 38.2%, 50% and 61.8% of the move before it — and says price is more likely to turn there. That is a claim about counting, so it can be checked by counting. 25,934 pullbacks, 15 markets, four timeframes, measured only on the half of each market's history the model never saw. Pullbacks are no more likely to end at those depths than anywhere else.
Where pullbacks actually end
If a level attracts, pullbacks should pile up at it. For each level, the share of all pullbacks that ended within ±1% of it, against the average share in the bands immediately either side — same width, same sample, just not the magic number.
| Level | Ended at it | Ended nearby | Ratio |
|---|---|---|---|
| 23.6% | 2.09% | 1.95% | 1.07 |
| 38.2% | 1.85% | 1.94% | 0.95 |
| 50.0% | 1.50% | 1.61% | 0.93 |
| 61.8% | 1.31% | 1.32% | 0.99 |
| 78.6% | 1.11% | 1.06% | 1.05 |
A ratio of 1.00 means indistinguishable from any other depth. Every level is within a few hundredths of 1.00, and they miss in both directions — which is what noise looks like, and not what an effect looks like.
And what happens next is no different either
A weaker version of the claim survives if price merely resumes more often from a level. It does not. Below: how often the trend resumed for pullbacks ending within ±2.5% of each level, against the bands just short of and just past it.
| Level | At the level | Just short of it | Just past it | Pullbacks |
|---|---|---|---|---|
| 23.6% | 95.4% | 97.6% | 89.8% | 1,332 |
| 38.2% | 84.5% | 88.1% | 82.8% | 1,227 |
| 50.0% | 80.6% | 82.1% | 77.8% | 1,049 |
| 61.8% | 74.7% | 75.6% | 73.9% | 861 |
| 78.6% | 71.0% | 70.7% | 64.1% | 692 |
At four of the five, the depth just short of the level resumed more often than the level did. At the fifth the level is ahead by 0.3 of a percentage point, which on 692 pullbacks is nothing. That is the shape of a smooth curve sloping down, sampled at arbitrary points — not the shape of a level doing something.
What does predict: depth, continuously
The thing the levels are drawn on top of is real. How often a trend resumed, by how deep the pullback went as a share of the move before it:
| Retracement | Trend resumed | Pullbacks |
|---|---|---|
| 0–10% | 100.0% | 307 |
| 10–20% | 99.0% | 2,044 |
| 20–30% | 93.5% | 2,590 |
| 30–40% | 86.8% | 2,513 |
| 40–50% | 82.4% | 2,254 |
| 50–60% | 77.1% | 1,902 |
| 60–70% | 74.0% | 1,648 |
| 70–80% | 70.6% | 1,451 |
| 80–90% | 65.3% | 1,189 |
| 90–100% | 65.3% | 1,097 |
| 100–110% | 65.3% | 972 |
| 110–120% | 63.1% | 832 |
| 120–130% | 62.0% | 760 |
| 130–140% | 64.6% | 701 |
| 140–150% | 60.0% | 640 |
It falls from 100.0% at the shallowest to 60.0% at the deepest, without a step anywhere. Drawing three lines across that curve throws away everything between them, and the lines land on nothing.
The zone is not where the pullbacks are
The 38.2–61.8% band is where a Fibonacci user expects to be working. Only 18.9% of pullbacks end inside it, and the median pullback retraces 68% of the prior move — past the whole zone. 34.5% go beyond 100%, which means the move being retraced was not a pullback at all but the start of a reversal.
Worse for the idea: pullbacks shallower than 38.2% — the ones that never reach the zone — resumed 93.4% of the time, against 80.3% inside it. The zone is not the best place on the curve. It is a worse place than the part before it.
How this was measured
Same pullbacks, same engine and same definitions as the pullback study: a correction opens when price retraces a quarter of θ from a trend extreme, and closes when the trend makes a new extreme (resumed) or gives up θ against itself (reversed). θ is fixed per timeframe on each market's training half, so the numbers above come from history the threshold never saw. The published study measures depth as a share of θ; a Fibonacci level is a share of the previous leg, so that is what is measured here. Same corrections, different denominator — and the denominator is the entire question.
Two honest limits. Pullbacks retracing more than three times the prior leg are left out (20% of the sample): when the leg is very short the percentage explodes, and a "700% retracement" is a fact about the denominator rather than about price. And this measures the levels as a Fibonacci tool draws them — from one swing to the next, as the engine marks swings. It cannot speak for a level someone drew somewhere else by eye.
25,934 pullbacks is the out-of-sample half. The full set the engine walked, both halves together, is 94,347 — the same corrections /pullbacks reports, from the same run.
Where this fits
Turnmarks marks the hour a push runs out on 15 charts, and publishes what happened after every past marker, per instrument, so the tool can be judged before it is paid for. It does not forecast, and it does not draw levels. How a turn marker is decided, or open the charts — free plan, no card.