Key takeaways
- The confirmed setup is random. 11,990 decided reclaim trades won 48.7% against a 49.4% control (n = 288,760). z = −1.60, p = 0.111. Not significant in any asset-class cut we tried.
- The unconfirmed setup is not random — it is bad. Fading the level on touch, without waiting for a reclaim, won 40.9% of 18,088 trades. z = −22.30, p < 0.001.
- So the confirmation rule earns its place. Not by creating an edge, but by removing a large negative one. That is a real and useful finding, and it is not what the setup is sold on.
- No instrument survives correction. 2 of 22 reach p < 0.05 individually, in opposite directions. The Bonferroni threshold for 22 tests is p < 0.0023. Neither clears it.
- At 2R it gets worse. 30.8% against a geometric fair value of 33.3%, z = −5.61 — roughly the cost of the spread.
First, in plain English
Yesterday's high and low are the most watched prices on the chart. Stop-loss orders pile up just beyond them, because that is where a lot of traders agree "I am wrong." The liquidity grab story says a large participant needs those orders to fill a position, pushes price through the level to trigger them, and then lets price snap back. The snap-back is the trade.
It is a good story. It is also testable, which is rarer in this corner of trading than it should be. Here is exactly what we tested.
Step 1: mark yesterday's high and low. For every day, take the previous trading day's highest and lowest price. Call them PDH and PDL.
Step 2: wait for a sweep. Between 07:00 and 21:00 UTC, find the first moment price trades beyond one of those levels. We found 20,335 of these across 22 instruments.
Step 3: wait for the reclaim. The setup is not "price went through the level." It is "price went through the level and came back." We gave it eight M15 bars — two hours — to close back on the original side. 13,096 of the 20,335 sweeps did that. The other 7,239 just kept going, which is the part of the sample the screenshots never show.
Step 4: take the trade, mechanically. Enter at that reclaim close. Stop at the extreme of the sweep. Target the same distance away, a clean 1:1. Then walk forward bar by bar and see which one it hits.
Step 5 — and this is the step almost nobody does: compare it to nothing. A win rate on its own is meaningless. 50% at 1:1 is what a coin does. So we also ran a control: enter every eighth bar of the same window, on the same instrument, with the same typical stop distance, in the same direction mix. If the liquidity grab contains information, it should beat that.
The answer: it does not. The liquidity grab won 48.7% of its decided trades. The control won 49.4%. The setup that is supposed to show you where institutions are positioning performed a fraction of a point below entering at arbitrary moments. The gap is well inside the margin of error — the honest statement is not "it loses" but "it does nothing."
How this was measured
- Data. Dukascopy bid-side M15 OHLC for 22 instruments — 13 FX pairs, gold, silver, Brent crude, four equity indices and two crypto pairs — spanning 2021-12-27 to 2026-08-14. Each series is capped at 100,000 bars, so the FX pairs begin in August 2022 and the two crypto instruments only in October 2023.
- The levels. The previous trading day's high and low, computed from the same M15 series. Weekend bars are excluded for everything except crypto.
- The sweep. The first bar between 07:00 and 21:00 UTC that trades beyond the level. Sweeps deeper than 0.5 × ATR20 are discarded — that is a trend day leaving, not a stop run, and including them would flatter the result by mixing in obvious continuations.
- The reclaim. A close back on the original side of the level within 8 bars (two hours). Entry is that close, the stop sits at the swept extreme, and the target is one risk away.
- The control. An entry every eighth bar of the same window with a stop set to the median reclaim risk for that instrument. 288,760 decided control trades. This is the number the setup has to beat, and it is the reason this page can say anything at all.
- Ties go against the trade. When a single M15 bar contains both the stop and the target, the trade is counted in the denominator and not counted as a win.
- Significance. Pooled two-proportion z-test, the same one used in our candlestick pattern study and indicator signal study, so the three are directly comparable.
The result
Sorted by edge over the control. The last column is the same instrument traded without waiting for the reclaim — the comparison that turns out to matter most.
| Instrument | Sweeps | Reclaim win rate | Control | Edge | z | Significant? | No confirmation |
|---|---|---|---|---|---|---|---|
| AUD/CAD | 521 | 55.2% | 49.5% | +5.7 | +2.45 | * | 39.7% |
| EUR/USD | 604 | 52.9% | 49.4% | +3.5 | +1.62 | ns | 42.1% |
| Gold | 549 | 52.3% | 49.5% | +2.8 | +1.20 | ns | 36.8% |
| GBP/USD | 632 | 50.8% | 49.5% | +1.3 | +0.60 | ns | 43.1% |
| EUR/JPY | 546 | 50.6% | 49.4% | +1.2 | +0.53 | ns | 37.3% |
| EUR/CHF | 656 | 50.6% | 49.5% | +1.1 | +0.52 | ns | 49.6% |
| US30 (Dow) | 740 | 49.2% | 49.0% | +0.2 | +0.09 | ns | 45.4% |
| US100 (Nasdaq) | 695 | 48.8% | 48.8% | +0.1 | +0.03 | ns | 43.7% |
| AUD/USD | 503 | 49.0% | 49.5% | -0.4 | -0.19 | ns | 34.6% |
| Silver | 546 | 48.6% | 49.5% | -0.9 | -0.38 | ns | 37.0% |
| Brent crude | 712 | 48.4% | 49.5% | -1.1 | -0.56 | ns | 40.6% |
| NZD/CHF | 545 | 48.4% | 49.6% | -1.2 | -0.54 | ns | 45.2% |
| AUD/JPY | 481 | 48.1% | 49.5% | -1.4 | -0.57 | ns | 34.4% |
| USD/JPY | 490 | 47.2% | 49.1% | -1.9 | -0.78 | ns | 35.3% |
| EUR/GBP | 656 | 47.6% | 49.6% | -2.0 | -0.96 | ns | 44.6% |
| CAD/CHF | 632 | 47.2% | 49.5% | -2.3 | -1.08 | ns | 47.5% |
| BTC/USD | 480 | 46.8% | 49.5% | -2.8 | -1.14 | ns | 36.3% |
| US500 (S&P) | 729 | 46.1% | 49.1% | -3.1 | -1.54 | ns | 43.7% |
| ETH/USD | 466 | 46.6% | 49.8% | -3.1 | -1.26 | ns | 34.5% |
| GER40 (DAX) | 689 | 46.3% | 49.6% | -3.3 | -1.62 | ns | 41.3% |
| GBP/JPY | 556 | 46.1% | 49.4% | -3.3 | -1.46 | ns | 37.6% |
| USD/CAD | 668 | 45.1% | 49.6% | -4.5 | -2.21 | * | 44.2% |
Pooled across all 22: 13,096 reclaim setups, 11,990 decided, 48.67% won, against a control of 49.42% on 288,760 decided trades. Edge −0.74 points, z = −1.60, p = 0.111.
Two of twenty-two is what noise looks like
AUD/CAD at +5.7 points and USD/CAD at −4.5 both clear p < 0.05. It is tempting to read the first as a discovery. It is not, for two reasons.
First, they point in opposite directions. A real mechanism does not make the liquidity grab work on the Australian dollar against the Canadian and fail on the US dollar against the same Canadian dollar.
Second, running 22 tests at a 5% threshold is expected to produce roughly one false positive by chance alone. We got two. Correcting for the number of tests — Bonferroni, p < 0.0023, or |z| > 3.05 — neither survives, and nothing else comes close. The correct reading of that column is that no instrument in the sample behaves differently from the control.
This is the mechanism behind every convincing liquidity-grab backtest you have seen. Test one pair, over one period, and roughly one in twenty will look excellent. Publish that one.
The finding that is significant
There is one number on this page with real statistical weight, and it is not the headline setup.
Alongside the reclaim arm we measured the naive version: enter the reversal the moment price touches the level, no waiting, no confirmation, a fixed 0.25 × ATR stop and a 1:1 target. That is how most people actually trade a level, and it produced 40.87% winners across 18,088 decided trades. Against the same 49.42% control that is z = −22.30, p < 0.001. Overwhelmingly significant, and overwhelmingly negative.
Put the two arms side by side and the conclusion is more interesting than either alone:
- Fade the level on touch — 40.9%, decisively worse than random.
- Wait for the reclaim — 48.7%, statistically identical to random.
The confirmation rule is worth roughly eight percentage points. It is doing real work. What it is not doing is creating an edge — it is climbing back to the baseline from well below it. The smart money framework gets the instruction right and the explanation wrong: waiting for the reclaim protects you from a bad trade, rather than revealing an institutional footprint.
One caveat stated plainly, because it cuts against us: the naive arm uses a fixed 0.25 × ATR stop while the reclaim and control arms use structure stops of roughly 0.155 × ATR. The arms are not perfectly matched. The direction of that mismatch favours the naive arm — a wider stop pays proportionally less spread and is harder to clip — and it still lost by more than eight points.
What happens at a 2:1 target
The obvious rescue is to stop taking 1:1 and let the reversal run. We measured that too. At a 2R target the reclaim arm won 30.81% of 10,978 decided trades. The geometric fair value for a 2:1 structure is 33.33%, so it came in 2.5 points light, z = −5.61.
That shortfall is roughly what the spread costs over this sample, which is the least surprising result on the page: a setup with no edge does not acquire one by moving the target. It is worth noting that this arm is compared against a geometric constant rather than a matched control, because no 2R control was run — so treat it as indicative rather than as strong as the 1R result.
One number we are deliberately not reporting as a finding
The dataset contains a tempting statistic: sweeps that never reclaimed — the continuation cases — "won" 98.54% of 6,912 trades. That is not an edge, it is a definition. If the level is never reclaimed, then price by construction kept moving away from it, and any trade defined as "price kept moving away" wins. We are flagging it because it is exactly the kind of number that gets lifted out of a dataset and turned into a 98%-win-rate claim. It means nothing.
What this does not say
It does not say liquidity is fake, or that stops are not clustered above the previous day's high. They plainly are. What the data says is narrower and firmer: this stated mechanical version of the setup, entered this way, on these levels, in this window, contains no information a random entry does not.
A practitioner will object that real application adds higher-timeframe bias, a killzone filter, a fair-value-gap confluence, an order block at the reclaim. That objection is fair, and this study cannot refute it — discretionary confluence is not testable without turning it into rules. But it cuts both ways. Each of those filters now has a number to beat: 49.42%. A filter is worth keeping only if it clears the control by more than the spread, and the burden of showing that sits with whoever is teaching it.
The honest places to look next, in rough order of how testable they are:
- Sweep depth. We discarded sweeps beyond 0.5 × ATR20. The shallow ones may behave differently from the very shallow ones.
- Time of day. A 08:00 UTC sweep and a 19:00 UTC sweep are different events with different amounts of session left.
- Which level. PDH and PDL are only two of the levels the framework names. Weekly highs, session highs and equal highs are all mechanically definable.
- Higher-timeframe alignment. The most commonly cited filter, and the easiest to encode: does the reclaim agree with the daily direction?
- News. A sweep driven by a scheduled release is a different animal — our release volatility study measures how different.
Limitations
Stated plainly, because a study without them is marketing.
- M15 resolution. A stop and target inside the same 15-minute bar cannot be ordered; those trades count in the denominator and never as wins, which biases every arm slightly downward.
- Coverage is capped at 100,000 bars per instrument, so this is a roughly four-year sample for FX and under three years for BTC/USD and ETH/USD. It is not a multi-decade result.
- No spread, commission or swap applied to either arm. Costs move the setup and the control down together, so the comparison holds, but every absolute win rate here is optimistic.
- Bid-side data. Ask prices are not replayed, so a level measured on the bid is marginally early for a long and late for a short.
- One reclaim window (8 bars) and one stop rule (the swept extreme). Both are defensible and both are choices; a different window is a different study.
- Sweeps deeper than 0.5 x ATR20 are excluded as trend days. That threshold is a judgement call, and a different one would change the sample.
- The 2R arm has no matched control, only a geometric baseline.
Download this dataset
Every figure on this page as raw CSV and JSON, regenerated whenever the archive updates. Free for any use, including commercial, with attribution to FxBacktest.
- liquidity-grab-win-rate-by-instrument.csvLiquidity grab win rate by instrument
- liquidity-grab-win-rate.jsonAll tables plus metadata, one file
Cite or republish this data
This research is free to quote, screenshot, or republish — in an article, a video, a newsletter, or a course — as long as you credit FxBacktest with a link back to this page. Copy the attribution below.
Frequently asked
Does the liquidity grab actually work?
Not measurably. Across 11,990 decided sweep-and-reclaim trades on 22 instruments, the setup won 48.7% at a 1:1 risk-reward against a matched control that won 49.4%. That is a difference of -0.74 points with a z score of -1.60 and p = 0.111, which is not statistically significant. The liquidity grab is indistinguishable from entering at random with the same stop distance.
Is waiting for the reclaim worth it?
Yes, but not for the reason it is usually taught. Entering the reversal on the touch of the level without waiting for a reclaim won only 40.9% of 18,088 trades against the same 49.4% control - a z score of -22.30. Waiting for confirmation does not make the trade profitable; it rescues it from being clearly negative.
What is a liquidity grab in trading?
A liquidity grab, also called a liquidity sweep, stop hunt or turtle soup, is a move that pushes through an obvious level where stop orders are clustered - most commonly the previous day's high or low - and then reverses back through it. The claim is that the push exists to fill large orders against those stops, and that the reversal is the tradable event.
Which instruments had the best liquidity grab win rate?
AUD/CAD at 55.2% and EUR/USD at 52.9% were the highest of the 22 measured; USD/CAD at 45.1% and GBP/JPY at 46.1% the lowest. Only two of the 22 reached p < 0.05 individually, in opposite directions, and neither survives a Bonferroni correction for 22 tests. That pattern is what noise looks like.
What data was used?
Dukascopy bid-side M15 OHLC for 22 instruments, spanning 2021-12-27 to 2026-08-14. Coverage per instrument is capped at 100,000 bars, so FX pairs run from August 2022 and the two crypto instruments only from October 2023. Sweeps are accepted between 07:00 and 21:00 UTC and sweeps deeper than 0.5 x ATR20 are excluded as trend days.