Do Liquidity Sweeps Really Reverse? 3.5 Years of Gold Data

Nobody ever attaches a number to it.

So we did. We measured every first sweep of eight major liquidity levels on Gold — 2,434 sweeps across 3.5 years of 5-minute data — and asked one simple question: after a sweep, how often does price actually close back on the other side of the level?

The short answer: about half the time. A coin flip. The “sweeps reverse” story, stated unconditionally, is a myth.

The longer answer is more interesting, because a handful of conditional edges are real — including one level that does the exact opposite of what the folklore says. Let’s go through it properly.

How we measured it

Every statistic below comes from one study, run on one dataset, with definitions fixed in advance:

• Data: XAU/USD spot (Dukascopy), M5 bars, 3 January 2023 → 2 July 2026 — 807 trading days.
• Levels tracked (8 families): previous day high/low (PDH/PDL), previous week high/low (PWH/PWL), London session high/low (LdnH/LdnL), Asia session high/low.
• Sweep definition: a level only becomes “sweepable” after its formation period ends (you can’t sweep a London high while London is still building it). We count only the first sweep of each level per period — the event the stop-run narrative is actually about.
• Reversal definition: at the horizon (30 minutes, or end of day), price closes back on the original side of the level. Sweep PDH, close back below PDH by EOD → reversal. Close above → continuation.
• Regime: classified from the previous day’s daily ADX14/DI — Trend up, Trend down, Range, or Transition — so it’s known before the sweep happens, not fitted after.
• Uncertainty: every rate carries a Wilson 95% confidence interval. If the interval contains 50%, we do not call it an edge, full stop.

Sample size: n = 2,434 sweeps. Big enough that the aggregate numbers are stable; small enough in some sub-buckets that we’ll tell you exactly where the data gets thin.

The headline result: sweeps do NOT systematically reverse

Across all families, the end-of-day reversal rate sits at roughly 50%. Here is the full breakdown by level family:

Read that table slowly, because it kills two popular ideas at once.

First, no family shows an unconditional fade edge. PDH at 46% [41–51], PDL at 50% [44–55], London highs and lows in the 47–48% zone — every confidence interval either contains 50% or sits below it. If you have been fading every PDH sweep on principle, the data says you’ve been flipping a coin at best.

Second — and this is the one that should sting — the previous week high is the anti-sweep level. PWH sweeps reverse by end of day only 36% of the time [27–46], n=98. The upper bound of the confidence interval doesn’t even reach 50%. When Gold takes out last week’s high, the majority outcome is that it stays out. A weekly-high break on this dataset behaved more like a breakout than a trap — which makes sense once you say it out loud: clearing a weekly extreme is a bigger statement than clearing yesterday’s wick.

The 30-minute horizon isn’t the escape hatch either

“Fine, but the reversal happens fast — EOD is too slow.” We checked the 30-minute horizon too:

PDL at 30 minutes is the only bucket that pokes above the line — 56% [50–61] — and its interval touches 50% at the boundary. Suggestive, not bankable. Everything else is coin-flip territory again.

Where the real edges hide: condition on regime

The aggregate ~50% is the average of very different market states. When we split sweeps by the previous day’s regime (ADX14/DI on the daily), a few buckets separate cleanly from 50% — CI fully clear, minimum n of 25:

Two things stand out.

The only fade-shaped edge in the whole study is the Asia Low in a Transition regime: 65% [53–76] of those sweeps close back above the level by end of day. That’s the classic “run the overnight stops, then reverse” pattern — but it shows up in one specific context, not everywhere the folklore claims.

The PWH result inverts the playbook entirely. In a Range regime, a swept previous-week high reverses just 29% of the time [16–47], n=31. Fading that sweep means fighting a 70%+ historical continuation rate. The same level the sweep narrative tells you to short is, conditionally, the strongest continuation signal we found.

Timing of the sweep matters less than you’d think: sweeps during the London morning reverse 44% [38–49] of the time by EOD (n=304), and NY-morning sweeps land at 49% [44–53] (n=459). No session magically makes sweeps reverse.

What this changes at the screen

If you trade sweeps on Gold, three practical conclusions fall out of the table:

1. Stop fading sweeps by default. The unconditional baseline is ~50%. A strategy built on “sweep = reversal” has no statistical foundation on this dataset — whatever edge such a trader has is coming from their filters, not the sweep itself.
2. Treat the level family and regime as the signal, not the sweep. The same event — stops taken above a high — means opposite things at a PWH in a range (29% reversal → continuation setup) and at the Asia Low in transition (65% reversal → fade setup). Context isn’t a nuance here; it’s the entire edge.
3. Respect small n. The Asia-Low edge rests on 63 events, the PWH-in-Range on 31. Real, CI-clear, and worth knowing — but nobody should size up as if these were laws. They’re historical frequencies with honest error bars, and the intervals are wide.

This is exactly the gap our Sweep Outcome Stats indicator (Pro) closes on the chart: for every sweep on your own loaded history, per level family, it prints how often price closed back through the level within 20 bars (Back ≤h), how often it was still reversed at the horizon bar (Rev @h — the definition used in this study — with its Wilson 95% margin), the average extension in ATR, and the sample size next to every number. The two numbers differ on purpose: on XAU/USD M5 about 8 sweeps in 10 print a close back inside within 20 bars, but only about half are still reversed at the horizon — both are true, and only the second one is the reversal rate of this study. The regime-conditioned rates above come from the study itself; on the chart, read the regime from the SLR dashboard (Core) or the Hurst Regime Classifier (Pro) next to the panel. Liquidity Levels (Core) draws and tracks the eight level families themselves. And session context — which regime you’re actually in when Asia hands off to London — is the subject of its own study: Asia doesn’t predict London, and we checked 821 sessions.

The honest limits

Every study has edges of validity. Here are this one’s, stated plainly:

• “Reversal” is a close-based observation, not a simulated trade. We measure whether price closed back across the level at the horizon — no entries, no stops, no targets, no spread or slippage. A 65% reversal rate does not translate into a 65% win rate for any actual trade plan; it tells you which way the distribution leans, nothing more.
• Regime comes from daily spot data (previous day’s ADX14/DI). A different regime definition would carve the buckets differently.
• First sweeps only. Second and third touches of a level are a different animal and are not covered by these numbers.
• One instrument, one window. XAU/USD spot, 2023–mid-2026. Gold in a different macro era — or any other instrument — needs its own study. Historical frequencies are not guarantees, on this or any dataset.
• Some cells are thin. We enforced n ≥ 25 and CI-excludes-50% before calling anything an edge, but n=31 is still n=31.

Bottom line

Across 2,434 sweeps and 3.5 years of Gold data, liquidity sweeps reverse about half the time — the unconditional “sweep and reverse” story is a coin flip wearing a trench coat. The previous week high actually continues after a sweep far more often than it reverses (36% EOD reversal overall, 29% in a Range regime). The one textbook-shaped fade — the Asia Low in a Transition regime, 65% [53–76] — is real, narrow, and comes with wide error bars.

That’s what verifiable probabilities look like: less exciting than the myth, and considerably more useful. For how the event-day version of this logic plays out, see what actually happens on FOMC days, and for when Gold actually trades, the volume-by-hour numbers.

Delta Terminal ResearchStudies run on 3.5 years of XAU/USD spot data, developed hand-in-hand with a full-time Gold day trader. We publish the negative results too.