Short answer: 21 of the 288 versions passed our statistical bar, and every one of them was a loss. All 21 were the same trade: selling gold after a recent high was taken and reclaimed. Buying after a recent low was taken showed nothing at all. The most reliable thing we found in nine years of gold sweep data is a way to lose money, which is worth knowing, because it is the trade the concept is most often used for.
Read this first: what this test can and cannot tell you
- Gold only (XAUUSD), Deriv's feed, candles on the UTC clock, January 2017 to September 2026 (9 years 8 months). Gold rose roughly threefold over this period. That matters enormously for anything on the sell side, and it is dealt with explicitly below rather than left as a caveat.
- One precise, written definition, set out below.
- Every trade pays the spread (at least 15 points, from the broker's own per-candle record, once per round trip).
- Entries are market orders at the next candle's open.
- A record of what one rule did on past data. Not advice, not a signal, not a forecast.
How to read the numbers (skip if you already know)
| Term | What it means here |
|---|---|
| bp (basis point) | 0.01% of price. At $4,300 gold, 1 bp ≈ $0.43 per ounce. |
| The spread you pay | Charged once per round trip, at least 15 points, from the broker's own record. Measured across this sample it averages 0.89 bp per trade, and it has been shrinking as gold has risen: 1.19 bp in 2017, 0.79 in 2023, 0.37 so far in 2026, because 15 points is a smaller share of $4,300 than of $1,250. |
| t | Distance from zero in units of the result's own noise. Under 2 is ordinary randomness. A large negative t means a reliable loss, not a weak result. |
| The bar (3.75) | 288 versions were tried, so the threshold rises to match (Bonferroni, two-sided 5%). |
| Alpha | What is left after subtracting gold's own move over the same holding time in the same year. This is how a short is judged fairly in a rising market. |
| "Passed" | Cleared the bar, kept its sign in ≥70% of years and in both halves, and kept it after gold's drift was removed. It does not mean "profitable", a reliable loss passes too, and 21 did. |
What we tested
The idea, in plain words. Stop orders pile up just beyond obvious highs and lows. The claim is that price is pushed through those levels to trigger them, a "sweep" or "stop raid", and then turns around, so the sweep marks the end of the move rather than the start of one. The trade is to go the other way once price closes back inside.
The rules, exactly.
- Sweep of highs: let PH be the highest high of the previous N candles (N tested at 20, 50 and 100). Candle i trades above PH. Within K candles (K tested at 1 and 3) one closes back below PH.
- The trade: SELL at the next candle's open. Exit at the close 4, 12 or 24 candles later. Lows mirrored → BUY.
- Charts: M1, M5, M15, M30, H1, H4, D1, W1. One position at a time per version.
- Grid: 8 charts × 3 lookbacks × 2 reclaim windows × 3 holding times × 2 directions = 288 versions.
Results by chart: the standard version (50-candle extreme, reclaimed in 1 candle, exit after 12)
| Chart | Direction | Trades | Win rate | Net per trade (bp) | t | Passed the bar? |
|---|---|---|---|---|---|---|
| M1 | buy (low swept) | 48,147 | 46.2% | −0.68 | −14.21 | no, loses |
| M1 | sell (high swept) | 51,993 | 45.2% | −0.92 | −21.34 | no, loses |
| M5 | buy | 9,841 | 50.0% | −0.59 | −2.43 | no |
| M5 | sell | 11,110 | 47.6% | −1.38 | −6.69 | no, loses |
| M15 | buy | 3,221 | 51.7% | −0.69 | −0.98 | no |
| M15 | sell | 3,820 | 48.2% | −2.74 | −4.50 | no, loses |
| M30 | buy | 1,475 | 53.1% | −0.84 | −0.58 | no |
| M30 | sell | 1,844 | 46.3% | −4.80 | −3.99 | no, loses |
| H1 | buy | 755 | 52.8% | +0.20 | +0.07 | no |
| H1 | sell | 967 | 45.1% | −9.61 | −4.22 | yes, a reliable loss |
| H4 | buy | 198 | 52.5% | +8.84 | +1.02 | no |
| H4 | sell | 284 | 49.3% | −5.53 | −0.67 | no |
| D1 | buy | 33 | 63.6% | +59.48 | +0.98 | too few trades |
| D1 | sell | 53 | 41.5% | −63.70 | −1.12 | too few trades |
| W1 | buy | 2 | , | , | , | too few trades |
| W1 | sell | 14 | 35.7% | −497.95 | −2.74 | too few trades |
On the random control series: 0 of 288 passed, maximum |t| 2.78. The machinery is not manufacturing results.
The 21 versions that passed: every one a loss
| Chart | Lookback | Reclaim | Hold | Trades | Win rate | Net (bp) | t | Alpha t |
|---|---|---|---|---|---|---|---|---|
| M30 | 20 | 3 | 4 | 5,683 | 46.3% | −2.52 | −6.30 | −2.87 |
| M30 | 20 | 1 | 4 | 4,753 | 46.4% | −2.72 | −6.07 | −3.02 |
| M5 | 100 | 1 | 24 | 5,879 | 47.2% | −2.25 | −5.48 | −2.15 |
| M30 | 20 | 3 | 12 | 3,508 | 46.3% | −4.39 | −5.30 | −2.51 |
| M30 | 20 | 1 | 12 | 3,125 | 46.2% | −4.70 | −5.26 | −2.68 |
| H1 | 20 | 3 | 12 | 1,765 | 45.4% | −7.82 | −4.90 | −2.56 |
| H1 | 50 | 1 | 12 | 967 | 45.1% | −9.61 | −4.22 | −2.44 |
| … 14 more, all short, all negative |
They sit across three different charts, three lookbacks, both reclaim windows and all three holding times. That consistency is the point: this is not one unlucky configuration, it is the whole sell side of the concept.
And it is not simply "shorting gold in a bull market". The alpha column already removes gold's own drift over the same holding times in the same years, and it stays negative with a t beyond 2 on every one of the 21. Selling the sweep did worse than being short gold generally.
Does the sweep itself matter? The near-miss control
For every sweep we also took the near miss, a candle whose high came within 0.1 ATR of the prior extreme, closed back below it, but never actually took it. No stops triggered, no liquidity grabbed, otherwise the same picture.
| Chart | Direction | Sweep (bp) | Near miss, level not taken (bp) | Difference | t |
|---|---|---|---|---|---|
| M15 | sell | −2.74 | −2.13 | −0.67 | −0.68 |
| M15 | buy | −0.69 | −1.38 | +0.70 | +0.56 |
| H1 | sell | −9.61 | −6.95 | −2.64 | −0.70 |
| H1 | buy | +0.20 | −5.21 | +5.47 | +1.01 |
Selling into the highs of the last 50 candles lost whether or not the level was actually taken, and the difference between the two is inside chance. So the honest reading of the 21 losses is not "the sweep traps sellers", it is that on gold, over these years, selling a push into recent highs on an hourly chart lost money, sweep or no sweep.
Year by year: selling the swept high on the hourly chart, exit after 12 candles
| Year | Trades | Win rate | Net per trade (bp) |
|---|---|---|---|
| 2017 | 91 | 42.9% | −8.36 |
| 2018 | 88 | 54.5% | +0.23 |
| 2019 | 93 | 35.5% | −11.99 |
| 2020 | 118 | 39.8% | −13.19 |
| 2021 | 88 | 52.3% | −1.28 |
| 2022 | 92 | 42.4% | −9.02 |
| 2023 | 93 | 54.8% | −3.97 |
| 2024 | 107 | 43.0% | −4.35 |
| 2025 | 133 | 42.1% | −18.33 |
| 2026 | 64 | 48.4% | −26.02 |
Negative in nine years out of ten, and getting worse, not better. Whatever the last two years have been, they have not been kind to anyone fading gold's highs.
With a stop and a target
Stop beyond the swept extreme, target twice the risk, up to 200 candles. A market with no pattern reaches a 2:1 target before its stop about one time in three.
(Descriptive: added after the main test, not pre-registered.)
| Chart | Direction | Trades | Target before stop | Chance gives | Mean result |
|---|---|---|---|---|---|
| M15 | sell | 8,307 | 25.8% | 33.3% | −0.23 R |
| M15 | buy | 6,235 | 29.9% | 33.3% | −0.10 R |
| H1 | sell | 2,058 | 28.8% | 33.3% | −0.15 R |
| H1 | buy | 1,482 | 31.4% | 33.3% | −0.06 R |
| H4 | sell | 579 | 29.2% | 33.3% | −0.15 R |
| H4 | buy | 379 | 34.1% | 33.3% | +0.03 R |
The sell side is the worst plan in this entire study: 25.8% of trades reached a 2:1 target before their stop on the 15-minute chart, against 33.3% from chance alone, over 8,307 trades.
The obvious follow-up, and why we are not calling it a finding
If selling the swept high loses reliably, does buying it win? We ran exactly that: same signals, same candles, opposite direction.
| Chart | Trades | Win rate | Net per trade (bp) | t | Alpha (drift removed) | Years positive |
|---|---|---|---|---|---|---|
| M5 | 11,110 | 46.6% | −0.37 | −1.78 | +0.26 (t +1.24) | 7/10 |
| M15 | 3,820 | 48.7% | +0.95 | +1.56 | +1.10 (t +1.81) | 6/10 |
| M30 | 1,844 | 51.6% | +2.92 | +2.43 | +2.38 (t +1.98) | 8/10 |
| H1 | 967 | 53.1% | +7.74 | +3.41 | +5.54 (t +2.44) | 8/10 |
This does not pass. The bar for this grid is 3.75 and the best cell reaches 3.41; the 15-minute and 5-minute versions are ordinary noise. It was also not pre-registered, it was run after seeing the sell-side losses, which is precisely the circumstance in which patterns appear because you went looking for them.
What it does mean is that the sell-side losses are not an accounting artefact. Something in the data has price continuing upward after a high is taken, on the slower charts, more strongly in 2025 (+17.42 bp) and 2026 (+25.35 bp) than before. On the random series the same test reads +0.89 bp with a t of 0.13, nothing.
By our own rules that makes it a candidate, not a result: it needs its own written registration, its own control, and data that did not exist when we found it before anyone should act on it. We are publishing it here rather than quietly keeping it because a page that only reports failures is not being honest about what the data actually showed.
What this page does not say
- It does not say sweeps aren't real. Stops do cluster beyond obvious levels, and price does trade through them, the detection here found tens of thousands of them.
- It does not say the reversal never works. It says that as a mechanical rule, on gold, over these years, with costs charged, the reversal side lost, and lost reliably enough to clear a demanding statistical bar 21 times.
- It does not test sweeps qualified by anything else, a session, a higher-timeframe level, a news release. Each of those is a different concept and needs its own pre-registered test.
How we tested
- The rules were written and dated before the scanner existed. Nothing was tuned afterwards.
- Closed candles only; market entry at the next open; one position at a time per version.
- 288 versions, so the bar rises to match (Bonferroni). A version must also hold in ≥70% of years, in both halves, and survive the removal of gold's own drift.
- The identical grid on a random price series: clean, 0 survivors, maximum |t| 2.78. Honest limit: that control is strongest on the fast charts where it has tens of thousands of trades; on the hourly chart it carries about 1,500, enough to reveal a 10 bp effect but not a 5 bp one.
Reproduce it: every rule is written out in full above, so anyone with gold price data can rebuild this test and check our numbers. The candles are the broker's and are not ours to redistribute.