Free research · Support and resistance

Support and Resistance on Gold: 78 Versions, 9 Years, and a Control That Beat Them

Short answer: buying support did nothing measurable. Selling resistance lost reliably, 8 versions cleared our statistical bar and all 8 were losses. And when we compared support and resistance entries against random entries using the same stop and the same target, on the same days, the random entries did better on every chart tested.

Tested onGold (XAUUSD)
DataJanuary 2017 to September 2026 (9 years 8 months)
Rule versions tested78
Win rate we measured (after costs)44.3% over 1,523 trades (M1, the biggest sample)Across every chart in the main table: lowest 44.3% on M1 (1,523 trades), highest 53.0% on H1 (1,182 trades). Winning more often is not the same as making money; the tables below show both.

What is support and resistance?

Price is supposed to stall and turn at levels where it stalled and turned before. The cleanest, least arguable version of a level is the previous session's extreme, because everyone can see it and nobody has to draw it.

Short answer: buying support did nothing measurable. Selling resistance lost reliably, 8 versions cleared our statistical bar and all 8 were losses. And when we compared support and resistance entries against random entries using the same stop and the same target, on the same days, the random entries did better on every chart tested.


Read this first: what this test can and cannot tell you

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 Once per round trip, at least 15 points. Across this sample it averages 0.89 bp per trade, falling from 1.19 bp in 2017 to 0.37 so far in 2026 as gold's price rose.
t Distance from zero in units of the result's own noise. Under 2 is ordinary randomness. A large negative t is a reliable loss.
The bar (3.41) 78 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, the fair way to judge a short in a rising market.
"Passed" Cleared the bar and held its sign across years and halves. It does not mean profitable: a reliable loss passes too, and 8 did.

What we tested

The idea, in plain words. Price is supposed to stall and turn at levels where it stalled and turned before. The cleanest, least arguable version of a level is the previous session's extreme, because everyone can see it and nobody has to draw it.

The rules, exactly.


Results: previous day's levels, exit after 12 candles

Chart Trade Trades Win rate Net per trade (bp) t Passed?
M1 buy support 1,340 50.0% −0.02 −0.04 no
M1 sell resistance 1,523 44.3% −2.79 −7.18 yes, a reliable loss
M5 buy support 1,335 51.2% +1.03 +1.28 no
M5 sell resistance 1,513 46.0% −2.31 −3.11 no
M15 buy support 1,315 51.3% +1.18 +1.06 no
M15 sell resistance 1,485 44.9% −2.23 −2.24 no
M30 buy support 1,301 50.7% +0.15 +0.10 no
M30 sell resistance 1,452 46.8% −2.72 −1.97 no
H1 buy support 1,182 53.0% −1.02 −0.46 no
H1 sell resistance 1,333 47.9% −5.45 −2.97 no
H4 buy support 751 52.2% +4.57 +0.92 no
H4 sell resistance 806 48.4% −6.65 −1.39 no

Results: previous week's levels, exit after 12 candles

Chart Trade Trades Win rate Net per trade (bp) t Passed?
M1 buy support 208 46.6% +1.25 +0.58 no
M1 sell resistance 278 41.7% −5.05 −3.67 yes, a reliable loss
M15 buy support 208 49.0% −4.13 −1.07 no
M15 sell resistance 278 47.1% −1.66 −0.77 no
H1 buy support 204 51.5% −5.73 −0.87 no
H1 sell resistance 273 43.2% −9.51 −2.12 no
D1 buy support 124 55.6% +62.57 +2.49 no (small sample)
D1 sell resistance 149 50.3% −15.85 −0.55 no

8 of 78 versions passed the bar. Every one was "sell resistance", and every one was a loss, the 1-minute and 5-minute charts on day levels, the 1-minute on week levels. On the random control series: 0 of 78, maximum |t| 2.38.

It is not only gold's rise. The 1-minute day-level short keeps a negative alpha of −1.85 bp with t −4.77 after gold's drift is removed. Selling the first touch of yesterday's high did worse than simply being short gold over the same twelve candles.


The control that matters: random entries, same stop, same target

For every support and resistance entry we also took one random entry per trading day, and gave both the identical plan, a stop 1 ATR away and a target 2× the risk. If levels are special, entries at levels should beat entries at nothing.

Chart At a level Win rate Random entry, same plan Win rate
M15 buy support (1,335) 30.6% random long (2,878) 31.1%
M15 sell resistance (1,510) 28.7% random short (2,862) 30.4%
H1 buy support (1,298) 30.7% random long (2,575) 32.7%
H1 sell resistance (1,473) 29.8% random short (2,610) 30.9%
H4 buy support (1,097) 32.3% random long (1,733) 33.8%
H4 sell resistance (1,218) 30.6% random short (1,771) 32.2%

Six comparisons, and the random entry wins all six. The differences are small, one to two points, and we are not claiming random entries are better in any strong sense. The claim is the one the table supports: entering at the previous session's extreme did not beat entering at an arbitrary time on the same day.


Year by year

Selling the first touch of yesterday's high, 1-minute chart, exit after 12 candles, the version that passed:

Year Trades Win rate Net per trade (bp)
2017 137 39.4% −2.17
2018 143 43.4% −1.53
2019 155 41.3% −2.34
2020 164 42.7% −4.48
2021 163 47.9% −1.94
2022 150 45.3% −1.88
2023 158 55.1% −2.77
2024 174 40.8% −3.36
2025 172 37.2% −3.81
2026 107 52.3% −3.32

Ten years, ten losses. Note 2023 and 2026, where more than half the trades finished green and the average was still negative, the losers were bigger than the winners. That is the shape of most "high win rate" selling into strength.

Buying the first touch of yesterday's low, 15-minute chart, the best-looking version on the page, and still not a result:

Year Trades Win rate Net per trade (bp)
2017 123 56.1% +2.09
2018 152 46.7% −0.04
2019 128 48.4% −1.45
2020 122 53.3% +0.12
2021 144 55.6% +5.05
2022 150 50.0% +0.85
2023 146 43.8% −1.73
2024 134 57.5% +2.96
2025 116 56.9% +2.42
2026 100 46.0% +1.97

+1.18 bp per trade, positive in seven years of ten, t +1.06, comfortably inside the range that 78 coin flips produce. It is the kind of table people build strategies on, and it is exactly why the bar is set where it is.


The asymmetry, said plainly

Every concept we have tested on gold shows the same shape: the buy side ranges from neutral to mildly positive, the sell side loses, and the sell side loses more in the years gold rose fastest. Part of that is simply the market. But here, and in the liquidity-sweep test, the sell side keeps losing after gold's own rise is subtracted.

The honest summary is not "support and resistance works for buying". It is: over these nine years, mechanically selling gold into a prior high was a reliable way to lose money, at levels and not at levels alike, and buying prior lows was indistinguishable from doing nothing in particular.


What this page does not say


How we tested

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.


Request a backtest

Want a concept, a variation or another market tested the same way? Tell us what to test and where to send the result. It is free.

We only use your contact to send you this result.

Disagree? Bring results.

We publish numbers, not opinions, and we answer with numbers. If you believe a concept works and our test missed it, send us your exact rules, the market and timeframe, your data source and dates, the costs you charged, and your result with its sample size. We will run your version and reply with the actual statistics: every trade, the win rate, the result after costs, year by year. If your test holds up, we correct this page, dated, with credit to you if you want it.

Claims without results will not change a page. A screenshot of a winning trade is one trade; the tables here are thousands.

Send it to support@fxsynapseai.com with the subject “Research correction”, or use the request form above.