Over the past few weeks I ran a research program on gold (XAUUSD). It tested more than 400 strategy variants: popular YouTube setups, classic book strategies, chart patterns and open-source EAs. Almost all of them failed. This post shows what failed and why, so you can skip the same dead ends.
How I tested
I wrote the rules down before looking at any result:
- Data: gold M5 with real bid and ask prices (2016–2026), gold H1, and daily bars for 12 markets (2000–2026).
- Costs: the real spread, $0.07 per ounce commission, and slippage on stop fills.
- Settings fixed in advance: every test's variants were written down before the run, at most 24 per test.
- Two halves: 2016–2021 and 2022–2026. A real edge should work in both.
- Random entries: the same stops and exits, placed on random bars. If random entries do as well, the signal adds nothing.
- A luck ceiling: with hundreds of tests, some look good by chance. So a result needed t ≥ 3.3, not the usual 2. (t measures how far a result sits from zero compared with its noise.)
Results below are in R. 1R is the risk on one trade, so −0.20R means losing 20% of your risk per trade on average.
What failed
These are my translations of each strategy's public rules, tested on the data shown. The authors may trade them with discretion or on other markets.
| Strategy | Tested on | Result |
|---|---|---|
| Andrea Unger hourly breakout | Gold H1, 2016–2026 | 4,340 trades, 0.000R. Random entries in the same hours did better. |
| Larry Williams volatility breakout | Gold M5, 2016–2026 | About 230 trades a year, 0 to −0.03R. Every variant lost in 2016–2021. |
| Linda Raschke "Holy Grail" | 12 markets, daily, 2000–2026 | As written it almost never fires: 18–40 trades in 25 years. Relaxed: 407 trades, −0.20R. |
| Rob Hoffman inventory retracement bar | 12 markets, daily | −0.27R per trade |
| Qullamaggie breakout | 12 markets, daily | 690 trades, −0.13R. Random entries with the same stop did better. |
| SMB "9 EMA" continuation | Gold M5 | 844 trades, −0.19R, 27% win rate |
| An SMC "first candle" rule from YouTube | Gold M5 | 1,454 trades, −0.20R. A coin-flip direction did the same. |
| An open-source FVG gold EA | Gold | Claimed +48.7% in 6 months. Lost −0.13R per trade on real data. |
| Malaysian SnR | Gold H4 | −0.10R to −0.23R in every variant |
| Opening-range and Asian-box breakouts | Gold M5 | An open-source gold ORB EA failed. The Asian box made +0.035R: real, but too small to trade. |
| Bulkowski rectangle top and Eve & Eve double bottom | 12 markets, daily | Only 30–120 trades in 25 years, at zero or below |
| A machine-learning filter choosing which trades to take | Gold and indices | Out-of-sample skill score (AUC) 0.49–0.54. A coin flip scores 0.50. |
| Daily trend following | 12 markets, 2000–2026 | Sharpe 0.00 after overnight swaps |
Five lessons
- Most intraday ideas are zero before costs. Costs then make them negative. My next post is about this.
- Test against random entries. Gold rose from about $1,060 to $4,260 over my data, and that lifts almost any long strategy. One free gold breakout EA made +0.21R per trade, but random bars with the same exits made +0.14R. Most of its profit was the bull market.
- Check both halves. Many gold breakouts lost in 2016–2021 and only worked in 2022–2026.
- A high win rate means nothing on its own. One gold EA won 76% of its trades and still lost money. Its stop was 1.2 and its target 0.35, so it needed 77% just to break even.
- Rare patterns can't be proven. Thirty trades in 25 years is a story, not evidence.
What I do differently now
I write the rules and the test plan before I see any result. I count every test I run, because more tests mean more lucky-looking results. And I check costs before I code anything.
Not everything failed. But anything that passes still needs a live forward test before I'd call it an edge. If you trust a gold strategy, test it this way first: real costs, both halves, and random entries.
