We tested 288 versions of the gold breakout strategy on 18 years of data

The short answer

The gold breakout strategy that many popular gold EAs use did not have a real edge in our test. Versions that looked great in the past lost money later. The good results in many gold backtests come mostly from gold's big rise since 2017, not from the strategy.

The idea

When the price of gold breaks above its recent high, buy, because it may keep going up. When it breaks below its recent low, sell. Every trade has a stop loss (an automatic exit if the trade goes wrong) and often a take profit (an automatic exit at a target). Many of the best-selling gold EAs on the Market work this way and advertise "no grid, no martingale, a stop loss on every trade".

How we tested it

We let a computer trade these rules on 18 years of real gold prices, from 2008 to 2026, with trading costs included. The strategy has many settings, for example how far back to look for the high and how far away to put the stop loss, so we tried 288 different combinations.

Then we did the most important thing in any test: we split the history in two. We picked the best settings using only the first part (2008 to 2016), and then checked how those same settings did in the second part (2017 to 2026), which they had never "seen". That is exactly your situation when you buy an EA: its settings were chosen on the past, and you trade the future.

What happened

• Only 14 of the 288 versions made money in both parts. Pure luck would give about 15
• The typical version lost about 12% a year
• The 10 best versions of the first part made 9.5% to 11.8% a year there. In the second part every one of them lost money, between 0.3% and 8.9% a year, and at their worst they were down 42% to 65% from their highest point
• The versions that did best in the second part (up to +9% a year) only bought, never sold, and they lost money in the first part

Why? From 2017 to 2026 the price of gold went from 1,150 to 4,155 dollars. Almost anything that only buys looks good in that market. From 2008 to 2016 gold only went from 833 to 1,150, with a long fall in between, and the same buy-only versions lost.

What it means for you

• If a gold EA's backtest starts in 2020 or 2023, it mostly shows gold going up. Ask the seller for results from before 2017, or test it yourself on 2008 to 2016
• Be careful with very exact settings. If an EA only works with one precise number and stops working when you change it a little, it was fitted to the past
• Be careful with very high win rates. Even the best versions in our test won only 44% to 46% of their trades. A gold EA that wins 70% to 90% of its trades often has a very distant stop loss or a hidden grid

Our own TrueRisk Strategy EA has a gold breakout part too, and its page shows both periods: -8.0% from 2017 to 2022 and +121.5% from 2023 to 2026: https://www.mql5.com/en/market/product/200817

Whatever gold EA you run, protect the account with a loss limit. TrueRisk Account Guard closes all trades before your daily or maximum loss limit is broken: https://www.mql5.com/en/market/product/200814

Details for those who want them

• Data: XAUUSD hourly (H1) from Dukascopy, January 2008 to September 2026
• Entry: buy stop at the highest high of the last 10, 20 or 50 hours, sell stop at the lowest low (or buys only)
• Stop loss 1.5 or 3 times ATR(14) (ATR = the average size of one candle); take profit at 2 times the risk, or none; stop moved to the entry price after a profit equal to the risk, or not; trailing stop or not; trades closed after 24 or 120 hours; with or without a 200 EMA trend filter
• 1% of the account risked per trade, one trade at a time
• Costs: 0.02% per side, plus 5% a year swap (the fee for holding a trade overnight)
• If the stop loss and the take profit were both hit inside the same hour, we counted the loss
• 88 of the 288 versions made money in 2008 to 2016, and 49 in 2017 to 2026

Have a strategy, an indicator or an EA you want tested like this? Write it in the comments.

Next week: support and resistance levels on gold, real levels against random ones.