We generated thousands of gold strategies. Six survived. Here are the filters.

Every EA page says the strategy was "carefully developed and tested". That sentence carries no information. What tells you something is the process behind it: how many candidates were generated, how many were rejected, and what exactly killed them. This post describes the process we used for our gold project, filter by filter.

Step one: generate without opinions

We did not start from a market thesis. We started from systematic generation: thousands of candidate strategies built on XAUUSD data, exploring different entry logics, exit rules and timeframes. The point of generating at scale is to remove our own bias from the search. If you hand-build three strategies, you will fall in love with all three; if you generate thousands, you are forced to build machinery that rejects them. At this stage the results look wonderful, and that is precisely the problem. With enough candidates, some will fit the historical data by pure chance, the same way some coin-flippers in a stadium of thousands will land ten heads in a row. A beautiful equity curve at this stage proves nothing.

Step two: out-of-sample or nothing

The first real filter is out-of-sample selection. Each candidate is built on one portion of the data and then judged on a portion it has never seen. A strategy that only performs on the data it was fitted to is memorizing the past, not reading the market. This single filter removes the large majority of candidates, and it should. The rejection rate is the health check of the whole process: when almost everything survives your out-of-sample test, the test is too soft, not the strategies too good. If a seller cannot tell you how their strategy behaved on unseen data, you are looking at a curve-fit.

Step three: cut for correlation

Here is the filter almost nobody talks about. The survivors of out-of-sample testing tend to be siblings: variations of the same idea that win and lose in the same weeks. Ten correlated strategies are not diversification, they are one bet repeated ten times. So we compared equity curves and kept only strategies with low correlation to each other. Many good individual performers were cut at this step, not because they were weak, but because the portfolio already owned their behavior. This is the least intuitive filter, because you are deleting strategies that passed every other test. It is also the one that decides whether your drawdowns stack on top of each other or take turns.

What survived: six engines

After all filters, six engines remained. Each one runs on its own timeframe (M30, H1 or H4) and comes from a different logic family: breakout, volatility bands, parabolic SAR, momentum. Each engine opens every position with a fixed stop loss and take profit. No grid, no martingale, no averaging down, ever. Six different answers to the same market, deliberately chosen to disagree with each other, so that a regime that starves one engine still feeds another.

Combined in a single MT5 backtest, the six engines produced 1,425 trades over seven years (2020 to 2026), a net profit of 4,512 USD with a 0.01 lot per engine, a profit factor of 1.67 and a maximum drawdown of 326 USD, with all seven years closing positive. To be clear: these are backtest results, and past performance does not guarantee future results.

Why we publish the filters

Because the filters are the product. Anyone can show a green curve; the question is what was thrown away to get it. The same three questions work as a checklist for any EA you evaluate on the Market, ours included: how many candidates were discarded, was the selection done out-of-sample, and are the components actually uncorrelated? A seller with a real process can answer all three without hesitation. A seller without one will answer with an equity curve.

The six engines will ship together as Wonder Gold Portfolio, one EA running all of them on a single XAUUSD chart. It is coming soon. If you want to see how we already trade gold with fixed rules and mandatory stops, Wonder Gold is live on the Market, and if you prefer to start with something free, Wonder 8 costs nothing and follows the same principles on USDJPY.