Search for trading-strategy rules in your browser: how the AIStrategyMiner Strategy Miner works

AIStrategyMiner — Fri, 06 Nov 2026 16:35:34 +0000

The AIStrategyMiner Strategy Miner searches price history for entry rules and tests each candidate against a series of checks. It runs entirely in the browser. There is nothing to install.

This post walks through a run, and shows one number we measured that changes how we read the results.

What it does

You pick a market, a timeframe, a period and a trading direction.
The Miner tries thousands of random combinations of entry rules.
Candidates that pass the training period go through further checks.
The survivors are listed with their stats. In Pro, each one opens in the visual Builder and exports to MetaTrader 5 (MQL5) source, with more targets available.

A run in five steps

1. Choose a market. A currency pair (or gold, silver, indices and oil as add-ons), a timeframe such as H1, and 2 to 5 years of history.

2. Choose a style. Trend following (moving average, MACD, Donchian), counter-trend (RSI, Stochastic, Bollinger Bands), breakout (Bollinger Bands, Donchian), or all six.

3. Choose how deep to dig. Light, Standard and Thorough try 10,000, 50,000 and 200,000 combinations. Standard takes about a minute on an ordinary PC.

4. Start mining. Counters show trials tried and raw finds as the search runs. You can cancel at any time.

5. Read the results. The trial run lists each survivor with a star rating, its PF (gross gains divided by gross losses) and its trade count. Pro adds the equity curve, the largest peak-to-trough fall, and a pass or fail for every check.

The checks

The first 70% of the history is used for the search. The last 30% is used only for checking. A candidate must also survive:

a neighbour test: 10 variants with parameters shifted by 10–20%, at least 70% still ending in the black;
a Monte Carlo test: 200 runs with 10% of trades dropped, and the worst 5% must keep the peak-to-trough fall within 20%;
a cost test: still ending in the black with double the spread;
year-by-year checks: most calendar years must end in the black, and no single year may carry the result.

Duplicates are removed by shape, entry-bar overlap and month-by-month result correlation.

What we measured on gold

We ran the default settings on XAUUSD H1 with four years of data, 2,000 trials per run, ten seeds, and counted every survivor.

Price series	Passed the search period	Passed all checks
Real gold	7,845	55
Bars shuffled (same candle shapes, random order)	6,666	40
Shuffled, with the average drift removed	4,410	35

Totals over ten runs. On shuffled prices, where the order of moves carries no structure, 6 to 7 out of 10 of the real count still passed every check. The per-run counts overlap (real: 3–9, shuffled: 2–5), and the gap is not statistically clear with ten runs.

How we use this: the number of survivors alone does not tell you whether a rule captured something real. We treat the checks as a filter, not as proof, and we confirm any strategy on data the search never saw. In Pro, Leave out the last 12 months keeps the most recent year unseen by the search for exactly that purpose.

Free and Pro

Free: a trial search at a smaller fixed size, with stats and pass/fail for each candidate.
Pro: full search settings, the rule behind each candidate, the vault, and export to MetaTrader 5 and other targets.

Results describe past data only and say nothing certain about the future.

Try it

https://aistrategyminer.com/strategy-miner

This article was drafted with AI assistance and reviewed by the AIStrategyMiner team.