Research
Before an idea goes on sale we run hundreds of tests, and only the ideas that fully pass reach you. Below are six findings backed by numbers.
60.5 %3-month windows that ended in profit (of 2,223)
We split the gold robot's 2020-2023 and 2024-2026 real-tick tests into rolling 91-day windows: 60.5 % of the 2,223 windows ended in profit. With a stop loss on every trade and 1.2 % risk per trade, the maximum drawdown in the 2020-2023 test was only 3.26 %.
Our figures are measured on real ticks, so you know exactly what you get.
PF 1.90 → 1.581-minute model vs real ticks, same robot, 2020-2023
When price touches the trigger there is a spread between ask and bid. The 1-minute OHLC model does not simulate it, so for the same robot and the same period it showed 24 % more profit (6,503 vs 5,254 USD).
We publish only real-tick figures with the real spread, and always state the tick model.
+21 % / +33 %profit / max drawdown, risk 1.2 % → 1.6 %
Raising per-trade risk from 1.2 % to 1.6 % in the 2020-2023 test lifted profit 21 % (5,254 to 6,350 USD) but lifted the maximum drawdown 33 % (3.26 % to 4.34 %). That is why 1.2 % gives the better balance of profit and drawdown.
We do not chase profit by raising your risk.
×2.2 and −20 %adding the 08:00 candle: 2024-2026 and 2020-2023
Also trading the 08:00 candle multiplied profit 2.2 times in 2024-2026 but cut it 20 % in 2020-2023. This is how we catch and drop settings that shine in one period and fail in the next.
Every setting is tested separately in a trending and a trendless era.
0 tradesin the first two weeks of September
Our gold robot opened no trades in the first two weeks of September: gold was below its 200-day average (≈4,400 vs 4,536), so the buy-only strategy waited on purpose. Replaying those days with the 1-minute model produced 2 trades, both losers: by waiting, the robot avoided them.
The robot now writes to the Journal why it skipped each candle.
13 familiesof faults that only happen on a live account
In a backtest the terminal never restarts, the connection never drops, the broker never rejects an order and the server clock never drifts. On a live account all of these happen. We list 13 families of such faults (restarts, duplicate instances, account switches, lost connection, trading permissions, clock, missing data and more) and check the robot's code against each one.
We check and fix the faults that only show up on a live account before they reach you.
Unless stated otherwise: XM GOLD#, “Every tick based on real ticks”, $100,000 account, 1.2 % risk per trade, the 08:00 and 16:00 H4 candles. Get new research on our Telegram channel →
Method
An idea passes these stages before it goes on sale. Only a robot that clears all six reaches you.
Before a test runs we write down what is being tested and which result kills the idea, so we cannot move the goalposts after seeing the numbers.
Many variants are tried quickly on the 1-minute model to drop the clearly weak ideas. Figures from this stage are never published.
The survivors are re-tested on the broker's real ticks with the real spread. The figures we publish come only from here.
Tested separately in a trending (2024-2026) and a trendless (2020-2023) market, so a robot has to hold up in both.
The code is checked against 13 families of live-only conditions: terminal restarts, lost connection, duplicate instances, account switches and more.
After every code change the new build is compared trade by trade with the previous one, and it ships only once no decision changed by accident.