@ea_architect

Fair challenge — my heuristic: re-run each window using the neighboring window's optimal parameters. If the OOS result stays within 20%, the drift is tolerable; the strategy is not standing on one exact number. This EA passed 7 of 8 of those swaps. I call it the parameter-neighbor test — not…

I keep a 10-point overfitting checklist in the Backtesting category (posting in English, hope that's ok). For your case the top three to check first: profit factor above 3 in-sample, parameter "islands" where only one input combination works, and no untouched out-of-sample period. PF 2.3 isn't…

The every-tick-instead-of-new-bar wiring is the single most common generator bug I see, across every tool. My habit: attach any generated EA to the visual strategy tester and just watch it trade one month of ticks. Logic errors become obvious in about two minutes. For what it's worth, my…

Agree completely. 50ms vs 5ms changes nothing when your average trade lasts 3 days. Spend the savings on better data or a second demo account for forward tests — both pay better than shaved milliseconds.

Also: real ticks versus generated ticks matter more than the % label, and variable spread matters more than both. Variable spread alone has killed most M1 scalpers I've seen that looked brilliant on fixed spread.

The survivorship bias is the real killer for public data. The 10% that survive get screenshotted and sold; the 90% get silently deleted. Any survival statistic built from vendor marketplaces is measuring the winners' podium, not the race.

Honest timelines: FXDreema gets you a first working EA in a weekend, then you plateau around month 3 when you want custom logic. MQL5 from zero takes 2-3 months to competence, but there is no plateau. Your strategy (trend + ATR stop) is squarely in FXDreema territory. If you already know you'll…

Class-per-symbol plus a timer loop it is. Thanks both — I'll post the refactored skeleton here when it's done in case it helps the next person.

Anchored vs rolling matters too. My filter: anchored WFO, 8 windows, and reject any strategy where more than 2 windows come in under PF 1.2 out-of-sample. Strict, but my live results got boring in a good way.

We've all paid this tuition. The honest framing that finally fixed my process: a backtest is a hypothesis, not a result. The forward test is the experiment. Most retail flow skips the experiment entirely.