从零写出你自己的 MT5 EA

Your First EA Places No Trades

Your first Expert Advisor should read the market and publish what it thinks — nothing more. Build the signal-only EA that fills our signals page, from its real source, and learn why every read uses the last closed bar.

One File Decides, Another File Acts

Where you draw the line between deciding and trading is the most consequential decision in an EA. A strategy module you can call with plain numbers, a verdict that explains itself — and the failed strategy that proved it was worth it.

The Numbers Are Not in the Code

A backtest belongs to a set of values, not to the source that produced it. Input schemas, three kinds of version, why every number should carry its reason, and the .set file format trap that turns generated presets into junk.

Every Failure Must Fail Toward Zero

Sizing is the only code in an EA that can end the account in one order. Lot size from real tick values, why risk-based sizing must refuse rather than round up, the floating-point bug that eats a whole volume step, and what happens when the broker widens the stop you already sized for.

A Bad Day Is Not a Bad Strategy

Three layers decide whether an EA may trade at all, watching three different lengths of time: this bar, today, and this quarter. Session and spread gates, account-level caps that count currencies rather than symbols, and the rolling profit factor that notices when a regime edge stops working.

A Backtest Is Not Evidence

The MT5 strategy tester will always find settings that would have worked. The nine-rule gate we run our own presets through, the two rules we had to add after our own results slipped past the first seven, and why every rejected tune stays published.

The Track Record Starts at One

Live is the one window you cannot re-run or retune. The bug that made every backtest run a different program, a reporting pipeline that cannot lose a trade, the divergences nobody warns you about — and our own record, published from its first trade rather than its thirtieth.

Retuning Is Not Diagnosis

When a result disappoints, the optimiser will always find better parameters — which is why it is the wrong first move. Measurement artifacts, clock assumptions, control arms, improvements that are only leverage, and the case where the losses are the mechanism.

Do Not Test This Again

Pairs, sessions and timeframes are finite sets, and finishing one is a result. How a timeframe curve tells you the axis is closed, how four studies of scheduled events composed into a single rule, and why a rejection note should say what would have to change before you test it again.

What We Actually Know

The balance sheet after nine modules of method: one mechanism edge across four pairs, four instances explicitly labelled weaker, eighteen rejections, a kill switch that trims rather than rescues, and the list of things we still do not know.