An Expert Advisor (EA) is a program that trades for you on MetaTrader, following rules written in code. I develop them, run them on my own accounts, sell some utility ones on MQL5 and publicly dismantle others — including my own. This guide tells you what the "miracle EA" industry needs you not to know.
The fundamental theorem of automation
Automation executes; it doesn't invent. If the rules have an edge, the robot executes it with perfect discipline — no fear, no fatigue, no revenge trading. If the rules have no edge, the robot loses with the same perfect discipline — just faster than you, because it never hesitates.
Everything follows from this: the question "does this EA work?" is always the question "do these rules have an edge, and how was that demonstrated?". The software is the garnish.
What automation genuinely solves
- Execution errors: wrong sizes, forgotten stops, duplicate orders. Code doesn't botch calculations at 2am.
- Psychology: an EA doesn't average losers "because it has to bounce", doesn't widen stops, doesn't revenge-trade. It's constraint engineering in its final form.
- Limit discipline: daily loss limits and max drawdown can be written into code with a non-negotiable latch — the guardian module of my systems.
- Measurability: every decision is logged; the system's statistics are a fact, not a memory.
- Speed and constancy: 24/7 monitoring (crypto), millisecond reactions where needed (my BTC bot), identical execution on trade #1 and trade #1,000.
What automation can never do
Create an edge from nothing. Predict the future. Turn a pretty backtest into real money if the backtest was overfitted. And above all: relieve you of responsibility — a production EA must be supervised, because markets change and a system that changes rhythm is a system changing nature (my gold EA went from 60 to 250 trades a year when the underlying tripled: it was becoming a different system, never tested by anyone).
Buying an EA: the anti-scam checklist
The EA market lives on the fact that almost no buyer runs these checks — each takes one minute on a public Myfxbook page:
- Real or demo? A demo track record proves nothing: instant fills, laboratory spreads, zero slippage.
- Avg W/L vs win rate: if the average loss is 5x the average win, the break-even win rate is 84% — and "90% winning trades" becomes a sentence, not a virtue. Calculator here.
- Pip expectancy: if it's negative while the currency gain is splendid, size is being manipulated (progressions, grids).
- Z-score: strongly negative = wins in correlated clusters = grid/martingale. Not a hint: a confession.
- Drawdown and its date: a "small" DD with no hostile regimes crossed is lack of testing, not merit.
I demonstrated it live on my own +641% demo account: the full anatomy of the deception.
Building your own: the real path
- Hypothesis with declared edge source — one sentence: who pays, why they should keep paying.
- Lab verification (Python): does the effect exist gross? Does it survive costs? The full pipeline with free data. Here — correctly — most ideas die, in hours instead of months.
- Formal spec: every discretionary concept translated into a measurable threshold — the full case study.
- MQL5 implementation with modular architecture and risk guardian — the mistakes to avoid. With AI as the implementer, the method and prompts are here.
- Honest backtest (real ticks, real costs, in-sample/out-of-sample) and trade-by-trade analysis.
- Demo or shadow mode with a promotion criterion written first — my standard: 30 trades with positive expectancy before one real euro.