Some expert advisors work; the large majority don't. An EA is only as good as the edge it automates, and most are optimised until they look perfect on past data. The deciding test is whether performance holds on data the strategy was never fitted to.
The typical EA is built by adjusting parameters until the backtest equity curve looks beautiful. That process — curve fitting — guarantees a great past and says nothing about the future. Change the date range slightly and the magic disappears, because the settings were tuned to noise rather than to a repeatable market behaviour.
A genuine EA automates a behaviour that has a reason to exist — liquidity being taken, a session range breaking, an imbalance being filled — and it survives being tested on data it never saw. It also survives realistic costs. If a strategy's edge disappears once you add spread and slippage, it was never an edge.
Walk-forward analysis: optimise on one slice of history, test on the next unseen slice, roll forward, repeat. You end up with many out-of-sample results instead of one flattering in-sample curve. Our own sweep strategy passed 21 of 32 walk-forward folds — a real but modest edge, which is what honest results look like.
Ask for out-of-sample results, not a backtest. Ask whether costs are modelled. Ask for the losing periods. Ask how many trades the results are based on — thirty trades tells you nothing. Anyone unwilling to show the failures is showing you marketing, not evidence.
No EA reliably produces outsized returns. A good one automates a modest edge consistently; anything promising more is selling, not testing.
Almost always curve fitting, unmodelled costs, or both. A backtest tuned to past data doesn't generalise.
Hundreds, ideally across different market conditions. Thirty trades is well within the range of luck.