Direct answers on walk-forward testing, profit factor, curve-fitting, spread modelling and whether expert advisors actually work — grounded in our own measured results.
Because the backtest was optimised until it looked good on data it had already seen. Curve-fitting, ignored spread, and survivorship in the parameter search all inflate results. A strategy that has never been tested out-of-sample has not been tested.
You optimise on one block of data, then test on the untouched block that follows, then roll the window forward and repeat. Every result is out-of-sample. It is the closest offline approximation of trading a strategy you have not yet seen the future of.
Above 1.0 is profitable; the further above, the more margin for error. Treat anything extreme with suspicion — a profit factor of 3 on a few dozen trades usually indicates curve-fitting rather than an edge, especially if costs were not modelled.
More than most people run. A few dozen trades cannot separate skill from luck at any realistic win rate. Look for a sample large enough that a handful of outlier trades cannot flip the result, and check what happens when you remove the best few.
Change a parameter slightly. A real edge degrades gently across neighbouring values — a curve-fit collapses. Choose settings from a broad plateau of decent results, never the single peak, and confirm on data the optimiser never touched.
Always, and it changes conclusions. We subtract 30 points round-trip spread subtracted from every trade on XAUUSD M15. Strategies that looked viable before costs stopped clearing them afterwards — which is exactly why we retired one of ours rather than keep publishing it.
Some do, most sold online do not. The distinguishing feature is not the logic but the evidence: out-of-sample results, modelled costs, and published losing periods. An EA advertised only with an equity curve and no drawdown figure is telling you what it omits.
The same discipline as manual trading, applied consistently: an amount whose worst realistic losing streak you can sit through. Automation removes the emotion from entries, not the maths of ruin.
Percent risk keeps exposure proportional as the account changes, which compounds gains and softens drawdowns. Fixed lots make results easier to compare across tests. Many traders develop on fixed lots for clean comparison, then deploy on percent risk.
Spread, execution and the tick data the tester uses all differ. Results built on synthetic ticks are especially fragile. Re-run on your own broker's data before trusting any published figure, including ours.
Larger than most people plan for. The question to ask of any system is not its return but the worst peak-to-trough decline in testing — and then whether you could hold position size through twice that.
Long enough to accumulate a meaningful sample, which for most intraday systems means months rather than weeks. Forward results that disagree with the backtest after only a handful of trades are usually noise, not proof either way.