The Strategy Tester optimiser runs your EA across hundreds or thousands of parameter combinations and tells you which set scored best on past data. Used with care, it narrows a sensible range of settings. Used carelessly, it hands you the single luckiest combination in the sample and calls it a strategy. The whole skill is telling those two outcomes apart.
What the optimiser does
In MT4's Strategy Tester you tick Optimisation, then in the Inputs tab you give each parameter a Start, Step and Stop. The tester runs a backtest for every combination in that grid and ranks the results by one metric. Optimise three parameters with ten values each and that is a thousand backtests — at least one of which will look extraordinary by chance alone.
A careful walkthrough
- Optimise few parameters at once. Two or three that matter, not every field. More dimensions means more ways to fit noise.
- Rank by the right metric. Net profit alone rewards reckless sets; read profit factor and maximum drawdown beside the profit, or use a measure that already accounts for risk.
- Use good data. "Every tick" modelling and quality history, or the result describes a market that never existed.
- Set honest ranges and steps — wide enough to learn something, coarse enough that you are not fitting to the third decimal.
- Read the whole surface rather than the top row alone.
Choose a plateau, not a peak
This is the part most people skip. Sort the results and you see a top score, often a sharp spike where 14 works brilliantly and 13 and 15 are mediocre. That spike is fragile; the live market will not land on exactly 14. Look instead for a broad plateau where a whole neighbourhood of values all did reasonably well. A setting surrounded by other good settings is far likelier to hold up than a lone peak.
The best row in an optimisation is usually the luckiest, not the best. If moving one parameter a single step either way collapses the result, you have found an accident, not an edge.
The over-fitting trap
Every extra parameter, every finer step, every re-run tuned to lift last year's curve pulls you towards curve-fitting: a set that describes the past perfectly and predicts nothing. The optimiser cannot warn you, because a fitted set is exactly what you asked it to find.
Validate out of sample
Optimise on one slice of history, then test the chosen set on a later slice it never saw. Better, run a walk-forward analysis that repeats this rolling forward in steps. If the set only works on the data it was tuned on, it is not a strategy.
A worked example
An optimiser hands you a set showing £9,200 profit on two years of GBPUSD with a tidy curve. Run it forward on the next six months it never saw and you get £180, then a £700 drawdown. The backtest was real; it was also fitted. Had you picked the plateau set — £6,100 on the same two years, but stable one step either side — the forward test would more likely have survived.
The Karnek note
Karnek shows the live results of whatever set you finally choose, read-only from your terminal, so you can hold the real account against the backtest that sold you on it. It never optimises, changes a parameter or places a trade. Seeing the two side by side is usually where the over-fitting shows itself.
Written and reviewed by the Karnek Research team. Last updated August 2026.
Educational content only - not financial advice. Past performance does not predict future results. Trading carries significant risk.