Is my EA over-optimised? The one-step test
EA Trade TesterDemo labHow we testAll guidesFor AI agents
An over-optimised EA works at the exact settings it was tuned on and falls apart a little way off them. A simple test is to move each setting one small step and see whether the result survives. What our test cannot do: it moves at most six numeric settings of our rebuilt copy, one at a time, so it cannot explore every combination of your inputs, and a setting that is not a number (a choice of indicator, a time zone, a yes / no switch) is not moved at all.
How does the neighbour map move a setting?
For each numeric setting of the rebuilt rules (a period, a level, a stop or target distance) the lab runs the strategy twice more: once with the setting one step lower and once one step higher. A step is 10% of the value, and at least 1 for a whole number, so a 14-period setting is tested at 13 and 15, and a 50-pip stop at 45 and 55. Up to six settings are moved, one at a time, everything else unchanged.
When is a result called FRAGILE?
We compare each neighbour with the setting you sent, on the two later data periods and with extra costs. If your setting makes money there but one step away loses money, and the gap is more than 0.05 R a trade, the lab writes FRAGILE: a small change ruins it and names the setting. A fragile result cannot be PROMISING: the lab gives FAIL with that reason, because a real edge rarely sits on one exact number.
A worked example
Say your settings make +0.06 R a trade with extra costs. Period 15 of a moving average gives -0.02 R, and period 13 gives +0.04 R. Moving to 15 turns a profit into a loss, a gap of 0.08 R, so the result is FRAGILE at period = 15. If 15 had given +0.01 R instead, nothing would be flagged: smaller, but still on the same side of zero.
Why not just optimise again on more data?
Because every extra optimisation is more tries, and the best of more tries looks better by chance alone. The neighbour map asks a different question: not "which value is best?" but "does the result depend on one value?". It costs no extra versions on the luck bar, because the neighbours are measured, not picked.
What does a smooth neighbour map tell me?
That small changes in the settings do not change the sign of the result. It does not say the strategy has an edge: the luck bar, the extra costs and the drift control still decide that. It is one check among several, not a pass on its own.
Related pages
- How many backtests? Overfitting explained
- Is my EA backtest real or luck?
- Does my EA make money every year?
- How we test
- All EA testing guides
- EA Trade Tester home: test your EA, bot, indicator or report
Questions people ask
How many settings are moved?
Up to six numeric settings of the rebuilt rules, each one step either side, one at a time.
Does a fragile result mean the EA is useless?
It means the profit depends on one exact value, which is the typical sign of tuning to the past. Test it on new prices before trusting it.
Are the neighbour runs counted as versions you tried?
No. They are measured for every test the same way, and none of them is chosen as the result.
Is this financial advice?
No. It describes a test of a rebuilt copy. Not financial advice.