A finished backtest shows the EA that survived development. It does not show the versions that were discarded, the filters that were tried, or the losing periods that prompted another change. That missing history matters when an attractive equity curve is the main reason to consider a purchase.
A high profit factor (PF) or a small drawdown is welcome, and neither is, by itself, a sign of overfitting. The difficulty is working out how much of the result reflects a repeatable trading idea and how much reflects decisions made after inspecting the same data.
When a sensible change is not yet a convincing improvement
Consider an EA that loses heavily during a particular trading window. The developer excludes those hours and the backtest improves. A weak weekday is removed next, followed by a filter that eliminates another cluster of losses. Each change can be explained, and each produces a better report.
The improvement is real within that sample. What remains uncertain is whether the exclusions capture something useful about the market or merely remove losses that happened to occur there. The data that suggested the change is also being used to approve it.
This is why development and evaluation need different roles. Experimentation is necessary, but a better result on the sample used to make a change is not independent confirmation of that change. Repeatedly refining the same history can make the report more impressive without settling the question of robustness.
AI-generated illustration of how historical results influence decisions; not a product performance record.
Selection uses information too
Parameter optimization is only one way to fit a strategy to history. Entry and exit rules, trading hours, filters and symbol choices can all be influenced by the results. So can the choice of the final EA.
Suppose 100 candidates are tested and the one with the highest historical PF is selected. Its parameters might never have been edited. Even so, the backtest determined which candidate survived. A larger search offers more opportunities to find a genuine edge, but also more opportunities to find an unusually good fit to that sample.
For that reason, “no parameter optimization” is an incomplete description of the selection process. It helps to know what was compared, which data influenced those comparisons, and whether anything remained outside the search for a later check. The candidate count does not invalidate the result. It tells us how widely the developer searched before finding it, which helps put the winning backtest in context.
AI-generated illustration of how historical results influence decisions; not a product performance record.
What to ask beyond the report
When reviewing an EA, separate the evidence into what happened before the final configuration was chosen and what happened afterward. Before selection, historical data may have influenced rules, settings and candidate rankings. After selection, the useful evidence is how the fixed candidate responded to data or conditions that had not already supplied the same answers.
A practical starting point is to ask the developer which periods were used for construction and selection, what was fixed before later testing, and whether any rules changed after those later results were seen. You do not need the proprietary trading recipe to understand those boundaries.
AI-generated illustration of how historical results influence decisions; not a product performance record.
Different checks then answer different questions. A separate period examines behavior outside the development interval. Nearby parameter values reveal dependence on a narrow setting. Higher spreads or other costs show how much execution margin remains. Monte Carlo analysis can explore how the drawdown changes when the sequence of trades is disturbed.
These tests do not make the future predictable, and a test name is not a substitute for its settings and results. Their value is that they challenge assumptions which the final equity curve cannot reveal by itself.
Before relying on the report, ask which decisions used the development data and which tests were performed only after those decisions were fixed. You do not need every detail of the trading logic, but you do need to know whether the final chart was used to develop the system or to evaluate it.
Product details and testing conditions
When reviewing our EAs, apply the same distinction between a headline result and the process behind it.
Specifications, published historical results and operating limits: EdgeDriven Gold Portfolio — XAUUSD · EdgeDriven Dollar Yen Portfolio — USDJPY.
The examples in this article explain evaluation methods; they are not test results for those products. Historical simulations do not guarantee future results. Leveraged trading can cause substantial losses.
