Original Backtest
The EA is first tested on the original historical market data under a defined scope: EA version, settings, symbol, timeframe, broker/data assumptions and historical window.
Methodology
AntiOverfit does not ask whether the original backtest was beautiful.
It asks whether the behavior survives when the path changes.
The method starts from a simple premise: a historical backtest is one observed market path. A robust Expert Advisor should not depend excessively on that single path.
Synthetic Market Paths
AntiOverfit generates synthetic market histories designed to remain statistically compatible with the original data while changing the exact path order and structure.
Metric Distributions
The EA is tested across synthetic worlds. Key metrics such as Profit Factor, Expected Payoff, Drawdown, Recovery and Trade Count are compared against the original result.
Robustness Score
The score summarizes how the original result compares against the synthetic distribution. It is not a profit forecast and not a guarantee.
Warnings
Warnings can override a superficial interpretation of the score. Examples include abnormal flat distributions, near-constant trade counts or suspicious metric behavior.
Scope Limitations
Every result is conditional. AntiOverfit evaluates robustness under the tested scope, not the universal quality of an EA under all possible settings and brokers.
The practical interpretation is simple.
High clean score: stronger evidence. High score with warning: caution. Limited or failed result: do not rely on the backtest alone.
StrongBehavior remains acceptable across worlds.Useful evidence, still scope-dependent.
AcceptableBehavior is not perfect, but survives enough variation.Requires normal caution and scope discipline.
LimitedThe original backtest may be too optimistic.Buyer or developer should verify further.
LowPath variation exposes substantial fragility.Do not trust the original curve alone.
Methodology matters because claims without scope are just marketing.
Read public records or apply the method to an EA before assigning trust to the curve.