EA Overfitting: What Forward Testing Can and Cannot Tell You

Overfitting happens when parameter choices fit the history used to select them but fail to generalize. Reserve an untouched period, define evaluation rules before tuning and keep a record of every experiment—not just the winning settings.

Why searching more combinations changes the question

Suppose you test 200 parameter combinations and present only the best one. The selected result reflects both the strategy and the search process. A strong-looking winner may have benefited from features peculiar to that sample.

That observation is not proof that the strategy is useless. It is a reason to ask how the candidate was chosen and what evidence remains independent of the selection. Record the search ranges and number of trials along with the final preset.

Separate development from evaluation

MT5 provides a forward-testing option that reserves the later part of a historical interval and evaluates selected optimization results there. See the official optimization and forward-testing guide. This is a historical holdout test; it is different from running a demo account as new prices arrive.

For an illustrative research plan, develop on January–December of one year and reserve the next six months. These dates are an example, not a universal split. The available history, strategy frequency and changes in market conditions should inform the design.

Write the evaluation rules before looking

• Which measures will be reviewed besides profit?
• Which costs and position-sizing rules remain fixed?
• What behavior would make you reject or investigate the candidate?
• How will a low number of trades or incomplete data be reported?

Do not repeatedly adjust parameters after examining the “untouched” period and still describe it as independent. Once it influences your decisions, it has become development data. Keep that history in the record and seek fresh evidence for the next evaluation.

Challenge a narrow winning setting

Try nearby parameter values as a sensitivity exercise. If changing a threshold from 20 to 19 or 21 causes a dramatic reversal, investigate why. A wider region of similar behavior can be more informative than one isolated peak, although neither proves future reliability.

Then vary one execution assumption at a time. A candidate that depends on unusually favorable costs deserves closer inspection. Keep any deliberately stressed settings labelled so they are not confused with the original broker specification.

Use demo observation for a different question

Historical holdouts ask whether selected rules behave similarly on other history. Demo observation can reveal operational issues such as scheduling, logging and reconnect behavior. The MQL5 tester reference documents differences between the testing environment and normal operation.

Build a short research log with candidate, selection period, holdout period, assumptions, outcome and next action. Keep failed candidates too. Continue with execution-cost analysis when reviewing a promising result, and avoid treating either forward testing or a demo account as a profitability certificate.

Back to top