A perfect fit can be a warning
Draw ten scattered points on a chart. A smooth straight line may capture the broad relationship while leaving some errors. A sufficiently wiggly curve can pass through every point exactly. On the old sample, the wiggly curve wins. On new points, it may fail badly because it learned every accidental bump.
That is overfitting: treating noise as if it were repeatable structure. Market data is especially vulnerable because it contains many variables, changing conditions and a great deal of randomness.
How research becomes too tailored
Rules become more flexible as filters, thresholds and exceptions are added. A researcher might adjust dates, sectors, indicator counts and strength levels until the old result looks impressive. Each choice gives the method another chance to match quirks in the archive.
Complexity is not automatically wrong. A detailed rule can reflect real market structure. The warning is that every extra degree of freedom needs stronger evidence and a convincing reason that existed before the outcome was inspected.
Unseen data applies pressure
One defence is to freeze the rule and test it on data that did not influence its design. Simpler models, cross-validation, sensitivity checks and penalties for complexity can also help. None creates certainty, but each makes memorising the past harder.
The unseen period must genuinely be unseen. Splitting an archive after the researcher has already explored the full period does not restore innocence. The later slice can show chronology, but it is an exposed replay rather than clean validation.