The condition where a strategy's rules have fitted the noise of one historical sample rather than a repeatable relationship, so the backtest describes the past instead of predicting anything.
The tell is fragility. Move a parameter slightly and the result collapses, shift the sample by a year and the curve changes shape, or the strategy carries a dozen filters that each entered to repair one bad month. A real relationship survives small changes in how you look at it.
symptoms: parameter fragility · sample sensitivity · filter accretion · in-sample and live divergence
Overfitting mostly does not feel like cheating, and it is not one. It is the default outcome of iterating on the same data, because every look is a chance for noise to influence a decision. The searching does the damage, not the intent: enough variants evaluated on one sample will produce something impressive whether or not any edge exists, a result formalised by Bailey, Borwein, Lopez de Prado and Zhu in their work on backtest overfitting.
The defences are procedural rather than clever: hold data out and touch it once, fix the rules before the holdout is opened, record every trial, and deflate the headline figure for the trial count you recorded. Each defence exists because a specific habit defeats it, which is why the checklists in this section read like a confession list.
The cost asymmetry is worth stating plainly. An overfit strategy looks identical to a sound one in every backtest report. The difference only becomes visible with money at risk, and by then the evidence you needed was the evidence the search consumed.
Covered in depth in the Strategy research fundamentals pillar hub.