Overfitting

Overfitting is the failure mode where a strategy is tuned so closely to a specific historical dataset that it models the noise rather than a durable, repeatable edge. It typically comes from too many parameters/filters, excessive re-optimisation, and a small sample. The standard defences are validating on out-of-sample data, walk-forward analysis, forward testing, and keeping the rule set simple. Fitting a strategy to data is normal; overfitting is fitting it so hard that it no longer generalises.

In plain English — Overfitting (or "curve-fitting") is when you optimise a strategy until it looks almost perfect on historical data — but it has really just memorised the random noise of that exact period. The moment it meets new data, the magic disappears. A backtest that looks too good is the classic warning sign.

A system with eight filters shows a flawless equity curve on 2019–2021, then loses money in 2022. The filters were tuned to the past, not to a real edge — a textbook overfit.