In plain language
In trading, curve fitting describes a strategy tuned so closely to historical data already seen that it captures noise and coincidences rather than a relationship that survives new data.
Calibration or excessive fitting
Choosing parameters is part of building many models. Risk rises when many variants are tried, only the best is retained, and the same sample is used to invent and judge the rule. Parameter calibration does not automatically prove overfitting.
What to observe
Record the number of trials, discarded rules, in-sample period, genuinely out-of-sample tests, costs, slippage, and stability across markets or windows. A smooth backtest alone does not prove curve fitting; selection procedure and out-of-sample fragility provide the relevant evidence.
Practical consequence
The issue is not an aesthetically “too good” curve, but an optimistic estimate created after many opportunities to choose. Overfitting covers techniques and limits; automated systems also require continued testing and monitoring after release.
Sources
- Bailey and López de Prado, How Backtest Overfitting in Finance Leads to False Discoveries — Explains why historically selected strategies may become statistical mirages out of sample.
- Bailey et al., The Probability of Backtest Overfitting — Proposes a method for estimating overfitting probability in backtest selection.
- U.S. SEC, Robo-Advisers — IM Guidance Update 2017-02 — Includes testing, backtesting, and post-implementation monitoring among algorithmic-code controls.