Parameter Sensitivity
How to evaluate Turtle-style parameters without overfitting to the best historical result.
Parameter sensitivity asks whether a system works as a rule family or only at one lucky setting.
The wrong question
The wrong question is: which number produced the highest backtest return?
That approach often selects noise. A parameter that was perfect in history may fail because it captured accidental features of the test period.
The better question
Ask whether nearby values behave reasonably. For example, if a breakout window of 55 days looks good but 50 and 60 days collapse, the result may not be robust.
What to record
- tested parameter ranges;
- out-of-sample periods;
- drawdown behavior;
- trade count;
- sensitivity to fees and slippage;
- reason for choosing the final rule.