Plot performance against a parameter and you get a landscape. A plateau is a wide flat area; a spike is one tall cell with cliffs on both sides. Choosing the top of a spike guarantees disappointment, because live data will land slightly off the peak.
The practical rule is to choose the centre of the widest acceptable plateau rather than the highest point anywhere. If lookbacks 40 through 90 all return 9% to 13%, take 65 and expect around 11%.
A plateau is evidence but not proof. Nearby parameter values share most of the same trades, so their results are correlated. A plateau of forty settings is not forty independent tests.
Related: parameter-sensitivity, parameter, overfitting, robustness