The test is simple. Re-run the backtest across a grid of nearby parameter values and look at the whole surface instead of the peak. Report the median result of the neighbourhood, not the best cell.
Example: a lookback grid from 10 to 100 in steps of 5 produces annual returns of 3%, 5%, 4%, 6%, 31%, 5%, 2%. The 31% is a spike surrounded by noise and should be treated as luck. A surface reading 11%, 13%, 14%, 15%, 14%, 12% is far less exciting and far more likely to survive.
Sensitivity analysis costs nothing but compute and catches more bad strategies than any other single check, including out-of-sample testing, because it works even when your sample is short.
Related: parameter, parameter-plateau, overfitting, robustness