Optimisation is a mapping tool. Run it, then look at the whole surface: where are the decent regions, how wide are they, do they move between periods. Taking the single best cell is the classic route to overfitting.
Prefer a robust objective. Maximising raw return picks the luckiest path; maximising sharpe-ratio picks low-trade-count flukes; maximising the 25th percentile of returns across a neighbourhood of parameters and several data subsets picks something more likely to repeat.
Example: a grid of 400 combinations tested on 800 trades. The best cell shows a profit factor of 2.1, the median cell 1.15. A reasonable live expectation is closer to the median plus a little, not the maximum.
Related: grid-search, objective-function, parameter-plateau, overfitting