Parameter optimization overfitting is the default outcome when you search a large space on a finite history. Hyperopt overfitting freqtrade is a popular example, but the same issue exists in any grid search.

Rules that help

Reduce degrees of freedom: fewer parameters, simpler rules
Use walk-forward protocols (What is WFA?)
Report stability across neighborhoods (Parameter sensitivity, PSI)

Why optimizers find noise

Grid search, Bayesian search, and genetic algorithms all share one property: they maximize a score on a fixed dataset. If the dataset contains randomness, the optimizer will happily absorb it as long as your objective function rewards it. That is why the winning parameter vector is rarely the same as the vector that would win on next year's data.

A practical workflow that reduces harm

Split time before you choose what to optimize, and keep a final segment untouched.
Limit search breadth: fewer free parameters beats a huge search space with weak validation.
After you pick a candidate, run forward evaluation with realistic fees and slippage (Cost drag).
If you use Freqtrade, treat hyperopt output as a candidate list, not a certificate (Hyperopt).

When optimization is still worth it

Optimization is not evil. It is a discovery tool. The mistake is stopping at the leaderboard row without a second stage that is designed to reject most discoveries.

Count your trials honestly

If you try many parameter sets and only report the winner, you are running a multiple-comparison problem.

Track how many distinct candidates you evaluated, and treat the search budget as part of the evidence (PBO, Data snooping).

Separate discovery from confirmation

A clean split is:

Discovery: broad search allowed, but conclusions are weak
Confirmation: narrow tests on untouched time with locked rules

If discovery and confirmation share the same segment, your "confirmation" is theater.

Prefer simpler models when scores are close

If a complex parameterization beats a simple baseline by a small margin on noisy metrics, default to the simpler model unless stability evidence is overwhelming.

Complexity should pay rent in OOS robustness, not in backtest cosmetics.