RandomizedSearchCV can give worse results than manual tuning due to a few common reasons:

1. Too few iterations – n_iter=10 may not explore enough parameter combinations.

2. Poor parameter grid – Your grid might miss optimal values or be too coarse.

3. Inconsistent random seeds – Different runs can yield different results if random_state isn’t set.

4. Improper CV splits – Use StratifiedKFold for balanced class sampling.

5. Wrong scoring metric – Make sure scoring aligns with your real objective (e.g., accuracy, f1).