Why does RandomizedSearchCV sometimes return worse results than manual tuning in scikit-learn?

I'm working on a classification problem using scikit-learn's RandomForestClassifier. I tried using RandomizedSearchCV for hyperparameter tuning, but the results were worse than when I manually set the parameters based on intuition and trial/error.

Here's a simplified version of my code:

from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import RandomizedSearchCV

param_dist = {
    "n_estimators": [100, 200, 300],
    "max_depth": [None, 10, 20, 30],
    "min_samples_split": [2, 5, 10],
    "min_samples_leaf": [1, 2, 4],
}

clf = RandomForestClassifier(random_state=42)
random_search = RandomizedSearchCV(clf, param_distributions=param_dist, n_iter=10, cv=5, scoring='accuracy')
random_search.fit(X_train, y_train)

In multiple runs, this approach yields models with lower accuracy on my test set than my manually-tuned model.

What are common pitfalls when using RandomizedSearchCV?

How can I ensure reproducibility and robustness of the tuning process?

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).