Introduction

In scikit-learn, a random forest normally trains each tree on a bootstrap sample of the training set, which means sampling rows with replacement. When you set bootstrap=False, every tree sees the entire training set instead. The model is still a forest, but it loses one of the main sources of tree diversity that classic random forests rely on.

What Changes When bootstrap=False

With the default bootstrap=True, each tree is trained on a slightly different resampled dataset. That row-level randomness helps decorrelate the trees, which is a major reason the ensemble generalizes well.

When bootstrap=False, the trees still have some randomness because scikit-learn can still randomize feature selection at splits. But every tree is now looking at the same rows.

That means:

• less diversity from training data sampling
• potentially more correlation between trees
• no out-of-bag scoring

The last point matters because scikit-learn’s out-of-bag estimate depends on the fact that each bootstrap sample leaves some rows out. If bootstrap=False, there are no left-out rows for that mechanism.

A Minimal Example

Here is a simple classifier comparison:

1from sklearn.datasets import load_iris
2from sklearn.ensemble import RandomForestClassifier
3
4X, y = load_iris(return_X_y=True)
5
6rf_bootstrap = RandomForestClassifier(
7    n_estimators=200,
8    bootstrap=True,
9    random_state=42
10)
11
12rf_no_bootstrap = RandomForestClassifier(
13    n_estimators=200,
14    bootstrap=False,
15    random_state=42
16)
17
18rf_bootstrap.fit(X, y)
19rf_no_bootstrap.fit(X, y)
20
21print(rf_bootstrap.predict(X[:5]))
22print(rf_no_bootstrap.predict(X[:5]))

Both models train successfully. The difference is not whether the forest works. The difference is how the trees are diversified during training.

Why You Might Use bootstrap=False

There are a few legitimate reasons:

• you want to compare bagged and non-bagged behavior experimentally
• your dataset is small enough that throwing rows away in bootstrap samples feels undesirable
• you prefer to rely mostly on feature randomness rather than row resampling

In some cases, using the full dataset for every tree can improve training stability or slightly improve fit on a specific problem. But it is not automatically better. The usual strength of random forests comes from combining many decorrelated trees, and bootstrap sampling is one important way to create that decorrelation.

So turning it off is a deliberate modeling choice, not a default optimization.

What You Lose Operationally

The biggest practical loss is out-of-bag evaluation:

1from sklearn.ensemble import RandomForestClassifier
2
3model = RandomForestClassifier(
4    n_estimators=100,
5    bootstrap=True,
6    oob_score=True,
7    random_state=42
8)

This kind of estimate only makes sense when bootstrap sampling is enabled. If you set bootstrap=False, you should expect to use regular validation or cross-validation instead of OOB metrics.

You may also see the trees become more similar, especially if the dataset is not very noisy and max_features is generous. More similar trees often reduce the variance-reduction advantage that ensembles are supposed to provide.

It Becomes Closer to "Pasting" Than Bagging

Conceptually, bootstrap=False moves the model away from classic bagging and closer to a forest of randomized trees trained on the same full dataset. Feature randomness still matters, but the row-sampling randomness disappears.

That means if you want the traditional random-forest behavior described in many textbooks, bootstrap=True is the closer match.

If you turn it off, you are not breaking the estimator. You are changing what kind of ensemble it behaves like.

Common Pitfalls

The biggest mistake is assuming bootstrap=False merely makes training more data-efficient without tradeoffs. It also removes an important source of ensemble diversity.

Another issue is enabling oob_score=True mentally while forgetting that out-of-bag reasoning depends on bootstrap sampling. Without bootstrap, that diagnostic is no longer the right tool.

Developers also sometimes interpret better training-set performance as proof that bootstrap=False is superior. The real question is generalization on unseen data, not how tightly each tree can fit the training rows.

Finally, do not forget that randomness in feature selection still exists. Setting bootstrap=False does not turn the model into one deterministic giant tree; it changes only one axis of the forest’s randomness.

Summary

• 'bootstrap=False means every tree trains on the full dataset instead of a bootstrap sample.'
• The forest still has feature-level randomness, but it loses row-sampling diversity.
• This can make trees more correlated and removes the basis for out-of-bag evaluation.
• It is a valid experimental choice, but not the classic random-forest default.
• Judge the setting by validation performance, not by intuition alone.