I am searching for a theoretical or experimental estimation of the lower bound for the number of trees in a Random Forest classifier.

I usually test different combinations and select the one that (using cross-validation) provides the median best result.

However, I think that there may be a lower bound on the number of trees to use, given $m$ observations and $n$ features, to reduce the variance contribution to the error. Is there some test or reference I may check to correctly tune my classifier?

  • 3
    $\begingroup$ I doubt it. But a suggestion: the area of machine learning that studies performance in machine learning is called empirical risk minimization. You can try to add those keywords to your searches. $\endgroup$ Aug 9, 2016 at 10:10

1 Answer 1


This is not necessarily an answer to your question. Just general thoughts about cross-validating the number of decision trees within a random forest.

I see a lot of people in and cross-validating the number of trees in a random forest. I have also asked a couple of colleagues, and they tell me it is important to cross-validate them to avoid overfitting.

This never made sense to me. Since each decision tree is trained independently, adding more decision trees should just make your ensemble more and more robust.

(This is different from gradient boosting trees, which are a particular case of ada boosting, and therefore there is potential for overfitting since each decision tree is trained to weight residuals more heavily.)

I did a simple experiment:

from sklearn.datasets import load_digits
from sklearn.ensemble import RandomForestClassifier
from sklearn.grid_search import GridSearchCV
import numpy as np
import matplotlib.pyplot as plt

df = load_digits()
X = df['data']
y = df['target']

cv = GridSearchCV(
    {'n_estimators': np.linspace(10, 1000, 20, dtype=int)},
cv.fit(X, y)

scores = np.asarray([s[1] for s in cv.grid_scores_])
trees = np.asarray([s[0]['n_estimators'] for s in cv.grid_scores_])
o = np.argsort(trees)
scores = scores[o]
trees = trees[o]
plt.plot(trees, scores)


I am not saying you are committing this fallacy about thinking more trees can cause overfitting. You clearly are not since you have asked for a lower bound. This is just something that has been bugging me for a while, and I think it is important to keep it in mind.

(Addendum: Elements of Statistical Learning discusses this on page 596, and is in agreement on this with me.

"It is certainly true that increasing B [B=number of trees] does not cause the random forest sequence to overfit".

The author does make the observation that "this limit can overfit the data". In other words, since other hyperparameters may lead to overfitting, creating a robust model does not rescue you from overfitting. You have to pay attention when cross-validating your other hyperparameters.)

To answer your question, adding decision trees will always be beneficial to your ensemble. It will always make it more and more robust. But, of course, it is dubious whether the marginal 0.00000001 reduction in variance is worth the computational time.

Your question, therefore, as I understand, is whether you can somehow calculate or estimate the number of decision trees to reduce the error variance to below a certain threshold.

I very much doubt it. We do not have clear answers for many broad questions in data mining, much less specific questions like that. As Leo Breiman (the author of random forests) wrote, there are two cultures in statistical modeling, and the random forest is the type of model that he says that has few assumptions but is also very data-specific. That is why he says we cannot resort to hypothesis testing; we have to go with brute-force cross-validation.

  • $\begingroup$ Mh, I thought that increasing the number of trees would lead to an increase of the bias component of the error (which is somehow counter-balanced by a decreasing variance of the model). $\endgroup$
    – gc5
    Aug 10, 2016 at 11:06
  • 1
    $\begingroup$ I always thought that adding trees could never hurt, but in The Elements of Statistical Learning page 596, they claim that too rich of a forest can introduce extra variability, so I think your answer is still correct, but with a tiny grain of salt :) $\endgroup$
    – TBSRounder
    Aug 10, 2016 at 11:43
  • 1
    $\begingroup$ @TBSRounder Thanks for the link. The author is discussing too many things in such a short piece of text, so it's hard to distill him, especially confusing because he talks about tree depth and tree number at the same time. But, with regard to B (=#trees), he is saying that more trees makes the resulting model closer to the average tree, and, of course, the average tree may be biased, so you may end up with a biased model. I never claimed otherwise. I just said that B, by itself, does not increase bias. It just makes your model more robust (=reduces variance). $\endgroup$ Aug 10, 2016 at 13:03
  • $\begingroup$ Ah yes, It was a little confusing, thanks for the clarification. For some reason I thought the author was talking about B, not tree depth. $\endgroup$
    – TBSRounder
    Aug 10, 2016 at 13:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.