Currently, setting "auto" for the max_features parameter of RandomForestRegressor (and ExtraTreesRegressor for that matter) leads to choosing max_features = n_features, ie. simple bagging.

This is misleading if the documentation isn't carefully examined (in particular since this value is different for classification, which uses sqrt(n_features), actually leading to a proper random forest and not bagging). One may think they are training a Random Forest when they really aren't.

Furthermore, this is not the recommended value for regression. For example, in Hastie et. al., 15.3 (p. 587), the recommended value is n_features / 3.