oob_score :bool, default=False

[sunsunsun5114]

What happens if you set oob_score :bool, default=False to True in causal forest, and how many folds will the data become?

[0trm]

There are no folds. oob_score is scikit-learn's bootstrap mechanism: each tree is fit on n rows drawn with replacement, so every row is left out of roughly 37% of the trees ((1 - 1/n)^n), and its out-of-bag prediction averages only those trees. On 0.18 each tree also splits its sample again, since honesty=True is the default: half grows the structure and the other half estimates the leaves.

That said, from reading the 0.18.0 code I'd expect oob_score=True to raise once the trees are trained. The forest receives y from _prepare_data as an n × groups matrix that is NaN outside each row's own treatment group, and the line right before the OOB step calls sklearn's type_of_target(y), which rejects NaN in float targets. Even if it got through, sklearn's OOB score is the R² of the out-of-bag predictions against y, and here the predictions are treatment effects while y holds outcomes, so it wouldn't tell you how good the CATEs are.

For a generalization check I'd hold out a split and score the forest's predictions on it. In 0.18.0, causalml.metrics.dr_score does that against doubly robust pseudo-outcomes (lower is better), and qini_score or auuc_score work if what you care about is ranking users by uplift.

I haven't run oob_score=True myself, so treat the error as my reading of the code.