fastcpt(metric = ...) now defaults to NULL, which auto-selects the test statistic: out-of-bag backends ("forest", "ferns") at leaveout = 0 use "rate" (the out-of-bag classification accuracy rate); all other cases use "probability". This makes the headline statistic for the default OOB forest and ferns the OOB accuracy rate. An explicit metric = is always respected. balance() inherits this (e.g. class.method = "forest" reports "rate"). Note: this deviates from cpt::cpt(), whose default is "probability", but matches cpt’s documented OOB-rate behaviour when metric = "rate".
fastcpt() forest backend (class.methods = "forest") now defaults to 100 extremely-randomized trees (splitrule = "extratrees"), ~5x faster than the previous 500-tree gini default with equivalent size and power (validated over 6 core + 3 realistic DGPs at perm.N = 1000). It auto-falls back to gini when the data contain NA (extratrees cannot handle missing values).
Any ranger::ranger argument can now be forwarded via classifier.args (e.g. splitrule, num.random.splits, min.node.size, sample.fraction).
write.forest is skipped in the out-of-bag path (leaveout = 0) for speed; the leave-out predict path now restores trained column names before prediction (fixes a pre-existing failure with unnamed covariate matrices).
MLbalance 0.2
Initial CRAN submission.
balance(): unified function for covariate balance assessment and treatment effect estimation supporting binary and multi-arm treatments.
fastcpt(): fast classification permutation test with support for random ferns, ranger forests, and fast logistic regression.
random_check(): balance permutation test using boosted regression forests with diagnostic plots.
vip(): variable importance convenience function for grf model objects.
