NEWS
quantregForest 1.4-0 (2026-07-03)
- New maintainer: Nicolai Meinshausen <[email protected]>.
- Substantial speed improvements, all without changing results:
- predict() now computes quantiles with a single vectorized
sort per prediction matrix instead of calling quantile()
once per observation (predictions on new data are now
considerably faster; results are unchanged, type-7 quantiles
as before).
- Fitting with keep.inbag=TRUE previously used an O(n^2) loop
per tree to sample out-of-bag responses; this is now fully
vectorized. The overhead of keep.inbag=TRUE over a plain fit
is reduced by more than an order of magnitude for larger n.
- quantregForest() no longer computes the terminal-node matrix
of the training data twice when keep.inbag=TRUE.
- Out-of-bag predictions via predict() without newdata are much
faster.
- Removed a broken, unreachable duplicate definition of
quantregForest() in R/convert.R that masked the real function
during package collation.
- quantregForest is now a pure R package. The bundled copy of the
C/Fortran sources and R helper functions of the randomForest
package (used only for fitting with nthreads > 1) has been
removed; parallel fitting now calls randomForest() directly in
each fork and merges the forests with randomForest::combine().
This removes the compiler dependency, keeps the parallel code
path automatically in sync with upstream randomForest, and does
not change the fitted forests.
- RColorBrewer is no longer attached (it was needed only by the
removed copies of randomForest plotting helpers).