Default QRF grows trees to purity, giving miscalibrated quantiles and inert sample weights

[vahid-ahmadi]

The default QRF configuration grows trees to purity, so leaves hold a single training point and the "forest" returns essentially the same one or two neighbours for every query. The resulting quantiles are badly miscalibrated.

Coverage on Y | X ~ N(2X, 1), n = 5000:

nominal        0.10   0.25   0.50   0.75   0.90
OLS            0.103  0.256  0.498  0.748  0.905
QuantReg       0.104  0.254  0.505  0.751  0.898
QRF msl=1      0.296  0.376  0.502  0.616  0.697   <- default
QRF msl=50     0.123  0.270  0.500  0.738  0.881

At the default, the nominal 10th percentile covers 30% of the data and the 90th covers 70%. The distribution is far too narrow in the tails, which is exactly where imputed wealth or asset variables matter.

A second consequence: sample weights become inert

With one training point per leaf there is nothing for weights to reweight:

min_samples_leaf=1    unweighted +5.35   weighted +5.35   (identical)
min_samples_leaf=20   unweighted +4.32   weighted -4.58   (follows weights)

The weight plumbing is correct; the default defeats it.

Also: DEFAULT_MODEL_PARAMS is dead code

config.py documents min_samples_leaf: 1, max_depth: None, but DEFAULT_MODEL_PARAMS is never referenced anywhere in the package — the library simply inherits the underlying library's fully-grown defaults. So the documented configuration is not the one in force.

Fix

Wire DEFAULT_MODEL_PARAMS in, and default min_samples_leaf to something in the range 10–20. QRF is the method the benchmarking paper recommends, so its default configuration is worth getting right.