Verdict.knn scores highest, but
linear, hist-gbm, rf, dummy are statistically indistinguishable
from it at this sample size. Choosing between them on these numbers alone
is not supported.
Checks
warn
4 model(s) are not statistically distinguishable from knn: linear, hist-gbm, rf, dummy
How this was measured
cross-validation
5-fold, stratified, shuffled with seed 0
preparation
numeric: impute median · categorical: impute most_frequent + onehot · standardise (scale-sensitive models only) — fitted inside each fold
baseline
a dummy model is always included as the floor
significance
corrected resampled t-test (Nadeau & Bengio, 2003), Holm-adjusted across models
intervals
95% t-interval across folds — optimistic, since folds share training data
Intervals describe the spread of fold scores, not true generalisation error.
The significance test, which corrects for the overlap between training folds,
is what the verdict relies on.