The fitted models an interval is built from

Cross-validation already fits one model per fold and then throws them away. Asking tune to hand them back costs nothing, so the default ensemble is free: the only extra work a projection does is predicting from each member.

Usage

projection_ensemble(resampled, workflow, model_data, settings)

Arguments

resampled

the tune::fit_resamples() result, fitted with control_resamples(extract = )

workflow

the workflow to refit, for bootstrap

model_data

the data to refit on, for bootstrap

settings

from uncertainty_settings()

Value

a list of fitted workflows

Details

bootstrap refits instead, which does cost, and buys the one thing folds cannot give — enough members for a quantile to mean something. Ten folds support a spread; they do not support a 95% interval, because the 2.5th percentile of ten numbers is the smallest of them.