A multistep model's per-step outcome columns (e.g. lead_2_value) are
engineered by step_lead() from a single raw column, so they are not
present in a user's raw data the way genuine multi-output columns are.
This re-bakes the fitted recipe on new_data to recover the actual
future value at kerasnip_step_view()'s step, for calibration/interval
use. Rows too close to the end of new_data for the lead to be computed
return NA (dropped automatically by calibration routines that call
sort()/stats::complete.cases() on the result).
Arguments
- view
A
kerasnip_step_view.- new_data
A data frame of raw predictors (and the original outcome column
step_lead()was applied to).