Introduced on day 56
Block bootstrap
A resampling method that redraws whole blocks, such as weeks, instead of single rows so that dependence inside a block is kept when you compute an interval.
In detail
For time-ordered evaluation data, rows within a week are not independent. Resample the weekly blocks with replacement, recompute the paired metric difference each time, and read the interval from the percentiles. It captures which blocks happened to be in the evaluation sample for the trained model; it does not capture retraining noise, which seeds measure.