Bootstrapping is (in general) a method of assessing the accuracy of a statistic. It can also be used to estimate the skill of a machine learning model.
Briefly, this is the process:
- Choose a number $B$ of bootstrap samples to perform.
- Generate $B$ bootstrap sets of the desired size, using sampling with replacement.
- For each bootstrap set, either:
- Calculate the statistic on the set.
- Fit a model to the set and estimate the skill of said model.
- Calculate the mean of the statistics/model skill estimates.