I am not a statistician by training¹ but rather and engineer appreciating statistics and enjoying building data products. No surprise something as practical as bootstrapping fascinates me. Bootstrapping is a statistical technique that helps estimating the uncertainty associated to any population statistic of interest (mean, median…).

So, what can bootstrapping do for you? It allows to determine confidence interval in a a very general way. If you need deep concetration to describe what a “confidence interval”² is, keep reading on: knowing bootstrapping also improves the grasp of this concept.

What bootstrapping can’t do is improving your point estimate of the statistic. Let’s suppose you need to estimate the median of a population and you have 101 datapoints at hand. The median point estimate is obtained by ordering your data -ascending for instance- and get the 50th item in the list. By using bootstrapping, you can supplement your point estimate with uncertainties statement. Your number won’t be $^{x}$ anymore but somwhere between $[^{x}-s_1, ^{x}+s_1]$ with a given confidence level $c$, say 95%³.

A motivating example

Mario pushed his model to production yesterday ….

The core idea

The bootstrap core idea is that the we can estimate the dispersion a given statistic by running multiple experiments on a data sample.

Good stuff

1. https://speakerdeck.com/jakevdp/statistics-for-hackers?slide=138
2. https://yanirseroussi.com/2019/01/08/hackers-beware-bootstrap-sampling-may-be-harmful/
3. https://yanirseroussi.com/2020/08/24/many-is-not-enough-counting-simulations-to-bootstrap-the-right-way/
4. https://erikbern.com/2018/10/08/the-hackers-guide-to-uncertainty-estimates.html