ENH: Bootstrapping Confidence Intervals + Statistical Tests

[quillan86]

Is your feature request related to a problem? Please describe

Not a problem per se - new feature due to lack of ability to find this in commonly used python statistical packages

Describe the solution you'd like

I'd really like a way to do bootstrapping confidence intervals as well as statistical tests associated with them (e.g. one-sample and two-sample tests) within the package. I could contribute myself although I'm new in terms of open source stuff esp on a package on this level. Especially since R has standard packaging in this. In particularly relevantly for me I'd want to allow for weighted estimates in bootstrapping arbitrary point estimates

Describe alternatives you have considered

I found a BSD-licensed package on bootstrapping here - but this package is inactive:

https://github.com/facebookarchive/bootstrapped

In addition I hand-wrote some code to handle bootstrapping myself (mostly for practice - so obviously it's rough, not optimized and POC-like). In my particular use case I'd like to specifically account for weighted point estimates so any method should account for weighting in there optionally.

The R package bootstrap is here, but of course this is in R:
https://cran.r-project.org/web/packages/bootstrap/bootstrap.pdf

[WarrenWeckesser]

FYI: There is a SciPy pull request that is likely to be merged soon that adds a function to compute bootstrapped standard errors and confidence intervals for a given function that computes a statistic.

[bashtage]

The arch package also has a complete framework for bootstrapping arbitrary statistics:

https://bashtage.github.io/arch/bootstrap/bootstrap_examples.html

[quillan86]

Okay, thanks. I guess it's confusing trying to figure out which package has what. I haven't seen the arch package before.

[josef-pkt]

What's the role of weights in bootstrap? Which kind of weights?

I don't remember having seen weights there before and it might need additional support to what's available.

(statsmodels only has case specific support mainly when asymptotic results are not available)