0

Bootstrapping is such a great technique for ensuring that noise is correctly accounted for.

Why don't we always use bootstrapping for all statistical and machine learning techniques? It seems to apply to any use case.

  • 2
    But it does *not* apply everywhere! Like all other statistical procedures, there are circumstances where this one can be expected to give useful results, others where it definitely will not work or be misleading, and everything in between. Have you noticed that *all* the theoretical justifications for bootstrapping are asymptotic? – whuber Mar 07 '16 at 20:59
  • It doesn't hurt to apply it. It may be computationally expensive, however, in all practical applications, bootstrapping sounds like it would lead to better results. –  Mar 07 '16 at 21:00
  • Definitely not! It always can hurt to apply a procedure that is unjustified, because it can give misleading and wrong results. – whuber Mar 07 '16 at 21:01
  • Bootstrapping accounts for noise in the sample. By bootstrapping, we remove the effects of noise. I don't see where this could go wrong. If there was no noise in the sample, we'd never get a poorer result. But we will always get a better result if there was noise, and in most cases there will always be noise –  Mar 07 '16 at 21:03
  • When does bootstrapping give negative performance? –  Mar 07 '16 at 21:04
  • I believe I found a duplicate question that you will find informative. – whuber Mar 07 '16 at 21:06
  • 1
    "When does bootstrapping give negative performance?"...When the sample size is small or the model being bootstrap is relatively complex when compared to the sample size. When the sample is not iid and you aren't using the proper re-sampling technique to account for it. When there exists much more efficient parametric estimators for your data that you are not considering. When your data/model is so big/complex that bootstrapping would take an extremely long period of time and only improve negligibly over simpler estimators. – Zachary Blumenfeld Mar 07 '16 at 21:11
  • There are other related threads on the site as well (besides the noted duplicate). See [*What are examples where a “naive bootstrap” fails?*](http://stats.stackexchange.com/q/9664/1036) for one example. – Andy W Mar 07 '16 at 21:17

0 Answers0