I have a data set with about 3000 members that I'd like to use as a feature for a binary classification algorithm. The variable seems to be skewed by nature because the histogram is tailed at pretty much any sample. Below is the current population:

enter image description here

So I have a few questions that hope those of you who are much more experienced might be able to help with.

  • Can I "fit" these data to a more suitable distribution than a normal one? Perhaps a chi square?
  • For a non-normal distribution, especially a tailed one like mine, is there something analogous to standard deviation that I can use to characterize a data point?


  • $\begingroup$ Note: after posting, I realized I meant to add gamma as a potential fit $\endgroup$
    – R Vincent
    Dec 29, 2018 at 13:51
  • $\begingroup$ There are many relatively heavy-tailed, right-skewed distributions. Sometimes it is useful to identify the 'name' of such a distribution (in order to have a functional for for the PDF or CDF and to do inference), and sometimes not. // You might add Weibull, Rayleigh, and lognormal to your list of right-skewed distributions with support $(0, \infty).$ $\endgroup$
    – BruceET
    Dec 29, 2018 at 16:28
  • $\begingroup$ 1. Why would you need to identify a distribution if you're planing to use this as a feature? Aren't you just conditioning on it anyway? 2. It's unclear to me what you mean by "tailed at pretty much any sample"; you only appear to have one sample, and it's not clear to me what you intend by "tailed". $\endgroup$
    – Glen_b
    Dec 30, 2018 at 6:40

1 Answer 1


FYI I was able to find a fitting algorithm with the following results:

enter image description here

Turns out that the best fit is "Inverse Gaussian"!

Now I need to gain an understanding of the parameters.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.