My problem is to estimate the distribution some observed data. I am exploring what best distribution to use. The data is fat tailed most likely following a power law in the tails (the burr distribution has the maximal likelihood among a set of usual distribution)

I would like to produce an invertible (on $\mathbb{R}^+$) CDF function from observed data. It's important for the CDF to be invertible, that the inverse be fast to calculate (can be achieved computationally) and that I can use on all the domain $\mathbb{R}^+$. Also, the storing parameters or serialization is an important consideration.

Since I need it inversible and full domain usage => Empirical is not possible.

This leaves me with testing several parametric distributions or use a kernel density estimation.

For parametric estimation:

  • Which method should I use, given the fat tails? in R's fitdistrplus package they have maximum likelihood (mle), moment matching (mme), quantile matching (qme) or maximizing goodness-of-fit estimation (mge).

For nonparametric density kernel estimation - again given the fat tails:

  1. Does it estimate the power law tail correctly?
  2. Will it work on the whole domain even far outside what i have in the observations, or would i have to extend it ?
  3. Is it easy to store?

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.