lmoments as a new fit method?

[kikocorreoso]

Hi,

When I have done some EVA using some datasets, not specifically using pyextremes, I get that the MLE fit could be poor. In these cases I have used lmoments3 and parameters make more sense.

Would it be interesting to add a third fit method using lmoments3 to complement MLE and EMCEE?

https://github.com/Ouranosinc/lmoments3 (see potential license issues).

[georgebv]

Thank you for the suggestion, this could be a great addition. I don't know the theory to implement this myself, but I'd gladly review/accept a PR.

[kikocorreoso]

This is how I implemented it in scikit-extremes: https://github.com/kikocorreoso/scikit-extremes/blob/master/skextremes/models/classic.py#L440

I will have a look to the internals of pyextremes and check the best way to add it. I will come back when ready. This would add a new dependency...

[georgebv]

Thank you. It looks like only a subset of scipy distributions is supported, but we have the big 4 (GEV/Gumbel and GPD/Expon) so this should be enough for 99% of scenarios.

https://github.com/Ouranosinc/lmoments3/blob/main/docs/source/distributions.rst

[acrosby]

Scipy itself supports specifying "MM" for the Method of Moments approach to fitting a .rv_continuous distribution rather than the default "MLE." Is "Lmoments" similar? Is there an easy way to specify method="MM" in the .fit() call in pyextremes currently? Seems like I would need to monkeypatch some things.