How to get PPF for minima?

[nucflash]

I'm plotting my data, and I want to add a horizontal line, say, at 75% of the maxima, and at 75% of the minima, using the respective fitted models for maxima and minima.

I do

def eva(data, extreme_types):
    model = EVA(data=data)

    model.get_extremes(
        method="BM",
        extremes_type=extremes_type,
        block_size="365.2425D"
        errors="raise")
        
    model.fit_model()
    return model
    
model_high = eva(data, 'high')
model_low = eva(data, 'low')

high_75pct = model_high.distribution.distribution.ppf(0.75)
low_75pct = model_high.distribution.distribution.ppf(0.75)

assert high_75pct != low_75pct   # should not fail, but fails

high_75pct, and low_75pct come out exactly the same. I see that you do something with ranks in get_return_periods (L121-127), but I don't know how to tap into it. Any guidance is appreciated.

[georgebv]

This example is not reproducible because data is unknown.

You are also using distribution directly. Distribution is fit to transformed extremes - maxima left intact, minima flipped to behave like maxima. When EVA calculates return values it performs inverse transformation:

pyextremes/src/pyextremes/eva.py

Lines 1253 to 1259 in 3706ff8

	# Calculate return values
	return tuple(
	self.extremes_transformer.transform(value)
	for value in self.model.get_return_value(
	exceedance_probability=exceedance_probability, alpha=alpha, **kwargs
	)
	)

[nucflash]

Thanks George!

Happy to use some sample data to make this reproducible. Do you have a dataset to suggest?

Do I interpret you correctly if I say that for maxima, model.cdf() should match the output from model.distribution.distribution.cdfd()?

[georgebv]

To add some context - transformation is necessary because GEV has specific shape which represents extreme maxima. For minima we need this inverted. There are two ways to achieve that - either implement different distributions for maximima/minima or simply transform minima to maxima (flip around mode) and use the same distribution (we need to transform results back when interpreting results because of that).

[georgebv]

You can use numpy to generate a preudo-random data (don't forget to set a seed value to make it reproducible randomness). You can use scipy to sample from a specific distribution.

[georgebv]

eva.model.distribution points to instance of class Distribution specific to pyextremes. eva.model.distribution.distribution points to scipy.rv_continuous distribution which has default values for its fit parameters (shapes, loc, scale). These are different classes from different libraries with different interfaces.

eva.model.cdf(x) and eva.model.distribution.get_prop('cdf', x, eva.model.fit_parameters) would give the same result.

Because of that, in your example you are calling the same default scipy distribution with default parameters so results are always going to be the same. data is irrelevant here.

[nucflash]

I see. That makes a lot of sense. So I always need to use model.cdf() or model.distribution.get_prop(). Thanks for the clarification! And thank you again for taking the time and patience to answer my questions!