add summary_ to methodbase

[henrydingliu]

Description

outputting chainladder results can be a pain today, evidenced by our campaign to recreate friedland.

Is your feature request aligned with the scope of the package?

• Yes, absolutely!
• No, but it's still worth discussing.
• N/A (this request is not a codebase enhancement).

Describe the solution you'd like, or your current workaround.

adding a property to MethodBase that displays the following in a pandas dataframframe

index
column
origin
development
ldf_
cdf_
latest_diagonal_
ultimate_
ibnr_

not every use case will need all these series each time. but it's easy enough to call .summary_[[columns_you_want]]

Do you have any additional supporting notes?

No response

[genedan]

What do you think of adding it to the __repr__ of the estimator? Printing it doesn't have any meaningful information:

est = cl.Chainladder().fit(raa)
print(est)

Chainladder()

I'm not all that familiar sci-kit learn syntax though, or if they have summaries of any kind. But it doesn't look like sklearn's estimators print meaningful information either:

from sklearn.datasets import make_regression

X, y, true_weights = make_regression(
    n_samples=100,
    n_features=100,
    n_informative=10,
    noise=8,
    coef=True,
    random_state=42,
)


from sklearn.linear_model import LinearRegression

olr = LinearRegression().fit(X, y)
print(olr)

LinearRegression()

Statsmodels does have a summary though, and it's just via the summary method:

https://www.statsmodels.org/stable/glm.html

So I guess maybe for now I'm leaning towards your OP approach of summary_ .

[henrydingliu]

i think John intended for __repr__ to provide metadata about the estimator(s), rather than the result of the estimation. here is a more illustrative example.

on second thought, would it make sense for our summary() to return a triangle by default? then to_frame() can generate the df

[kennethshsu]

I'd like all the estimator models to have a .summary() that shows the latest_diagonal, ultimate_, IBNR_. I think this would be nice.

But the truth is there's no way to standardize an output easily no? Everyone needs different things.

[genedan]

Since the estimator is a separate class, I think it's OK to use summary() or summary_ (?) to produce the exhibit for a fitted estimator. You could extract the triangle with X_ and that would have its own triangle summary.

As far as standardization goes, I think transposing and concatenating the fields in the OP is okay, since we're just developing an input column (whatever it may be) to ultimate. So we can certainly include the column, the ultimate, origin, development and the LDFs and CDFs in there.

I think what would be out of scope would be further derived columns like IBNR or unpaid reserve, net/gross of reinsurance and/or sal/sub/expenses, which some companies may have their own custom formulas for.

[genedan]

Regarding those derived results, we could at ease things by making sure the summary frames are consistently formatted such that summaries with the same origin can be merged easily. For example if someone had separate models for incurred, reported, expenses, etc. and then wanted to combine them all, they can concat and then use standard pandas formulas to get stuff like IBNR.

[kennethshsu]

Regarding those derived results, we could at ease things by making sure the summary frames are consistently formatted such that summaries with the same origin can be merged easily. For example if someone had separate models for incurred, reported, expenses, etc. and then wanted to combine them all, they can concat and then use standard pandas formulas to get stuff like IBNR.

Yes! And if columns consist of the same info not to be showing up twice (i.e. latest_diagonal)!