How to link fit() and fit_transform() together?

[Xtina0]

Hi there,

I noticed that only fit_transform() has link_ratio as an attribute whereas fit() doesnt.

• Would you be able to pin point where in the Chainladder package fit_tranform().link_ratio is produced?
• I also have trouble understanding how fit_transform() links back with fit() (if at all) i.e. what functionalities do they share? Could you please elaborate on this?

Many thanks,
Xtina

[jbogaardt]

You might be comingling the concepts of Estimators and Triangles. Triangles have a link_ratio property, estimators do not. Estimators are all other classes that estimate something from a triangle such as Development, BornhuetterFerguson, TailCurve, etc.

fit, transform and fit_transform function in the same way they work in sklearn.

The fit method returns the model you are working on with fitted attributes. transform always returns a Triangle object augmented with any additional info passed on by the model you are working with.

So if you're using a Development model, and only call fit. The resulting object will be the Development estimator itself and will have all the Development fitted attributes, such as ldf_, cdf_, etc. If you transform a Triangle, then those attributes will be passed on to the triangle, along with the attributes the triangle already had, such as link_ratio`.

fit_transform is a shorthand for the two steps of fitting and transforming. That is

 Development().fit(X).transform(X)
# can be expressed as
Development().fit_transform(X)

Does that clarify?

[Xtina0]

Yes and thanks for your clarification!
If I may have two follow-up questions -

1. I am wondering where in the code controls the rounding (i.e. s.f.). For example, in Development model? Specifically I would like to make use of ldf_ but with a slightly different rounding.
2. While I have your attention here, Would you be able to elaborate on how each folder links to each other below? Appreciate this might be a rather vague question, I just want to have a better understanding of the code set up so I can make changes.
   

[jbogaardt]

1. There is no internal rounding on any calculations in chainladder. The display may show formatted rounding and that is here. If you want to change the precision of actual data underlying Triangle, just call its round method, e.g. cl.load_sample('raa').link_ratio.round(1) which works the same way as pandas.DataFrame.round.
2. The folders correspond 1-to-1 with the User Guide. The only renaming going on is that the core folder contains the Triangle object.