Expected Loss Method

[henrydingliu]

noticed this method was missing from the API reference.
https://chainladder-python.readthedocs.io/en/latest/modules/api.html#module-chainladder.methods

is it possible to do the following all within the cl package, without resorting to pandas manipulation?

1. get normal cl ultimate
2. derive ultimate loss ratio
3. for latest 3 origins, set final ultimate loss ratio to average of prior 4 origins

[kennethshsu]

I think the doc could use another round of revamp, there are some known issues outstanding, for example codes not compiling even though the source codes are fine.

Anyways...

Yes, 1 and 2 are relatively easy.

xyz = cl.load_sample("xyz")
cl_mod = cl.Chainladder().fit(xyz["Incurred"])
cl_mod.ultimate_ / xyz["Premium"].latest_diagonal

To do 3. We'll need to get the average prior LRs first, this can only be done with the help of pandas:

elr = (
    (cl_mod.ultimate_ / xyz["Premium"].latest_diagonal)
    .to_frame(origin_as_datetime=True)
    .iloc[4:-3]
    .mean()
)
elr.iloc[0]

Fit a new model using expected loss ratio:

el_mod = cl.ExpectedLoss(apriori=elr.iloc[0]).fit(
    xyz["Incurred"], sample_weight=xyz["Premium"].latest_diagonal
)
el_mod.ultimate_ 

Then, use cl.VotingChainladder():

model_weights = np.array([[1, 0]] * 8 + [[0, 1]] * 3)
vot_mod = cl.VotingChainladder(
    estimators=[("cl", cl_mod), ("el", el_mod)], weights=model_weights
).fit(xyz["Incurred"], sample_weight=xyz["Premium"].latest_diagonal)
vot_mod.ultimate_

[henrydingliu]

it's a shame that 3 has to be done outside of cl, when you consider that most of cl is just an organized way of taking averages of ratios.

would it make sense to expand the EL/BF/BT methods? something like, add an n_period parameter to the class. by default n_period = 0, we go to the current logic with the single apriori. n_period = 1,2, 3, ,,,n, -1, the average ratio against the sample weight is taken and applied to the sample weight.

[kennethshsu]

In this case, I don't think it makes sense to implement additional hyper parameters for the CL/BF models. If you think about the actual actuarial models, those parameters are not required or used. The CL model, is, just the CL model - it does not depend on how many "n"s to evaluate. This is why VotingChainLadder(...) was developed, so that we can mix and match models with varying weights.

If you look at my code above a bit more carefully, the only place that you would go outside of the chainladder package is to get the a priori loss ratio, if you have this else where (say judgmentally selected), you wouldn't have to go to pandas.

I'm curious to hear why you think exiting chainladder and using pandas is not convenient?

[henrydingliu]

If you think about the actual actuarial models, those parameters are not required or used.

same could be said about the age-to-age averages. you could always judgmentally select them.

If you look at my code above a bit more carefully, the only place that you would go outside of the chainladder package is to get the a priori loss ratio,

still a piece of code outside of the cl pipelines

I'm curious to hear why you think exiting chainladder and using pandas is not convenient?

not meta enough. sklearn doesn't make ppl exit and do preprocessing in np.

[kennethshsu]

Wait maybe I didn't understand you correctly.

Can you explain again what you are trying to accomplish?

[henrydingliu]

genins = cl.load_sample("genins")
genins_dev = cl.Development().fit_transform(genins)
genins_el = cl.ExpectedLoss(n_period = 4).fit(genins_dev)

[henrydingliu]

actually, i've been thinking about this all wrong. rather than mudding up bk, it would make more sense to generalize the generalized cape cod method to allow arbitrary specification of origin weights, with defaults to generalized cape cod. looking through the capecod source code, is it as simple as exposing weighted_exposure?

having said that, in the spirit of building flexible pipelines, maybe capecod shouldn't be an ibnr estimator at all, but an apriori estimator that combos with BF. an apriori estimator can produce a single apriori ratio estimate, or can predict ultimate for future periods if exposures for future periods is passed in.

[henrydingliu]

and if we are already at that point, just need to break out the trend parameter into a trend and a rate change vector, and chainladder can do indications

[jbogaardt]

This makes sense to me, we just need to think through the technical design.