[ENH] Meta-Estimators

[henrydingliu]

Description

we are facing two immediate tests of package maturity in the process of recreating Friedland

• Chapter 11 FreqSev is requiring a decent amount of manual workflows
• Chapter 12 Case Outstanding requires more averaging parameters than the current implementation

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.

@priyam0k suggested meta-estimators, which makes a lot of sense. for example, FrequencySeverity could be a meta-estimator that just multiples the ultimate_ of its constituent estimators. This should cover Approach no. 1 and no. 2 from Friedland.

Some additional meta-estimators we should look into

• refactor CaseOutstanding as a meta-estimator
• Approach no. 3 likely needs a separate FullTriangleFreqSev (pending better naming) meta-estimator

Before we get too far down the rabbit hole of meta-estimators, we should also provide QoL follow-ups

• enhance model_diagnostics to accept all meta-estimators
• probably a new public class for meta-estimators (we only have a private one in the package currently). to define properties that model_diagnostics can pick up from meta-estimators

Do you have any additional supporting notes?

Some code snippet to illustrate the idea

standard_pipe = cl.Pipeline(
    steps=[
        ('tri_sel', cl.TriangleSelector()),
        ('dev', cl.Development()),
        ('tail', cl.TailConstant()),
        ('model', cl.Chainladder())
   ]
)
exh1_ccc = clone(standard_pipe)
exh1_ccc .set_params(
   tri_sel__col = 'Closed Claim Counts',
   tail__tail = 1
)
exh1_rcc = clone(standard_pipe)
exh1_rcc .set_params(
   tri_sel__col = 'Reported Claim Counts',
   tail__tail = 1
)
exh1_rsev = clone(standard_pipe)
exh1_rsev .set_params(
   tri_sel__col = 'Reported Severity',
   tail__tail = 1
)
exh1_freq = cl.VotingChainladder(
   estimators = [
      ('ccc',exh1_ccc),
      ('rcc',exh1_rcc),
  ],
   weight = ...
)
exh1 = cl.FrequencySeverity(
   estimators = [
      ('count',exh1_freq),
      ('severity',exh1_rsev),
  ]
)

exh1.fit(cl.load_sample('friedland_auto_freq_sev'))

Would you be willing to contribute this ticket?

• Yes, absolutely!
• Yes, but I would like some help.
• No.

[priyam0k]

yes, that snippet is exactly what i had in mind. the ('count', ...) / ('severity', ...) naming resolves the ordering concern, and nesting a VotingChainladder inside the count branch reads cleanly.

since this is a fair bit of surface area, maybe we split it: i take FrequencySeverity + a small public meta-estimator base class (factored out of the private _BaseTriangleEnsemble) so model_diagnostics has stable properties to read, and validate against exhibit 1. then CaseOutstanding and the approach 3 estimator can follow as separate PRs. does that sequencing work for you?

[henrydingliu]

sequencing is fine. obviously no expectation that any of this is done in the next few months or even the next couple of years, i just wanted to capture our discussion for the future. my vote is for some manner of friedland recreation to be completed first, no matter the verbosity, and attract more users to this package.

[priyam0k]

sounds good, agreed. i'll keep pushing on the friedland chapters and we can revisit the meta-estimators later.