VotingChainladder enhancements

Would like to get additional model properties functioning with the VotingChainladder estimator prior to release.

import chainladder as cl
raa = cl.load_sample('RAA')

bcl1 = cl.Pipeline(steps=[('dev', cl.Development(average='volume')), ('model', cl.Chainladder())])
bcl2 = cl.Pipeline(steps=[('dev', cl.Development(average='simple')), ('model', cl.Chainladder())])

estimators = [('bcl1', bcl1), ('bcl2', bcl2)]
vot = cl.VotingChainladder(estimators=estimators)
vot.fit(raa)

assert abs((bcl1.fit(raa).named_steps['model'].full_expectation_ + bcl2.fit(raa).named_steps['model'].full_expectation_)/2 - vot.full_expectation_).max().max()< 1e-4

[jbogaardt]

2. if origins provided (e.g. through dict) check that all origin periods are allocated and apply the default for unallocated (in future perhaps provide choice of functionality for omitted origins, e.g. applying first estimator). Provide an exception on origins that don’t exist.

This may help reduce the complexity of the dictionary approach:

import chainladder as cl
raa = cl.load_sample('raa')

dict_weight = {'1989':(1, 0), '1988': (0.75, 0.25)}

# Convert dict_weight into a mapping dictionary
mapping_dict = {raa.origin[raa.origin==k][0] :v for k, v in dict_weight.items()}
# Fill in missing values
missing = {k: (0.5, 0.5) for k in raa.origin[~raa.origin.isin(mapping_dict.keys())]}

# Use it in same way as the callable method
raa.origin.map({**mapping_dict, **missing})

[cbalona]

There also appears to be some unaccounted for mutation/side-effect we should resolve. Notice the second “refit” of the estimator.

This is due to adding a new axis so that the weights can multiply with the ultimates. When refitting, it is adding a 6th axis. Should be a quick easy fix. It may also be avoided through better handling of the weights as per my previous comment.