BarnettZehnwirth does not support categorical valuation predictors

[danielfong-act]

Are you on the latest chainladder version?

• Yes, this bug occurs on the latest version.

Describe the bug in words

Discussed in #579

The BarnettZehnwirth estimator fails to convert some values of valuation to categories if given a formula that contains C(valuation) or some similar term. In particular, a PatsyError is raised: Error converting data to categorical: observation with value X does not match any of the expected levels, where X is the next valuation date after the latest diagonal converted to an ordinal.

How can the bug be reproduced?

import chainladder as cl
abc = cl.load_sample('abc')
cl.BarnettZehnwirth(formula='C(valuation)').fit(abc)

What is the expected behavior?

BarnettZehnwirth should estimate coefficients for valuations up to the latest diagonal. Future valuation categories could be dropped or have their coefficients set to zero.
Fitting data with BarnettZehnwirth also populates the resultant triangle with model expectations (and predictions) for each cell in Triangle.model_.triangle_ml_. The estimator should use valuation category coefficients for this estimation when it can.

I'm not sure how to do the above in a way that works with Patsy and fitted triangle calculations.

[henrydingliu]

@danielfong-act could you please elaborate on what you mean by "estimate coefficients for valuations up to the latest diagonal. Future valuation categories could be dropped or have their coefficients set to zero"? In other words, how you would use this hypothetical BZ estimator when future valuations are all dropped or zero?

[danielfong-act]

@henrydingliu sure, I'm basing most of my conclusions off of this publication by Barnett and Zehnwirth https://www.casact.org/sites/default/files/database/forum_08fforum_3barnett_zehnwirth.pdf.
In short, categorical valuation variables hold no predictive power because we assume no relation among them. They can hopefully capture some nonlinear trends in the valuation direction (e.g. non-constant inflation, significant changes in company practices) to help better fit variables in the origin and/or development directions. A model like this would use valuation variables in conjunction with origin and/or development variables to estimate cell values for observed data, and would rely only on the origin and/or development variables for the unobserved cells.

[henrydingliu]

A model like this would use valuation variables in conjunction with origin and/or development variables to estimate cell values for observed data, and would rely only on the origin and/or development variables for the unobserved cells.

I'm not seeing how this provides a motivation for a valuation-only model.

We don't have to limit this to a theoretical discussion. As an alternative, could you please provide a numerical example (possibly in Excel) of some made-up data and how a valuation-only model could be applied to that data?

[danielfong-act]

@henrydingliu

I'm not seeing how this provides a motivation for a valuation-only model.

It doesn't. The PTF family is presented very generally and includes such models, but I do not think the authors intended people to use the family that way. In much the same way, you can construct an origin only model by excluding development and valuation variables, but this would probably not be useful. The BZ estimator supports this behavior already, however, as a side effect of being able to fit formulas that include origin variables. Fixing this bug would likewise grant the BZ estimator the ability to fit a model that uses only categorical valuation variables. Still, users will likely not want to do this.

[henrydingliu]

i think i may have misunderstood. you are not looking just for formula='C(valuation)' to work, but for C(valuation) to work as part of any formula?

[henrydingliu]

I created a new branch in which you can now engineer new features into the BZ estimator. An example is below

abc = cl.load_sample('abc')

def valuation_is_0(X):
    return [1 if x == 0.0 else 0 for x in X["valuation"]]

def valuation_is_1(X):
    return [1 if x == 1.0 else 0 for x in X["valuation"]]

feat_dict = {
    'valuation_1':valuation_is_0,
    'valuation_2':valuation_is_1}

cl.BarnettZehnwirth(formula='C(valuation_1) + C(valuation_2)',feat_eng = feat_dict).fit(abc).coef_.T.values

Is this what you are looking for? If so, we can put in some more time to create a couple of utility functions for creating feat_dict fully and dynamically

[danielfong-act]

@henrydingliu yes, the C(valuation) term is intended to work as a part of more complex formulas. Your branch looks like it took care of the heavy lifting, so good work there. However, I'm starting to suspect that the real issue is in fitting LDFs.

[henrydingliu]

I added parameters to the feature engineering dictionary to allow more generalized functions. the update is in the same branch. the snippet below recreates what you were intending to achieve with C(valuation). do the estimated LDFs follow your expectation?

abc = cl.load_sample('abc')

def valuation_onehot(df,val):
    return [1 if x == val else 0 for x in df["valuation"]]

feat_dict = {f'valuation_{x}':{'func':valuation_onehot,'kwargs':{'val':float(x)}} for x in range(11)}

print(cl.BarnettZehnwirth(formula='development').fit_transform(abc).ldf_.values)
print(cl.BarnettZehnwirth(formula='development+' + '+'.join([f'C({x})' for x in feat_dict.keys()]),feat_eng = feat_dict).fit_transform(abc).ldf_.values)