Generalized Bondy - Excel version

[hoverat]

I am trying to confirm/document the generalized bondy tail. It seems like there might be a different level of precision between the calculations in Python and Excel. I would like to show the same CDFs between the two. Any suggestions are appreciated. Thank you!
 
Python/CL code
triangle = cl.load_sample('tail_sample')['paid']
dev = cl.Development(average='simple').fit_transform(triangle)
tail = cl.TailBondy(earliest_age=12).fit(dev)
tail.cdf_

Excel version
TFWP_Appendix_August2013_CL.xlsx

[jbogaardt]

I haven't had a chance to look too deeply yet at this and might not for a bit, but I suspect this has to do with the precision of each of the different solvers between chainladder-python (uses scipy under the hood), the TFWP paper (some older version of Excel), and your sample. Since solvers only provide numerical approximations, I would expect different implementations to have different values within rounding of each other. I think the only way to validate this hypothesis is to test each of the intermediate values of the overall bondy algorithm between the two.

[hoverat]

Thanks for the suggestions. I am trying to work backwards and noticed that I am getting different ldfs for the "earliest age"
triangle = cl.load_sample('tail_sample')['paid']
dev = cl.Development(average='simple').fit_transform(triangle)
dev.ldf_
84-96 = 1.0466
tail = cl.TailBondy(earliest_age=84).fit(dev)
tail.ldf_
84-96 = 1.0466
tail.earliest_ldf_
earliest_ldf = 1.046368

It seems like these 2 numbers should be the same and I haven't been able to figure out why they are not yet.