You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[DOC] Example: split the pooled versus standalone IBNR gap by company in the Development groupby example #1657
The Development docstring compares groupby="LOB" patterns with standalone patterns on clrd["CumPaidLoss"]. Its othliab row prints standalone IBNR of -14,285,874 against 1,640,597 pooled, and the text says "a handful of small companies with erratic development" cause the negative total. A per-company bridge shows that one company and one origin account for it.
For each company, the bridge multiplies each origin's latest paid by the pooled CDF minus the company's own CDF and sums over origins. The contributions add to the line's IBNR difference. On main at a59efc9, othliab gives these values.
companies 239
industry IBNR 1640597, standalone IBNR -14285874, difference 15926471
sum of per-company contributions 15926471
company own 12-24 latest paid contribution share
Dorinco Rein Co 299.8261 53301 16269828 1.022
State Farm Mut Grp 7.6392 1409719 -487530 -0.031
Grinnell Mut Grp 1.7848 116587 84887 0.005
Dorinco's 12 month paid for 1989 to 1996 sums to 46, and its 24 month paid for the same origins sums to 13,792. That gives an own 12-24 factor of 299.8261 and a 12 to ultimate CDF of about 1,602.5. Its 1997 origin shows paid of -10,225 at 12 months, and the CDF turns that into standalone IBNR of -16,364,285.
The medmal line shows why the bridge helps even without an outlier. Its total gap is -125,553, but Physicians Recip Insurers contributes -510,800 and Scpie Indemnity Co contributes +274,949.
I propose to extend the docstring example with the per-company gap for othliab, (ibnr_industry - ibnr_standalone).sum("origin") on the othliab rows, and to reword the sentence so it names the single company and origin that drive the negative total. The example needs no new API. A two-segment doctest could pin the arithmetic by hand, with 12 month denominators of 80 and 20, factors of 1.40 and 1.80, a pooled factor of 1.48, and segment deltas of -1.76 and +10.56 that sum to 8.8. That fixture reproduces exactly with Development(groupby=...) and TailConstant(1.10).
Would you take this in the docstring, or would you rather see it in a gallery example?
(This issue was written with Árbol, an AI assistant, under my direction.)
Activity
henrydingliu commentedon Nov 19, 2026
either docstring or gallery is fine