Experience Analysis

[alecloudenback]

@MatthewCaseres in case you were interested, I recreated the tutorial from actexp in this notebook:

https://github.com/JuliaActuary/Learn/blob/master/actexp_analysis.jl

[MatthewCaseres]

Thanks, I always appreciate seeing your code. Right now I'm thinking about the right way to implement the left and right truncation for study, the issue that is on roadmap for v1.0. Might be good to implement some policy year truncation as well, to reduce the amount of space it takes up to do something like a lapse-skewness analysis for a particular duration

[MatthewCaseres]

Following up on this, I opened some issues after doing a comparison of differences with actxps. I should have the time to resolve these and will begin working on a PR.

[alecloudenback]

Thanks for the analysis! Can you differentiate from the issues that cause a material change in the exposures versus ones that are off by a single day? I'd definitely want to fix, e.g. where an exposure is off by a whole unit versus one where it's +- a single day (e.g. 1/365 difference in exposure)

[MatthewCaseres]

My best guess is that there are no material differences for any of the issues raised. It could come down to the specifics of a dataset. If the decrement almost always happens on a policy anniversary, then the issue about decrements occuring on policy anniversary dates would be material.

The main thing I am wondering about is these decrements occuring on policy anniversaries. ExperienceAnalysis.jl default behavior is to back-date them. actxps maintainer closed issue as he thinks all terminations should be assumed to occur on the exact date given - mattheaphy/actxps#22. Overall I'd like some more clarity around the meaning of a date. Compatibility with DayCounts.jl makes my preference that the intervals should be inclusive like [from, to] not workable.

My overall desire is that experience study calculations mean a particular thing and that multiple implementations will produce the exact same result in terms of the exposure intervals that are created. And so my perspective is that anything causing a different number of exposure rows between two implementations would be considered a bug, even if there is no material impact.

I'm inclined to assume the design decisions from actxps as fundamentally correct because the maintainer seems to specialize in experience studies and I have been out of the industry for some time. You would also likely know better than I, and if you have differing opinions from actxps I am hoping that some consensus can be achieved.

[MatthewCaseres]

trying to be concrete, I think the following is true

if issue_date == 2019-01-01 and term_date == 2020-01-01 and termination is a death then ExperienceAnalysis.jl would need to have an exposure interval [2020-01-01, 2020-01-02) if death is not the decrement of interest and [2020-01-01, 2021-01-01) if death is the decrement of interest.

If actxps implements end dates for intervals, mattheaphy/actxps#21, then I think it would be nice if both packages used the same conventions, regardless of what that convention is. If the end dates are inclusive ] then maybe there can be a toggle in DayCounts like inclusive=true)?

[MatthewCaseres]

I opened a PR that makes some changes. If you could take a look at it and tell me what areas need work hopefully all the issues in ExperienceAnalysis.jl can be closed.

[MatthewCaseres]

Finished. Should close all the open issues. Performance tests show slightly improved performance.