What kind of projects would you like to see here?

[ResearchActuary]

As the title states. Our vision for this site is that it will become the de facto location for open collaboration on non-life actuarial software. To that end, we have a number of repos which address common actuarial use cases. Are there others? If so, what are they?

[Tom-Holmes]

I would like to see collaboration on creating modeling, data science, and coding best practices.

• A series of case studies highlighting common modeling problems and solutions. This can be GLM, Penalized Regression, GBM, Neural Net ... anything.
• Demonstration of implementing methodologies like cross validation and proper data splitting.
• Code with proper comments and explanations of best practices like where and how to document actuarial assumptions.
• Code checking best practices. For example, a few sample pull requests with errors that an actuary could find through checking and testing. Additionally, discussions on creating "test code" to verify that calculations or manipulations are behaving as intended.

Actuarial projects require robust review and documentation, so I think starting at best practices and knowledge sharing of data science concepts may be a good place to start to make sure that everything created is as standardized as possible and reviewed as thoroughly as possible.

[ResearchActuary]

Very good ideas! Do you have thoughts about how this could be done via GitHub? Would something like a blog/website achieve this?

[Tom-Holmes]

I think the case studies and demonstrations would be straightforward on GitHub with a ReadMe and good comments providing the background. Similarly, a few pull requests with errors to test could be provided (as long as we don't push the solutions).

You're right that a blog would be better for developing best practices as we could have small updates and discussions in comments along the way. In the end, I would envision codifying the advice in a website for easy browsing.

[KiY0On]

Upvoted this. I think it's hard for some people who are starting to use and adapt Python in their actuarial work projects to grasp what the best practices are. It will be really good to have blog posts for people to see.

[jbogaardt]

It would be awesome to see a non-life suite of packages guided by CAS similar to what the astronomy community has done with Astropy and its suite of interoperable affiliated packages. This would promote:

• Standards by which actuarial research can be implemented
• Reproducibility of actuarial research
• Access to tools for actuarial community members in less mature insurance markets
• Teaching tools for those who learn through tinkering

[ResearchActuary]

Mechanically, what approaches would support that in GitHub? Would something like submodules make sense?

[jbogaardt]

I didn't really have a specific GitHub feature in mind in facilitating interoperable packages. Submodules may have its place. Releases to CRAN and Pypi might be sufficient. I was thinking at a much higher level.

I see GitHub as simply a medium for a community with common interests (non-life actuarial) to collaborate on software for the broader good. I see the CAS account as imposing some de facto leadership on making that collaboration fruitful and sustainable. I see an interoperable suite of packages as a tangible manifestation of these points.

[genedan]

I think there's a lot of actuarial literature out there that hasn't been turned into usable packages yet. While the chainladder packages cover much of the reserving literature, at least to my knowledge it seems like most other actuarial domains (pricing, ecm, etc.) have been untouched. So any packages, or attempts to make them, covering these other areas would be a win in my opinion.

I think don't think packages have to focus on the latest/trendiest topics (such as DS/AI). There would be a significant benefit in tackling mundane, but still difficult/important problems. For example, simply storing information about reinsurance transactions is a struggle and a coworker and I joked that we could write a significant paper on just the naming conventions and standards that can be implemented in organizing these transactions. It wouldn't involve fancy statistics, but writing such a paper and building a package would be an important contribution from a practical standpoint.

[ResearchActuary]

Makes sense to me. One of the challenges for me is trying to suss out where the sweet spot is for actuarial methods. Some years back, I wrote an R package which would support fixed and random effects modeling for reserve analysis. It included S4 classes to define origin year and lag. The object orientation was pretty cool. But this necessitated an interface to functions like gls() or lmer(), which wanted a data frame. (The data frame was easy to come by; it was one of the slots in the S4 class.) These days, I'm unlikely to program it that way. I'd welcome some helper functions to create a skeleton data frame, but that's about as far as I'd take it. The model methods are well-established and there's little sense creating a non-data frame structure if that's not what the methods are using. Similar issue with ratemaking. GLMs are a solved problem and most ratemaking data will be in something like a data frame, which is what glm() expects.

Are there particular articles in the actuarial literature that look like they would benefit from an open source implementation? Mildenhall's aggregate package would be a good example. I'm sure there are others.

[genedan]

I think something cool would be a workers' comp package. I have no experience in the area, so it is unlikely that I myself would tackle it in the near future, but I imagine the process would go something like this:

An actuary decides that a good candidate paper to wrap in a package would be the Brosius paper on Table M construction. It's not too long to read, but the calculations are tedious and it looks like something that can be encapsulated into a collection of classes and functions to ease the burden. They make the package, publish it, and write about it. And then let's say another actuary gets inspired and decides they want the Teng paper codified so they start contributing, and yet another actuary decides to add features from Mahler's paper on wc excess ratios, etc., and then eventually there's enough momentum going that the effort morphs into a comprehensive package on WC pricing.

People in the industry start noticing the package, they try to use it, and they run into bugs and other issues on usability, so at this stage it's important to have a platform for discussion and ways to contribute, so that the package continues to evolve to better suit its users needs. A major downside is unlike papers, which allow authors to move on to other things after publication, packages like these require ongoing maintenance, either by the original author or volunteers who pick up the work, if they are to be used seriously at work (imo).