In our experience there is no substitute for doing – this resource allows you to get started implementing a neural network on individual claims data.
Developed initially as an internal framework for the working party, to introduce consistency and allow us to compare results and assess the success of different models, our series of blogs also provides an off-the-shelf package which is ready to go, saving individuals time by not having to write everything from scratch.
We are sharing it more widely to:
Enable reserving actuaries to gain knowledge and skills. We know there is great enthusiasm for learning machine learning in the actuarial community, and this is designed to make it much easier for many to start implementing a neural network.
Speed up our collective knowledge of how to use machine learning for reserving. We hope this will encourage ‘open source research’ where groups of us can share our findings to move things on more quickly. We are particularly keen that reserving actuaries try this out on their own data as we believe the real breakthroughs will come from insights learned on real rather than simulated data.
To get the most out of it we encourage experimenting. What happens when you try using different input variables, or altering various parameters of the model? In this way you can build up knowledge and judgement of how to use these models: What model works best on what data? Which hyperparameters are likely to work best in which situations?
The blogs should be seen as living documents, in the sense that the forum is being continuously updated to reflect the results of our own investigations. We would love to learn from you too. Feel free to get in touch or better still join the working party. Details of how to do this are shown at the end of this page.
The blogs
Explainer: Takes code to implement a neural network and steps you through / explains what it is doing.
Diagnostics: Provides examples of outputs you can use to understand what the model is doing and assess its effectiveness. Gives details of some of the different tools you can use to produce them, and steps you through how to create them. We intend to add to the Diagnostics blog over time.
Hyperparameters: Takes individual hyperparameters, shows what happens when you vary them and steps you through how to change them in the code so you can have a go yourself.
Points of note
Although a theoretical understanding of the working of neural networks is assumed, links to a number of learning resources can be found in the Explainer blog.
The Python programming language was selected since it can leverage all the major machine learning and artificial intelligence libraries, while having one of the lowest learning curves and being the most popular. That said, a knowledge of Python is not a prerequisite, as the Explainer blog takes you right from the beginning of how to set up your Python environment.
The examples make use of relatively simple simulated data, and arguably using a neural network on this data is overkill. This is an introduction and we hope you will start to use it on your own data where we think the real breakthroughs will occur. However simulated data is still useful as it allows one to experiment and find the relative strengths and weaknesses of a particular model.
Learning about machine learning and artificial intelligence is very much a hands-on process, where one has to set aside time to code the models, experiment and study the topic.
Other resources
To get more of an idea of what the blogs are about, you might find our presentation that we gave at the 2024 GI Spring Conference useful A practical introduction to neural networks.
Are you wondering why we have specifically used a neural network? See Jacky Poon’s February 2024 article in The Actuary Data Science lab - GBMs or neural networks.
Get in touch
See our blog site: Machine Learning in Reserving Working Party
Find us on Communities (a great place to ask us questions): MLR WP - General Insurance - IFoA Communities
LinkedIn: IFoA General Insurance Machine Learning in Reserving | Groups | LinkedIn
To join the working party simply apply through the IFoA’s volunteering page Volunteer vacancies (actuaries.org.uk) (scroll down to find us under ‘Ongoing vacancies’). All levels of experience are welcome, we just ask that you commit time and actively contribute to the working party.