Generalized Binomial Estimator: a bridge between Actuarial Science and Data Science

Generalized Binomial Estimator: a bridge between Actuarial Science and Data Science


Thanks! Share it with your friends!


You disliked this video. Thanks for the feedback!

Sorry, only registred users can create playlists.


For some years now, Data Science and Machine Learning has entered the world of actuarial science and is shaking up traditional practices.

Machine Learning is usually opposed to conventional statistical methods: in many actuarial works dealing with Data Science published in recent years where, often, an insurance issue is treated on the one hand with Machine Learning and on the other with "classic" modeling, the objective being to outperform traditional approachs.

The CNP Assurances R&Data Lab propose a new model to the actuarial community: the Generalized Binomial Estimator (GBE), based on an original approach combining maximum likelihood with empirical risk minimization approachs and whose objective is both to preserve the interpretability and robustness advantages of conventional models while integrating the performance and optimality components offered by Machine Learning. Rather than oppose "classical" actuarial science and Machine Learning, aim was to reconcile them. For application, we choose one of the most emblematic insurance issues of the actuarial profession: the construction of mortality table.

In a fast moving environment where insurers have more and more data to exploit and where regulatory and competitive requirements on the quality of modeling are becoming stronger, GBE has the following properties:

- It significantly improves the quality of the estimate by correcting number of biases present in traditional survival models.

- It also has a flexible structural form that can be identical to traditional survival models. This feature is very useful in practice since, by maintaining a structural form identical to the existing one, GBE provides table in a format directly exploitable by the insurer that do not require new developments in the projection models.


GBE is nevertheless more complex to develop than traditional models and it is necessary to have a good knowledge of development frameworks in Machine Learning before implementing it.

However, once these obstacles are overcome, GBE is much more efficient than traditional survival models.


Conceptualized and developed in 2017 and 2018 within the CNP Assurances R&Data Lab, GBE is now used in production by the company for the calibration of its mortality tables used into its Solvency 2 and MCEV calculations.

Post your comment

Sign in or sign up to post comments.
Be the first to comment