Investigation of machine learning approaches for proxy modelling in life insurance

Details

Supervisors

Hainaut, Donatien

Faculty

Faculté des sciences

Degree label

Master [120] en sciences actuarielles, à finalité spécialisée

Abstract

The valuation of the variable liabilities in life insurance is still a complex subject in the computation of the Net Asset Value and the Solvency Require Capital, required by Own Risk and Solvency Assessment. To avoid the robust computation of those variable liabilities with a Nested Monte Carlo, current research shows that different methodologies exist. The most used is the Least Square Monte Carlo which is the one used in the company analysed in this thesis. As the world of data science and artificial intelligence is growing up, some papers show that other machine learning tools can be efficient and give accurate results if they are used within or instead of the Least Square Monte Carlo methodology. The goal of this thesis is to analyse the results given by the Least Square Monte Carlo methodology and verify if they could be improved with a neural network or a generalized linear model. Firstly those tools will replace the regression part in the Least Square Monte Carlo process and secondly they will be taken as a complete substitute. Based on the limited validation data sets available, it is shown that the results obtained by the company are for now quite good. The results obtained by the new methodologies tested show that the neural network is also convincing both in time computation and in accuracy but that does not give global improvement. However, the generalized linear model is less accurate than the current methodology. Nevertheless, the empirical distributions obtained with the current and the tested methodology show striking differences. The first one is closer to a normal distribution, while the latter is closer to a gamma distribution. Based on those results, keeping the current Least Square Monte Carlo methodology does not lead to a loss of accuracy and is efficient based on the available validation data sets. However, a further study with a significant number of validation data could be interesting to point out the general curve of the variable liabilities. The case where the neural network completely replaces the Least Square Monte Carlo could be a good analysis but also needs more research and may never be completely trusted due to a lack of real world scenarios. Moreover, the management has to take into account the fact that the neural network still is a kind of "black box" to choose the best solution.