Grouping-in-insurance-with-neural-networks

Research project 2019-2020.
Accompanying code to paper "Grouping of contracts in insurance using neural networks" by M. Kiermayer and C. Weiß.
Paper available at https://www.tandfonline.com/doi/abs/10.1080/03461238.2020.1836676 .

Goal:

Create a surrogate portfolio (-> grouping) that is (in some sense) simpler than an existing portfolio, but displayed characteristics that are as close as possible the the original portfolio.
The main motivation for this objective is the Solvency II directive and computationally highly expensive simulations. Simplifying the portfolio leads to time savings and is in practice essential.
A naive approach for such a grouping is the unsupervised K-means clustering algorithm and its cluster assignment. We improve this baseline by including an economic supervision (by neural networks). Further, our approach can also be used for fuzzy clustering.

Methodology

This grouping procedure is based on a 2-step approach where
i) a prediction model for characteristics of contracts is fitted
ii) the grouping is performed, given a fixed prediction model that supervises the resulting aggregated characteristics

Content of the github-project:

1. Generation of (realistic) data
   

   • Term life insurance contracts
     + for training a prediction model (-> "SUB_DATA_TL.py")
     + for performing a grouping procedure (-> "SUB_DATA_TL_NEW_Portfolio.py")
     
   • Pension contracts, i.e. defined benefit plans
     + for training a prediction model (-> "SUB_DATA_Pensions.py")
     + for performing a grouping procedure (-> "SUB_DATA_Pensions_NEW_Portfolio.py")
     

2. Calibration and analysis of the prediction models
   

   • For term life insurance: MAIN_Prediction_TL.py
     
   • For pension contracts: MAIN_Prediction_Pensions.py
     

3. Grouping procedure
   

   • For term life portfolio: MAIN_Grouping_TL{}.py
     
   • For pension portfolio: MAIN_Grouping_Pensions.py
     

4. Sensitivity analysis of the investment surplus for the term life portfolio (-> MAIN_Sensis_TL_SCR.py)
   

Note on 1): See references for realistic distribution of the data in the paper.