Surrender analysis in life insurance

Research project 2019-2021.
Accompanying code for the paper "Modelling surrender risk in life insurance: theoretical and experimental insight" by M. Kiermayer.
Paper available on https://www.tandfonline.com/doi/full/10.1080/03461238.2021.2013308 (publication) or https://arxiv.org/abs/2101.11590 (preprint).

Goal

1. Perform extensive experiments to analyze the capabilities of several models (including logistic regression, random forest, XGBoost, neural networks bagged or boosted) to estimate surrender probabilities.
   
2. Check the effect of resampling on model performance and predicted surrender probabilities
   
3. Investigate a time-dependent evaluation of surrender rates, including confidence bands
   

Structure of the project

• The simulated data can be found in the directory "./Data", including the portfolio "Portfolio.csv" at time t=0.
  * Sub-directories "./Data/profile_{i}" include time-series data for years t>=0 wich is unique to surrender assumptions in the ith-profile, i = 0,1,2,3.
  * All data in these (sub-)directories is generated (and analyzed) by the scripts "_data{i}_{..}.py"
  

• The files "HPSearch_{..}.py" implement an automated hyperparameter-tuning (based on the python package 'hyperopt')
  

• The directories "./profile_{i}" contain the results for trials of HPTuning for all models and the resulting best model-parametrizations
  

• The main files analyzes all models (given the parametrization after HPTuning)

• Visual and statistical analyses of our experiments are saved in either "./Plots" or "./Tables"

• All helper-functions can be found in "./functions"