Deep learning in life insurance risk prediction

Abstract

This diploma thesis deals with the implementation of an artificial neural network to predict life insurance risks. At first, general terms of deep learning are declared and defined. A brief insight into life insurance risk, specifically into the crucial parameter, the probability of dying, is given. Consequently, the structure of a deep neural network, the different activation functions and optimisers are explained in detail. This also includes a precise explanation of the training algorithm of a deep neural network. Finally, the calculation of the probability of dying is performed. Several experiments are carried out to test different scenarios for the neural network and the simulations are thoroughly analysed. In conclusion, the calculation of the probability of dying via an artificial neural network worked exceptionally well. A model with four hidden layers, overall 640 neurons, the Adam optimiser and either the ELU, TanH or Softplus activation function yielded by far the best results for this problem.