EcoSta 2022: Start Registration
View Submission - EcoSta2022
A0633
Title: A residual network for valuing large portfolios of variable annuities Authors:  Heng Xiong - Wuhan University (China) [presenting]
Abstract: The valuation of large variable annuity portfolios is a central concern for insurers considering that the commonly used Monte-Carlo (MC) simulation is computationally intensive. A spatial interpolation method was developed recently to significantly reduce the calculation time for valuation. However, such a method relies heavily on a predefined distance function. Thus, it is replaced by a neural network (NN) strategy that could select the optimal distance function automatically. We present the residual portfolio valuation network (ResPoNet), which outperforms the traditional NN by adding a loss of weight item. ResPoNet maintains the universal meaning of the distance function. The high performance of ResPoNet is also due to the insertion of a residual connection into the network training process, which in turn enables the network to learn the attributes of insurance policies. Our numerical experiments illustrate that the proposed approach effectively smooths the training process and significantly improves the accuracy of the valuation when benchmarked to the original NN.