Efficient Machine Learning Methods for Risk Management of Large Variable Annuity Portfolios ∗

Variable annuity (VA) embedded guarantees have rapidly grown in popularity around the world in recent years. Valuation of VAs has been studied extensively in past decades. However, most of these studies focus on a single contract. These methods cannot be extended to valuate a large variable annuity portfolio due to the computational complexity. In this paper, we propose an efficient moment matching machine learning method to compute the annual dollar deltas, VaRs and CVaRs for a large variable annuity portfolio whose contracts are over a period of 25 years. There are two stages for our method. First, we select a small number of contracts and propose a moment matching Monte Carlo method based on the Johnson curve, rather than the well known nested simulations, to compute the annual dollar deltas, VaRs and CVaRs for each selected contract. Then, these computed results are used as a training set for well known machine learning methods, such as regression tree , neural network and so on. Afterwards, the annual dollar deltas, VaRs and CVaRs for the entire portfolio are predicted through the trained machine learning method. Compared to other existing methods (Bauer et al., Gan, Gan and Lin, 2008, 2013, 2015), our method is very efficient and accurate, especially for the first 10 years from the initial time. Finally, our test results support our claims.