Money-Back Tontines for Retirement Decumulation: Neural-Network Optimization under Systematic Longevity Risk

Battle Count: 72

The paper is primarily focused on retirement decumulation and actuarial product pricing rather than active trading strategies. However, it is relevant to quantitative finance in several ways: (1) the NN-based stochastic control framework for portfolio optimization with CVaR constraints is transferable to trading contexts; (2) the international diversification analysis with four-asset portfolios provides insights for multi-asset allocation; (3) the jump-diffusion and bootstrap simulation methods are standard in quantitative finance; (4) the CVaR optimization methodology applies broadly to risk-managed portfolio construction. The tontine/MBG product design aspects are more actuarial than trading-oriented.

High complexity due to: (1) NN parameterization of multi-dimensional state-dependent controls with constrained output mappings (scaled sigmoid, softmax); (2) 30-year horizon with annual decision times requiring sequential simulation; (3) stochastic mortality calibration (LC/CBD via StMoMo) with 256,000 simulated mortality sequences; (4) stationary block bootstrap for four-asset return simulation preserving cross-sectional and serial dependence; (5) transfer learning across a 14-point gamma grid; (6) separate Monte Carlo pricing of MBG overlay with death-time simulation; (7) contract pooling approximation; (8) benchmark validation against PIDE methods. Requires significant computational resources (256,000 paths, 30,000 training iterations, multiple gamma values).

The paper provides detailed NN architecture (Table A.1), training parameters, transfer learning procedure, full algorithm for MBG pricing (Algorithm 7.1), data construction details (Appendix C), and benchmark validation against PIDE methods (Appendix B). Data sources (HMD, Bloomberg, RBA, CRSP) are publicly available. However, no code repository is provided, and the 256,000-path Monte Carlo simulations and 30,000-iteration training require significant computational resources. The stationary block bootstrap implementation details are partially specified.