Machine-learning a family of solutions to an optimal pension investment problem

Battle Count: 75

The paper addresses optimal investment and consumption strategies under uncertainty, which is directly relevant to portfolio management. However, it focuses specifically on pension/retirement investment rather than active trading. The GRU-based approach to stochastic control and the loss scaling technique for parameter families could be adapted to trading strategy optimization across different risk preferences. The Black-Scholes framework and duality validation are relevant to derivatives pricing and hedging.

The implementation requires three jointly trained neural networks (main GRU, mean network, scaling network), online estimation of conditional moments, log-space transformations for numerical stability, and careful truncation of scale estimates. The gradient-based variant requires per-sample gradients which significantly increases computational cost. The architecture is deliberately over-engineered (not exploiting Markovianity) for generality. Training takes 24 hours with 131,072 scenarios per epoch.

The paper provides detailed network architecture (Figure 5), parameter ranges (Tables 1-2), full algorithm pseudocode (Algorithm 1), convergence proofs, and a provably convergent validation method (Appendix H). Training hyperparameters are specified (Adam optimizer, learning rate 0.001, 24-hour training, batch sizes). However, no code repository is mentioned, and the specific mortality tables and salary profiles used are not fully detailed.