Deep Least Squares Monte Carlo methods for the valuation of variable annuities with guarantees

Battle Count: 85

The paper is primarily focused on actuarial/insurance pricing rather than trading strategies. However, the LSMC methodology and deep learning approaches for stochastic control are transferable to optimal execution, dynamic hedging, and portfolio optimization problems relevant to quantitative trading. The neural network regression approach for conditional expectations in high-dimensional settings has broader applicability.

The methodology involves multiple algorithm variants (Realised Value vs Regression Surface, Regress Now vs Regress Later), control randomization techniques, and requires careful feature engineering for polynomial regression. Neural network implementation is more straightforward but requires GPU acceleration and hyperparameter tuning. The backward induction with path recomputation in Realised Value adds complexity. The paper provides detailed pseudocode which aids implementation.

The paper provides detailed algorithm pseudocode (Algorithms 1-5), specific parameter settings for both polynomial and neural network regressions, exact simulation schemes for the Vasicek model, and closed-form formulas in the appendix. However, no code repository is mentioned. The neural network architecture (3 hidden layers, width 128, SiLU activation, Adam optimizer, 2000 epochs) is fully specified. Computing times are reported.