Stratified adaptive sampling for derivative-free stochastic trust-region optimization

Battle Count: 55

Highly relevant for portfolio optimization with market impact, option hedging, and model calibration in quantitative finance. The paper explicitly addresses portfolio management problems (PM), CARA utility optimization, and option pricing calibration. The derivative-free nature is particularly relevant for complex financial models where gradients are unavailable. However, the method is general-purpose optimization rather than trading-specific, and the numerical portfolio example is relatively simple (single-period, two options).

The algorithm requires: (1) constructing a Rosenblatt/inverse transform map for the distribution of X, (2) implementing stratified sampling with proper strata boundaries, (3) building DQ-SPIM polynomial interpolation models, (4) adaptive sample size determination via sequential stopping rules, (5) trust-region update mechanisms. The data-driven PCA-based alternative (Algorithm 3) simplifies high-dimensional cases but lacks convergence guarantees. Overall, moderate-to-high implementation complexity requiring careful numerical handling of variance estimation and stopping criteria.

The paper provides detailed algorithms (Algorithms 1-3), explicit parameter settings for numerical experiments (Table 2), and clear mathematical formulations. However, no code repository is mentioned. The numerical experiments use simple toy problems (Ex1-Ex3) and a portfolio example (PM) with specified parameters, making reproduction feasible but requiring implementation effort.