This article addresses the valuation of issuer-callable exotic structured products featuring Bermudan-style optionality and contingent coupon and principal payoff mechanisms. These products are structurally complex, often involving multi-asset underlyings, layered barrier conditions, and highly nonlinear coupon rules. The standard valuation approach is the Longstaff–Schwartz algorithm, which uses least-squares regression to approximate continuation values but encounters difficulties in high dimensions. We propose a randomized neural network (RNN) framework where hidden layers are randomly initialized and only the output layer is trained. Numerical experiments on representative structured products show that the RNN approach achieves pricing accuracy comparable to that of fully trained neural networks while offering improved computational efficiency. The RNN-based method thus provides a scalable and effective solution for valuing complex callable structured products.
Deng et al. (Mon,) studied this question.