In this article we study the evaluation of American options with stochastic volatility models and the optimal fish harvesting decision with stochastic convenience yield models, in the presence of drift ambiguity. From the perspective of an ambiguity averse agent, we transfer the problem to the solution of a reflected backward stochastic differential equations (RBSDE) and prove the uniform Lipschitz continuity of the generator. We then propose a numerical algorithm with the theory of RBSDEs and a general stratification technique, and an alternative algorithm without using the theory of RBSDEs. We test the accuracy and convergence of the numerical schemes. By comparing to the one dimensional case, we highlight the importance of the dynamic structure of the agent’s worst case belief. Results also show that the numerical RBSDE algorithm with stratification is more efficient when the optimal generator is attainable.
Multidimensional optimal stopping problems arise in applications including the valuation of American and Bermudan-style derivatives, real options, and sequential investment decisions under uncertainty. Although simulation-based dynamic programming methods have substantially extended the range of tractable problems, their computational performance depends critically on the approximation space used to represent continuation values. Existing approaches typically rely on generic function spaces, including polynomial regression, splines, kernel methods, and neural networks, whose conditional expectations must be estimated numerically. This paper proposes a projection framework based on a tensor-product basis of European contingentclaim payoff functions. Rather than using generic approximation architectures, we construct an approximation space whose basis functions are analytically tractable contingent claims. The proposed basis reflects the option-like nonlinearities inherent in optimal stopping value functions while exploiting the analytical structure of the stochastic model. Motivated by the static replication methodology of Carr and Madan, the resulting approximation space coincides with the classical truncated-power spline basis, combining established approximation properties with analytical tractability. Consequently, the Bellman recursion separates into an analytical valuation step and a numerical projection step. We derive semi-closed-form continuation-value representations under correlated GBM using exponentially tilted multivariate lognormal moments and extend the framework to affine Gaussian statespace models. Numerical illustrations include three-dimensional Bermudan geometric put and basket call options, and a forestry harvesting problem with stochastic timber prices, convenience yields, and biological growth. By establishing a direct connection between static option replication, spline approximation, and dynamic programming, the proposed methodology opens a new direction for constructing approximation spaces in multidimensional optimal stopping and advocates designing approximation spaces that jointly exploit the geometry of value functions and the analytical structure of the underlying stochastic model.