. In this paper, we study a finite-horizon, discrete-time dynamic portfolio optimization problem with regime-switching and transaction costs. We formulate the problem as a POMDP and study the corresponding belief-MDP. We characterize some structural properties of the value function and the optimal policy and establish the existence of a no-action region. Building on these structural results, we develop a simulation-based, stochastic dual dynamic programming algorithm with adaptive belief-grid selection, where the granularity of the belief state is chosen based on the ambiguity level of hidden regimes. We conduct both in-sample and out-of-sample analysis and compare our proposed method with common heuristics and benchmarks from the literature. We identify when regime information is valuable and demonstrate the computational benefits of the adaptive belief-grid selection algorithm. We further demonstrate the scalability of our algorithm through increasing the number of assets, and examine its performance with

respect to different classes of utility functions.