A Learned Model-Based Simulation Approach to Enhanced Portfolio Management

The proposed approach introduces an innovative portfolio management strategy focused on single-asset selection, designed to overcome the limitations of conventional diversification approaches, which often dilute allocations to high-potential assets and hinder optimal return maximization. This framework leverages a Monte Carlo Tree Search (MCTS)-driven exploration mechanism to generate diverse market scenarios and systematically accumulate experiential data, thereby enabling the development of an optimized asset selection policy. Acknowledging the inherent complexities of portfolio management—where conventional MCTS architectures, such as those used in AlphaGo, may inadequately capture the complexity of experiential insights, the approach integrates foundational principles from the MuZero algorithm. By decoupling observational data from environmental state representations, it supports more precise and adaptive decision-making processes. Furthermore, the Simulation phase of the search tree is refined through a Dynamics-function-based predictive model, facilitating the anticipation of multiple potential future states for each action, thereby enhancing the system’s adaptability to dynamic market conditions. Empirical validation conducted on a sample of 29 Dow Jones Industrial Average (DJIA) stocks across various time horizons demonstrates the superiority of the proposed approach. Key performance metrics including return, Sharpe ratio, Sortino ratio, and Information ratio consistently reveal its ability to outperform conventional methods, underscoring its significant potential to advance portfolio management practices and achieve superior return optimization.