Cost-Efficient Asset Allocation: Graph-Based Machine Learning for Dynamic Portfolio Rebalancing.

This research introduces a novel approach to portfolio rebalancing by integrating Graph Neural Networks (GNNs) with Dijkstra's algorithm to optimize transaction costs in financial markets. GNNs are trained on historical stock data from major technology companies to predict future transaction costs, capturing complex dependencies between assets and market conditions. These predicted costs are then embedded as edge weights in financial asset graphs, enabling a dynamic representation of transaction expenses within the portfolio structure. Using this enriched financial network, Dijkstra’s algorithm is applied to determine the most cost-efficient paths for asset capital reallocation. By leveraging this hybrid framework, portfolio managers can systematically identify low-cost trading routes, reducing slippage and improving execution efficiency, particularly in high-frequency trading environments. Empirical results demonstrate that this approach significantly minimizes transaction costs compared to traditional rebalancing strategies, highlighting the synergy between machine learning and graph-based optimization in financial decision-making. The study underscores the potential of AI-driven portfolio management techniques in enhancing capital efficiency and reducing execution risk. JEL: C61, G11, C63, G17, C45.