Portfolio Optimization with Rebalancing: A Hybrid Approach with Genetic Algorithms and Reinforcement Learning

This work presents a framework for sequential portfolio optimization with discrete asset allocation, incorporating knowledge transfer between rebalancing periods. The problem is modeled as an extension of the Multidimensional Backpack with Repetition, aiming to maximize the Sharpe Ratio under realistic constraints of budget, cardinality, allocation, and transaction costs. Three approaches are compared: an exact one via mixed integer nonlinear programming, a discrete genetic algorithm, and a hybrid version with reinforcement learning. Experiments with real data from the Brazilian market demonstrate that the hybrid approach has greater robustness and superior performance in risk-adjusted return, especially in high volatility scenarios, evidencing its potential for practical applications in dynamic financial environments.