Does reinforcement learning outperform deep learning and traditional portfolio optimization models in frontier and developed financial markets?

Advancements in machine learning have opened up a wide range of new possibilities for using advanced computer algorithms, such as reinforcement learning in portfolio risk management. However, very little evidence has been provided on the superior performance of reinforcement learning models over traditional optimization models following the mean-variance framework in different financial market settings. This study uses two experiments with data from the Vietnamese and U.S. securities markets to justify whether advanced machine learning models could outperform traditional portfolios' cumulative returns while optimizing the Sharpe ratio. The results suggest that reinforcement learning consistently outperforms the established methods and benchmarks in both experiments, even when using a very similar degree of diversification in portfolio construction and the same input data. This study confirms the ability of reinforcement learning to provide dynamic responses to market conditions and redefine the risk-return standard in the financial system.