Reinforcement Learning para Optimización de Portafolios:

This work addresses the challenge of portfolio optimization using algorithms based on Multi-Armed Bandits (MAB), offering an alternative to classical approaches such as Markowitz theory and the CAPM. In particular, it analyzes the advantages of Orthogonalization methodologies with Adaptive Discounted Thompson Sampling (ADTS) and Combinatorial Adaptive Discounted Thompson Sampling (CADTS), which are adaptations of traditional Thompson Sampling designed to model asset correlation and the non-stationarity of returns. Through experiments with synthetic data, we show that under certain return distribution behaviors, ADTS can adapt to the non-stationarity of returns and learn to optimize in this context. We also apply the algorithms to real data and demonstrate that these methods can outperform traditional models under specific conditions.