Estudo de aplicação de técnicas de aprendizado por reforço no problema de otimização de portfólio

Abstract

The portfolio management problem, focus of this work, consists of determining the optimal asset allocation within a wallet, in order to maximize (or minimize) one or more objectives. These objectives are usually related to risk and return metrics. In financial economics literature, this problem has been solved using portfolio optimization models, such as Markowitz, CAPM and Black Litterman, which are executed for each instant when portfolio rebalancing is necessary. This decision process, incremental and under uncertainty, can be seen as a Markovian decision process, which makes modeling under reinforcement learning paradigm attractive, this being a recent trend discussed in machine learning literature. This work aims to investigate the use of reinforcement learning technics in portfolio optimization problem. A literature review is realized, with its main learnings. In a case study, a portfolio asset weight control problem is modeled as a Markovian decision process, and reinforcement learning algorithms are used to optimize it. The implementations are made in an incremental way, aiming to demonstrate the logic behind these algorithms developments. Finally, the model’s behavior is compared with Markowitz based strategies, and the result shows that these approaches hold good performances, and have a promising use for this kind of problem.