Seleção de portfólio de ações por meio de otimização metaheurística híbrida com redes neurais LSTM

Abstract

The portfolio optimization problem consists in the formation of a set of assets and their respective proportions so that the risk is minimized and the return maximized. The work aims to improve the portfolio optimization model, calibrating the cardinality and reducing the error of estimating the portfolio, through the hybrid Artificial Bee Colony (ABC) algorithm with Long Short Term Memory (LSTM) Neural Networks. Regarding the methodology, two systematic literature reviews were carried out to verify the state of the art on the research topic. Two applications were also made: in the first, portfolios are built via optimization using the Markowitz model and in the second, the Markowitz model with cardinality restriction is used. The results show that, for the analyzed data, the increase in the number of assets, in optimized portfolios, does not reduce its risk. In addition, attention is drawn to the importance of using simulations and experimentation as the main approach to be used to define the ideal amount of assets in a portfolio. The hybrid ABC algorithm with LSTM Recurrent Neural Networks proved to be a potential technique for reducing the estimation error and for optimization. The limitations were: the filters used in the literature reviews may have excluded important works; in addition, the choice of assets based on their liquidity and period of participation in the stock exchange, although justified, may have avoided the analysis of assets with good chances of gains in terms of return