Aplicação de aprendizado de máquina na construção de carteiras de ações de longo prazo: uma abordagem comparativa com modelos tradicionais

Abstract

Long-term stock investing attracts the attention of investors seeking effective and straightforward methods to select the securities to compose their portfolios, as well as to identify assets with the highest potential for future appreciation. Over time, many approaches have been developed for this purpose. In this context, this study aimed to apply Machine Learning (ML) techniques to estimate future stock values and, this way, construct a portfolio capable of outperforming the traditional equity market. With the goal of selecting assets for a portfolio and predicting their profitability, ML models can be used to perform stock price regression and recommend highly profitable investments. This approach enables the formation of a diversified investment portfolio that reduces risk while still outperforming the market. In this study, a methodology was developed, analyzed, and tested for constructing a stock portfolio with a one-year allocation period using the following ML models: Ridge Linear Regression (LR), Bayesian Regression (BR), Regression Tree, Support Vector Regression (SVR), and Gradient Boosting Regression (GBR). Additionally, all models were trained and tested both with and without hyperparameter tuning using the Grid Search CV method to determine whether this adjustment enhances model performance, thereby recommending even better portfolios. The portfolios generated by the ML models had their returns compared to traditional financial market models for asset recommendation, such as Modern Portfolio Theory and the S&P 500 market index. In the end, the results showed that, in the vast majority of tests conducted, the ML-based methodology proposed in this study outperformed the market-based models.