Aplicação de técnicas de machine learning para identificação de oportunidades de investimento no setor de energia elétrica brasileiro

Abstract

This study investigates the application of machine learning techniques to identify investment opportunities in the Brazilian electric power sector, focusing on stock behavior around the ex- dividend date. The research is justified by the sector's strategic importance, characterized by stable cash flows and a consistent record of dividend distribution, as well as the ex-dividend date's potential as a trigger for temporary price adjustments. The main objective is to assess whether stocks in the sector represent attractive buying opportunities after the ex-dividend date, employing Random Forest and XGBoost algorithms. The methodology followed the CRISP-DM model, including business understanding, data preparation, predictive modeling, and investment strategy evaluation. The dataset comprises historical events and stock prices of leading compa- nies from 2004 to 2023. The target variable indicates whether the dividend value was recovered within 30 days after the ex-dividend date. The models were evaluated using metrics such as accuracy, precision, recall, and F1-score, showing robust results-with XGBoost achieving 89% accuracy and an F1-score of 0.87 and for Random Forest 90% accuracy and F1-score of 0.88. The investment strategy simulation based on model predictions yielded a total return of 49.6% in the test period, an average return per trade of 4.9%, and a hit rate of 90.5%, outperforming traditional benchmarks. It is concluded that the proposed approach is effective for identifying events with high recovery potential and can enhance quantitative investment strategies, while conside- ring the limitations regarding sectoral scope, class imbalance, and the absence of transaction cost simulation.