Uma avaliação sistemática de técnicas de aprendizado de máquina baseadas em ensemble para previsão de índices do mercado de ações usando séries temporais financeiras

Abstract

The prediction of time series indices in stock markets has been garnering increasing interest, providing investors with valuable information about a country’s future economic trends in the short, medium, and long term. This type of prediction has been widely performed through machine learning models. However, due to the volatility, noise, and stochasticity of the data, single models often present limitations in terms of accuracy. To mitigate these issues, ensem ble techniques have emerged, combining multiple models to increase the robustness and accu racy of predictions. Ensemble approaches exploit the diversity of models and machine learning techniques, leveraging the individual characteristics of each to achieve more reliable results than those generated by isolated models. Thus, a systematic literature review was conducted focusing on stock index prediction using financial time series and ensemble approaches, with the aim of mapping the main articles, authors, types of techniques used, and gaps in the literature. The review reveals that many comparative analyses are limited to the use of traditional performance metrics, such as mean squared error, root mean squared error, mean absolute error, and mean absolute percentage error, which can introduce biases in comparisons. Furthermore, these analyses often omit robust statistical tests, focus on only one type of market,and adopt nadequate comparison protocols. This study investigates how different ensemble techniques can enhance prediction in financial time series, adopting a meticulous protocol to eliminate data biases and standardize comparisons between methodologies. In addition to tra ditional metrics, a cost-benefit analysis was introduced for a more comprehensive evaluation of ensemble architectures. The Wilcoxon hypothesis test was applied to validate the findings,along with the Friedman and Nemenyi tests. The results highlight the importance of running algorithms in different environments, such as the selected IBOVESPA and S&P 500, as there are divergences in the performance of evaluation metrics and statistical tests, emphasizing the need to use the cost-benefit metric to verify whether there is a real gain in the overall performance of the architecture compared to single models. This establishes a framework for constructing and evaluating ensemble approaches. With these results, the aim is to con tribute to the development of more robust and effective methodologies in the use of ensemble techniques for stock index prediction, ensuring a better understanding of the advantages and limitations of these techniques in different market environments.