Aplicação de Estratégias Híbridas de Aprendizado Profundo na Estimação de Preços no Mercado de Ações Brasileiro

• Ano de defesa: 2025
• Autor(a) principal: Orssatto, Angelo José
• Orientador(a): Brun, André Luiz
• Banca de defesa: Catarina, Adair Santa, Paetzold, Gustavo Henrique
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual do Oeste do Paraná; Cascavel
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação
• Departamento: Centro de Ciências Exatas e Tecnológicas
• País: Brasil
• Palavras-chave em Português: Séries Temporais; Bolsa de valores; Aprendizado de Máquina; Redes Neurais Convolucionais; Redes Neurais Recorrentes; Predição
• Palavras-chave em Inglês: Time Series; Stock Market; Machine Learning; Convolutional Neural Networks; Recurrent Neural Networks; Prediction
• Área do conhecimento CNPq: METODOLOGIAS E TÉCNICAS DE COMPUTAÇÃO
• Link de acesso: https://tede.unioeste.br/handle/tede/7816

In Brazil, B3 is the stock exchange where various types of assets are traded, such as company shares, commodities, currencies, property funds, among others. The time series of stock market data represents a sequence of information, which can include the opening, closing, high and low prices of a stock, collected over time. Many economists, investors and researchers try to understand its behaviour in order to forecast financial actions, which is a challenge due to its non-linearity and complexity. This work aimed to analyse the applicability of machine learning and deep learning strategies in predicting the values of a set of the most traded shares on B3. These strategies involved MLP, SVM, Bagging, Random Forest, XGBoost models, recurrent neural networks with LSTM and GRU cells, CNN and hybrid networks combining convolutional and recurrent networks. Temporal data and other financial indices was collected from the specialised website investing.com over a period from 2013 to 2023, on a daily basis, and this data was used to build the image base for feeding the CNN, in the form of a block representation, a line graph and a candlestick graph. The study aimed to create a robust and effective machine learning model, from the perspective of univariate and multivariate analyses, and to compare it with consolidated methods in the literature applied to the same problem, analysing the MAE, RMSE metrics and Pearson’s correlation coefficient. The results showed that the MLP and Bagging models achieved the bests metrics for all the scenarios tested, obtaining RMSEs below 0.4 and 0.7, respectively, and emphasising the importance of calibrating their hyperparameters. The block approach for hybrid networks proved to be very promising, and although it didn’t achieve the best error metrics, its correlations were significantly high in general, which makes it possible to explore this context further. In addition, the applicability of structured data as inputs to a CNN aimed at predicting the value of stocks was demonstrated.