Aplicações de redes neurais na previsão de operações day trade
Arquivos
Data
Autor(es)
Orientado(es)
Título da Revista
ISSN da Revista
Título de Volume
Editor
Abstract
This monograph addressed the development and analysis of four neural network mod els: Autoencoders, Convolutional Neural Networks, Long Short-Term Memory (LSTM) Neural Networks, and Multilayer Perceptron (MLP) Neural Networks, aiming to evaluate their performance in day trading operations. The results highlighted the complexity of implementing neural networks in the financial environment, with each model exhibiting unique characteristics and significant trade offs. The study emphasized the importance of adaptive and careful approaches in choosing and implementing neural networks in trading systems, recognizing the absence of a universal solution. In this work, these models were implemented, tested, and evaluated using metrics such as profit obtained, number of trades, percentage of profitable trades, profit factor, and recovery factor. The models were trained using data from January 2018 to January 2023, with 70% of the collected data allocated for training and 30% for testing. Five features were used as input for each implemented network: opening price, closing price, high, low, and a 21-period moving average. At the end of the study, it was concluded that the best neural network model among those analyzed was the autoencoder. It achieved the highest profit and, in most tests, the best results according to the adopted metrics. This does not indicate that the other models are poor or cannot be used in the financial market, but rather that for the 15-minute time frame, the autoencoder was the most effective model and the best option in this scenario.
