Previsão de preços de ativos listados na B3: um estudo comparativo entre modelos tradicionais e redes neurais
Data
2023-01-19
Autores
Moreira, Larissa Lopes 
Orientadores
Lucas, Edmilson Costa
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
O objetivo deste trabalho é comparar o desempenho de modelos econométricos tradicionais e de algoritmos baseados em inteligência computacional, a fim de averiguar qual das técnicas é a mais assertiva na previsão de preços de ações de três empresas: Weg, Magazine Luiza e Porto Seguro. O presente estudo compara o desempenho da aplicação conjunta dos modelos Arma e Garch com o desempenho obtido por redes neurais artificiais do tipo LSTM. Utilizou-se a série de preços de fechamento de cada ativo e os dados foram obtidos através de uma API do Yahoo Finance. Para escolher a melhor topologia da rede neural LSTM, utilizou-se o RMSE e o MAPE, enquanto que para a escolha do melhor modelo Arma-Garch, foram utilizados o AIC, o teste Ljung-Box, análises de significância dos coeficientes, além do princípio da parcimônia. Por fim, comparou-se o desempenho do melhor modelo de cada uma das duas técnicas através do RMSE e do MAPE. Os resultados indicam que as redes neurais LSTM apresentam maior capacidade preditiva que os modelos Arma-Garch para todos os ativos analisados e esse achado se alinha com outros trabalhos presentes na literatura.
The objective of this work is to compare the performance of traditional econometric models and algorithms based on computational intelligence, in order to find out which of the techniques is the most assertive in predicting the share prices of three companies: Weg, Magazine Luiza and Porto Seguro. This study compares the performance of the joint application of the Arma and Garch models with the performance obtained by LSTM artificial neural networks. The closing price series of each asset was used and the data was obtained through a Yahoo Finance API. To choose the best topology of the LSTM neural network, RMSE and MAPE were used, while to choose the best Arma-Garch model, AIC, the Ljung-Box test, significance analysis of the coefficients, in addition to the principle of parsimony were used. Finally, the performance of the best model of each of the two techniques was compared using RMSE and MAPE. The results indicate that LSTM neural networks have greater predictive capacity than Arma-Garch models for all assets analyzed and this finding is in line with other works present in the literature.
The objective of this work is to compare the performance of traditional econometric models and algorithms based on computational intelligence, in order to find out which of the techniques is the most assertive in predicting the share prices of three companies: Weg, Magazine Luiza and Porto Seguro. This study compares the performance of the joint application of the Arma and Garch models with the performance obtained by LSTM artificial neural networks. The closing price series of each asset was used and the data was obtained through a Yahoo Finance API. To choose the best topology of the LSTM neural network, RMSE and MAPE were used, while to choose the best Arma-Garch model, AIC, the Ljung-Box test, significance analysis of the coefficients, in addition to the principle of parsimony were used. Finally, the performance of the best model of each of the two techniques was compared using RMSE and MAPE. The results indicate that LSTM neural networks have greater predictive capacity than Arma-Garch models for all assets analyzed and this finding is in line with other works present in the literature.
Descrição
Citação
Moreira, Larissa Lopes. Previsão de preços de ativos listados na B3: um estudo comparativo entre modelos tradicionais e redes neurais. 2023. Trabalho de Conclusão de Curso (Bacharelado em Ciências Atuariais) - Universidade Federal de São Paulo, Escola Paulista de Política, Economia e Negócios, Osasco, 2023.