Long Short Time Memory in the forecast of financial indices in the Brazilian market (Ibovespa)
The present investigation seeks to evaluate different models of recurrent neural networks (LSTM, BLSTM and GRU) in comparison with the ARIMA model, whose purpose is to determine which of these models is capable of making a better forecast for the closing price of 5 steps forward in the stock index o...
| Autor: | |
|---|---|
| Tipo de documento: | dissertação |
| Estado: | Versão publicada |
| Data de publicação: | 2022 |
| País: | Brasil |
| Recursos: | Universidade de São Paulo (USP) |
| Repositório: | Biblioteca Digital de Teses e Dissertações da USP |
| Idioma: | inglês |
| OAI Identifier: | oai:teses.usp.br:tde-12042023-150020 |
| Acesso em linha: | https://www.teses.usp.br/teses/disponiveis/45/45133/tde-12042023-150020/ |
| Access Level: | Acceso aberto |
| Palavra-chave: | ARIMA model Ibovespa LSTM Modelo ARIMA Neural network Rede neural |
| Resumo: | The present investigation seeks to evaluate different models of recurrent neural networks (LSTM, BLSTM and GRU) in comparison with the ARIMA model, whose purpose is to determine which of these models is capable of making a better forecast for the closing price of 5 steps forward in the stock index of the Sao Paulo Stock Index (IBOVESPA). The optimization of the parameters allows to reduce the cost function, for this reason, 8 configurations with more than 720 simulations were studied, discovering that the ADAMAX optimizer has worked better compared to the other optimizers, presenting a lower cost function (mean square error). In the simulations of the different configurations, the average and the standard deviation of different models have been considered. The GRU model with the ADAMAX optimizer was more efficient in more than 90% of the results obtained. The final configuration was the GRU model with a batch size equal to 5, with 250 epochs, a learning ratio equal to 0.001 and with 30 neurons. This configuration presented a lower mean square error and therefore better forecasts. The LSTM, BLSTM models presented a lower cost function compared to the GRU model. Likewise, the ARIMA model did not have an optimal result compared to the recurrent neural network models. |
|---|