Comparativo entre redes neurais recorrentes GRU e LSTM para a predição de instrumentos financeiros

Abstract

This work aims to compare two models of recurrent neural networks, for the prediction of quotation of financial instruments, considering criteria such as coefficient of determination and correlation coefficient of predicted versus actual data. Comparative analysis of the LSTM (Long short-term memory) and GRU (Gated Recurrent Unit) model was performed based on a recurrent neural network. The work compared the two models in an identical recurrent neural network architecture for both models. Python-based algorithms using the Keras, Numpy Pandas and Scikit-learn libraries were used. The quotation of the Ibovespa Index for the period from April 28, 1993 to April 1, 2021, broken down by the daily quotation, served as input. The research results showed a better performance of the GRU model, which using the 200 previous closings of the Ibovespa Index as predictors, 300 times for training and using the Relu activation function was able to explain about 98% of the variations of the real data.