Modelos de deep learning e previsão de preço de ações: estudos de casos da bolsa brasileira.

Abstract

The objective of this work is to verify and compare the univariate and multivariate RNN deep learning Recurrent Neural Network models called in this work as RNN-U and RNNM and univariate and multivariate LSTM models called LSTM-U and LSTM-M, considering as a case study the actions of PETR3, PETR4, ITUB4 and ABEV3. Two were the contributions of this study: first, to propose and evaluate by the variance ratio test whether the selected stocks follow a Random Walk stochastic process or the validity of the HME in its weak form; second, to propose and apply the cross-validation technique for time series to assess the predictive power of univariate and multivariate RNN and LSTM models in predicting stock prices. The empirical results of the variance ratio test showed that the returns of shares of PETR3 and PETR4 present negative serial autocorrelation of the first order, while the returns of shares of ABEV3 and lTUB4 also present negative serial autocorrelation of higher orders. Indeed, as the RNN LSTM-U and RNN LSTM-M models have short (1 period) and long (multiple periods) memory, they reported greater predictive power in terms of the various metrics used than the RNN-U and RNN- M that have only short or 1-period memory, in the case of ABEV3 and lTUB4 stocks