Estudo de séries temporais e suas aplicações em econofísica usando redes neurais artificiais

Abstract

This work falls within the field of Econophysics and aimed to create and evaluate the effectiveness of an Artificial Neural Network (ANN) with a Long Short-Term Memory (LSTM) architecture in forecasting time series of stock prices in the Brazilian capital market. For this purpose, historical daily data were collected for the assets Petrobras (PETR4.SA), Vale (VALE3.SA), and Itaú (ITUB4.SA) over the period from 2005 to 2025. The methodology involved the use of advanced computational models to analyze complex financial systems and included, as part of the data preprocessing, normalization using the MinMaxScaler function and structuring the data into 60-day time windows to feed a deep LSTM model. This model consists of four layers and was regularized using the Dropout technique (with a rate of 0.2) to prevent overfitting. The results on the test set showed varied performance: high accuracy was achieved for the banking sector asset (Itaú), good trend forecasting ability for the oil sector asset (Petrobras), and a quantitatively significant failure for the mining sector asset (Vale), evidenced by a notable error and a "cluster"at the end of the training epoch, occurring only for the Vale asset. It is concluded that, although LSTM networks show great potential for modeling financial time series, their effectiveness strongly depends on the particular characteristics and stability of each asset. This reinforces the idea that the financial market is a complex and non-universal system.