Previsão do preço de ações da Petrobras utilizando redes neurais artificiais

Abstract

The stock market represents an investment alternative with high return potential, but permeated by inherent uncertainty and volatility. Conventional approaches, such as statistical methods and technical and fundamental analysis, prove to be limited in identifying non-linear relationships between the variables that affect stock prices and their upward and downward movements. Therefore, the application of more advanced techniques, such as Artificial Neural Networks (ANN), is necessary for a more comprehensive and accurate analysis of the stock market, which includes a wide variety of data, which is highly volatile and follows non-linear patterns over time, forming complex time series. Because of this, predicting market fluctuations becomes a significant challenge. In this study, the price of shares on the market serves as a source of information for investors in decisions relating to the allocation of resources, strategic planning and the execution of financial transactions. Thus, this paper investigates the application of artificiais neurais networks in forecasting the stock market of Petrobras, one of the world's largest energy companies. This is a challenging task due to the complex and highly volatile nature of the company, and the use of an ANN is justified by the difficulty of traditional statistical methods in achieving accurate long-term forecasts in economic time series. To this end, this paper describes the development and implementation of a recurrent neural network (RNN) model using the TensorFlow library, which employs a sequential architecture composed of bidirectional Long Short-Term Memory (LSTM) layers, interspersed with Batch Normalization and Dropout layers, in order to mitigate overfitting and promote maximization of generalization. The findings of this study reveal the significant potential of Artificial Neural Networks in predicting the performance of the Petrobras stock market. By capturing complex, non-linear patterns in financial data, it is possible to make informed financial decisions. The results obtained show the ability of a properly structured ANN to identify and represent intricate and non-linear mappings, often found in financial time series. The results of this study point to the RBF model as the most effective for predicting time series, due to its lower error and high consistency. The LSTM and ELM models also demonstrate high performance, combining accuracy and stability. Despite being a viable alternative, MLP has limitations in terms of robustness. ARIMA, on the other hand, has good stability but lower predictive reliability, making it a secondary option. RBF is therefore recommended as the preferred choice, with LSTM and ELM as promising alternatives.