Loading...
Publication
Modelos Estatísticos e Redes Neuronais na Previsão do Índice da Bolsa Brasileira
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 4.12 MB | Adobe PDF |
Authors
Abstract(s)
Os estudos na área de tecnologia para investimentos têm sido objeto de interesse no meio
académico e nos negócios. A facilidade em obter um histórico de dados em maior volume,
velocidade e variedade é um dos principais impulsionadores dos avanços nesta área. O
advento da inteligência artificial levou os investigadores a explorarem modelos preditivos
para compor a inteligência de negócios, mostrando grande potencial de apoio nas decisões
humanas, que podem ser melhor suportadas em estudos mais elaborados e desenvolvidos
recentemente.
Esta dissertação aborda a previsão dos valores da Bolsa de Valores Brasileira, centrando-se
no Ibovespa, que representa as principais ações negociadas na B3. O estudo tem como
objetivo prever os valores de fecho e o retorno do índice através da aplicação de técnicas de
aprendizagem automática, comparando os resultados obtidos com modelos estatísticos
tradicionais. A investigação envolve diferentes métodos, procurando identificar quais
abordagens oferecem melhores resultados preditivos. Estes resultados visam contribuir para
uma compreensão mais aprofundada das dinâmicas do mercado financeiro brasileiro,
podendo ser úteis tanto para o meio académico quanto para profissionais do setor financeiro
interessados em realizar investimentos na área financeira. Para a realização do projeto, foram
testados modelos estatísticos e modelos baseados em redes neuronais com o intuito de
comparar os resultados de diferentes abordagens preditivas. Entre os modelos estatísticos
testados estão os modelos ARIMA para valores de fecho e os modelos GARCH e E-GARCH
para retornos. Para os testes com redes neuronais, foram escolhidas as redes Long Short-
Term Memory (LSTM) e as Gated Recurrent Unit (GRU), as redes Multilayer Perceptron
(MLP) e o Neural Basis Expansion Analysis Time Series Forecasting (N-BEATS), todas
com diversas parametrizações. Após diversos testes, atingiu-se uma estabilidade com Mean
Absolute Percentage Error (MAPE) próximo de 1% para os valores de fecho na maioria dos
modelos.
Studies in the technology for investment area have been the subject of interest in both academic and business circles. The ease of obtaining historical data in greater volume, speed and variety is one of the main drivers of advances in this area. The advent of artificial intelligence has led researchers to explore predictive models to compose business intelligence, showing great potential to support human decisions, which can be better supported in more elaborate studies recently developed. This dissertation features the prediction of fluctuations in the Brazilian Stock Exchange, focusing on the Bovespa index, which represents the main stocks traded on B3. This study aims to predict the closing values and the return of the index by applying Machine Learning techniques, comparing the results obtained with traditional statistical models. The research involves different methods, seeking to identify which approaches offer better predictive results. These results aim to contribute to a deeper understanding of the dynamics of the Brazilian financial market, which can be useful both for academia and financial sector professionals interested in investing in technology for the financial area. To carry out the project, statistics and Machine Learning models were tested to compare the results of different predictive approaches. Among the tested statistical models are ARIMA models for closing values and GARCH and E-GARCH models for returns. For tests with Machine Learning models, the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) recurrent neural networks models, multilayer models such as the Multilayer Perceptron (MLP), and the Neural Basis Expansion Analysis Time Series Forecasting (N-BEATS) model were chosen, all with various parameterizations. After several tests, stability was achieved with a Mean Absolute Percentage Error (MAPE) near to 1% for closing values in most of the models.
Studies in the technology for investment area have been the subject of interest in both academic and business circles. The ease of obtaining historical data in greater volume, speed and variety is one of the main drivers of advances in this area. The advent of artificial intelligence has led researchers to explore predictive models to compose business intelligence, showing great potential to support human decisions, which can be better supported in more elaborate studies recently developed. This dissertation features the prediction of fluctuations in the Brazilian Stock Exchange, focusing on the Bovespa index, which represents the main stocks traded on B3. This study aims to predict the closing values and the return of the index by applying Machine Learning techniques, comparing the results obtained with traditional statistical models. The research involves different methods, seeking to identify which approaches offer better predictive results. These results aim to contribute to a deeper understanding of the dynamics of the Brazilian financial market, which can be useful both for academia and financial sector professionals interested in investing in technology for the financial area. To carry out the project, statistics and Machine Learning models were tested to compare the results of different predictive approaches. Among the tested statistical models are ARIMA models for closing values and GARCH and E-GARCH models for returns. For tests with Machine Learning models, the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) recurrent neural networks models, multilayer models such as the Multilayer Perceptron (MLP), and the Neural Basis Expansion Analysis Time Series Forecasting (N-BEATS) model were chosen, all with various parameterizations. After several tests, stability was achieved with a Mean Absolute Percentage Error (MAPE) near to 1% for closing values in most of the models.
Description
Keywords
Bolsa de valores brasileira Ibovespa Aprendizagem automática Previsão financeira Modelos estatísticos Volatilidade