Ensaios sobre o mercado financeiro brasileiro: aplicação de modelos preditivos e análise de anomalias

• Ano de defesa: 2024
• Autor(a) principal: Holanda Filho, Ivan de Oliveira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Mercado financeiro; Anomalia financeira; Machine learning; Financial market; Financial anomaly; CNPQ::CIENCIAS SOCIAIS APLICADAS::ECONOMIA
• Link de acesso: http://repositorio.ufc.br/handle/riufc/77379

This dissertation is composed of two essays that have the Brazilian stock market as their central theme. The objective of the first test is to apply Machine Learning techniques to predict movements in the Bovespa Index (Ibovespa), which is the most important indicator of the average performance of shares traded on the Brazilian stock exchange, B3 (Brasil, Bolsa, Balcão). To meet specific objectives, different supervised learning models and algorithms were applied with the purpose of evaluating and establishing a comparison of the predictive performance of these models. Using several studies in international literature as a reference, the work analyzes daily movements of the Ibovespa in the period between 2012 and 2022. For these purposes, the Python programming language was used and a database was formatted from different data sources. which include Yahoo Finance, the Central Bank of Brazil and IBGE. Ibovespa movements were modeled using Logit, LASSO, Support Vector Machine, Randon Forest and Artificial Neural Networks models, adopting a combination of predictor variables inspired by technical and fundamental analysis methods. The results highlighted the performance of the Lasso, Logit models and the Artificial Neural Networks algorithm, which showed good performance with accuracies close to 75%. In turn, the second essay is dedicated to analyzing and making inferences regarding the effect of calendar anomalies on the stock market in Brazil. The study is dedicated to reviewing different types of anomalies related to the calendar, of which the political cycle effects and the so-called Halloween effect, January effect and day of the week effect stand out. With the aim of making inferences about the statistical significance of these effects, logistic regression models were applied to study the effects on daily movements, and the ARIMA and ARIMAX time series models were applied to investigate possible effects on returns calculated based on the Ibovespa. The results show evidence that calendar variables do not have significant effects on the daily movements of the index. In turn, different versions of the models estimated to evaluate the correlation of these variables on returns showed different results, which were significant for some months of the year and days of the week. In this aspect, the knowledge produced on this topic becomes vital. In an increasingly globalized world in which markets and investment tools are more accessible to citizens, it is important to verify the occurrence of these events and produce evidence so that agents can define better operating strategies in these markets.