Um Estudo Sobre Aprendizado de Máquina Aplicado à Modelagem de Retornos de Ações

• Ano de defesa: 2015
• Autor(a) principal: Santos Junior, José Gilmar Alves
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Brasil; UFRN; Programa de pós-graduação em sistemas e computador
• Palavras-chave em Português: Aprendizado de máquina; Mercado de ações; Séries temporais financeiras; Máquinas de vetores de suporte
• Link de acesso: https://repositorio.ufrn.br/jspui/handle/123456789/26066

Stock prices behavior has been subject of research for over a century, and artificial intelligence techniques has been applied to forecast returns since the 1980s. In the present research, we examine the performance of support vector machines to forecast Brazilian stock returns and predictions intervals: based on past values of stock returns and volatilities, both extracted from prices series, we want to figure out if there is some gain over traditional statistical models. Our findings are based on analysis of linear, support vector and hybrid models applied to weekly, daily and intraday data. The empirical evidence suggests that (a) hybrid models performs better on forecasting volatilities series; (b) linear models performs better on forecasting stock returns expected values and prediction intervals; and (c) the approach of treating time series dynamics as a function may lead to results like the ones we’ve got (far bellow the best possible) if the time series is nonlinear with large additive noise.