Machine learning for intraday returns forecasting in the brazilian stock marketing

• Ano de defesa: 2020
• Autor(a) principal: Alexandre, Henrique Leone
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Elastic net; Lasso; Machine learning; PCR; PLS; Ridge
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/12/12138/tde-10052021-212420/

This paper applies different estimation methods, specialized in dealing with high data dimensionality, to make rolling five-minute-ahead return forecasts using high frequency data, 5 minutes. The methods used are ridge, LASSO, elastic net, PCR and PLS. The explanatory variables are only the lagged returns of their own and of all the other stocks on the Ibovespa index. More than just statistical, the economic sense behind these variables is that they can quickly capture the impact of new information about the companies. The aim of this paper is to perform a comprehensive comparison of out-of-sample forecast performance of stock returns among methods. The results show that Ridge Regression produces the best performance among all methods with a significant advantage. To assess the robustness of the results, different portfolios were formed. The returns obtained for the portfolio built with the most volatiles stocks and the portfolio that exploits the predictability of machine learning methods, even under a conservative assumption on transaction cost, suggest that these approaches appear to be promising for traders.