Formação de portfólios com o uso de redes neurais artificiais e distribuição de probabilidade com caudas longas

Abstract

In this experimental study, artificial neural networks (ANNs) and long-tail Probability Ranking are used to construct investment portfolios. The objective is to investigate whether portfolio formation can be seen as a classification problem, leveraging the inherent abilities of ANNs to capture complex relationships, allowing for more informed decisions about portfolio composition. The experiment was conducted using information from 70 randomly chosen assets, from the Brazilian and American markets, and a validation sample composed of all companies belonging to the S&P500 index. The study covers different periods from 2018 to 2023, with more than 585,650 asset observations per day. The technique was compared with other alternative techniques and market portfolios: Minimum Variance Portfolio, Maximum Sharpe Portfolio, Multifractal Trend Fluctuation Analysis (MF-DFA) Portfolio, Berkshire Hathaway Portfolio and S&P500 Index Portfolio. The results indicate that the proposed classification method using the asymmetric probabilities of the Student´s 𝑡���� distribution performs better than market portfolios and traditional portfolios. Furthermore, the results indicate that the combined approach of ANN and security classification based on their asymmetric leptokurtic probabilities demonstrates superiority over portfolios that only use signal classification. As contributions, a new way of forming an investment portfolio is presented, a new efficient market frontier and an R software package called ANNt.