Title: Advancing in portfolio management using machine learning in Brazil

Authors: Adriana Bruscato Bortoluzzo; Marcus Oliveira da Silva; Pedro Raffy Vartanian; Alvaro Alves de Moura Junior

Addresses: Insper, Rua Quata 300, 04546042, São Paulo, SP, Brazil ' Universidade Presbiteriana Mackenzie, Rua da Consolação 930, São Paulo, SP, Brazil ' Universidade Presbiteriana Mackenzie, Rua da Consolação 930, São Paulo, SP, Brazil ' Universidade Presbiteriana Mackenzie, Rua da Consolação 930, São Paulo, SP, Brazil

Abstract: The study aims to compare the performance of machine learning models against conventional linear models and explore their applicability in investment allocation strategies, including discerning factor significance and contributions to return predictions. We conduct a portfolio allocation analysis of Brazilian stocks returns, spanning from January 2007 to June 2022. We built five models using machine learning techniques: gradient boosted trees, random forest, LASSO and ridge regularisation, and a baseline linear model using size, price on equity value and momentum. Our results reveal significant economic benefits associated with the tree-based models, outperforming their linear counterparts. Notably, the long-short portfolio strategy combining the two superior models yields an annual Sharp Ratio of 0.24, demonstrating a remarkable 66% improvement over that of the Ibovespa index. Machine learning can assist in optimising investment portfolios by identifying the most attractive stocks and their respective weightings, leading to better returns for investors while managing risk.

Keywords: machine learning; asset pricing; forecast return; gradient boosted trees; portfolio allocation; Brazilian stocks.

DOI: 10.1504/IJMDM.2026.154603

International Journal of Management and Decision Making, 2026 Vol.25 No.4, pp.417 - 436

Received: 23 Aug 2024
Accepted: 24 Apr 2025
Published online: 07 Jul 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article