Análise de desempenho de portfólios que combinam modelos de machine learning e otimização multiobjetivo no mercado de ações brasileiro
Data
Autor(es)
Orientado(es)
Título da Revista
ISSN da Revista
Título de Volume
Editor
Abstract
Building an investment portfolio through careful stock selection is a crucial activity for investors. In this context, artificial intelligence methods emerge as fundamental tools to support investors' decisions. The main aim of this research is to compare different machine learning methods to perform asset pre-selection and to analyze the combination of these methods in evaluating the performance of optimized portfolios. These portfolios are integrated into a multi-objective optimization model that seeks to maximize return and minimize risk. To achieve this goal, the dissertation was divided into the following specific objectives, each structured as an article:i) Investment portfolio optimization: a bibliometric review (Bibliographic Product 1); ii) Asset selection and optimization of investment portfolios with artificial intelligence methods: a systematic and bibliometric review of the literature (Bibliographic Product 2); iii) Performance analysis of portfolios that combine machine learning models and multi-objective optimization in the Brazilian stock market (Bibliographic Product 3). The empirical article proposed in this dissertation was based on the two bibliometric review articles. Using machine learning models, namely: a) Random Forest, b) Multilayer Perceptron (MLP), and c) Extreme Gradient Boosting (XGBoost), each model was trained and validated to pre-select stocks based on financial indicators. Subsequently, after the selection of shares made by each model, portfolio optimization was conducted to determine the allocation percentage in each asset. It is important to highlight that the portfolios selected by the artificial intelligence models were tested through backtesting, in which the following performance indicators were calculated: i) Sharpe index, ii) Treynor index, iii) Jensen's Alpha and iv) VAR . The empirical article showed that the combination of machine learning models and a multi-objective model can generate results significantly superior to the market benchmark. Finally, as a technical product, the dissertation providesthe script and database used in the empirical article (Technological Product 1).
